key: cord-0018608-9i5gb9nk
authors: Vatansever, Sezen; Schlessinger, Avner; Wacker, Daniel; Kaniskan, H. Ümit; Jin, Jian; Zhou, Ming‐Ming; Zhang, Bin
title: Artificial intelligence and machine learning‐aided drug discovery in central nervous system diseases: State‐of‐the‐arts and future directions
date: 2020-12-09
journal: Med Res Rev
DOI: 10.1002/med.21764
sha: fcc4b1404f77d075f5c39eb8370aa1fad372976d
doc_id: 18608
cord_uid: 9i5gb9nk

Neurological disorders significantly outnumber diseases in other therapeutic areas. However, developing drugs for central nervous system (CNS) disorders remains the most challenging area in drug discovery, accompanied with the long timelines and high attrition rates. With the rapid growth of biomedical data enabled by advanced experimental technologies, artificial intelligence (AI) and machine learning (ML) have emerged as an indispensable tool to draw meaningful insights and improve decision making in drug discovery. Thanks to the advancements in AI and ML algorithms, now the AI/ML‐driven solutions have an unprecedented potential to accelerate the process of CNS drug discovery with better success rate. In this review, we comprehensively summarize AI/ML‐powered pharmaceutical discovery efforts and their implementations in the CNS area. After introducing the AI/ML models as well as the conceptualization and data preparation, we outline the applications of AI/ML technologies to several key procedures in drug discovery, including target identification, compound screening, hit/lead generation and optimization, drug response and synergy prediction, de novo drug design, and drug repurposing. We review the current state‐of‐the‐art of AI/ML‐guided CNS drug discovery, focusing on blood–brain barrier permeability prediction and implementation into therapeutic discovery for neurological diseases. Finally, we discuss the major challenges and limitations of current approaches and possible future directions that may provide resolutions to these difficulties.

To help the reader better understand AI/ML applications in CNS drug discovery, we provide a summary of AIbased algorithms that are widely used in drug discovery. AI uses a large variety of models to build up intelligent systems, which can be classified by learning procedures. AI is frequently used to denote ML algorithms-yet they are not the same. So, it would be worth clarifying both terms at first. In this review, we follow the US Food and Drug Administration's (FDA) definition of AI. They describe AI as "the science and engineering of making intelligent machines", while ML is "an artificial intelligence technique that can be used to design and train software algorithms to learn from and act on data", 5 adding that all ML techniques are AI techniques, but not all AI techniques are ML techniques. Here, we provide brief definitions of the basic learning algorithms in Table 1 , as these are most relevant in the context of drug discovery. AI-related learning techniques are broadly categorized as supervised, unsupervised, semisupervised, active, reinforcement, transfer, and multitask learning. Different algorithms are used in those learning architectures to perform specific tasks such as classification or clustering. However, success with AI requires more than training an AI model. A robust AI workflow involves (i) formulating a problem, (ii) preparing data, (iii) extracting features, (iv) selecting training and testing data sets, (v) developing a model, (vi) training the model and testing its performance (cross-validation), and (vii) applying the model to testing data sets and refining the model. Figure 2 displays the basics steps of building an AI architecture.

A key consideration in early drug discovery is to identify drug candidates with the desirable initial characteristics, which are then further developed into chemical structures with the desirable potency against the target molecule.

Molecular descriptors and fingerprints are used for quantifying such physicochemical characteristics of both chemical entities and their biological target molecules. Molecular descriptors are experimentally quantified or theoretically characterized properties of a corresponding molecule that represent the physical, chemical, or topological characteristics, while molecular fingerprints are more complex descriptors that are encoded as binary bit strings. 6, 7 Both molecular descriptors and fingerprints have crucial functions in ML-based applications in drug discovery processes such as target molecule ranking, 8, 9 similarity-based compound search, 10-15 virtual screening, 16, 17 QSAR analysis, 18, 19 ADME-T prediction of lead molecules. [20] [21] [22] [23] There are various tools for molecular descriptor and fingerprint calculation, and each has a different set of features. Here, we explain the molecular descriptors (i.e., target protein descriptors and compound descriptors) and compound fingerprints, and provide the highly used programs for generating them (i.e., sequence-based tools and structure-based tools) in the Supporting Information. Additionally, Chuang et al. 24 comprehensively discussed how AI-based methods (i.e., deep learning [DL]) could address limitations of molecular descriptors and fingerprints and thereby improve the predictive modeling of compound bioactivities.

A dominant approach to drug discovery is to design drug molecules that will reverse a disease course by modulating the activity of a target. 25 Drug development often begins with identification of a novel target whose modulation can lead to a therapeutic benefit with an acceptable safety margin. This is followed by validating the role of the selected target in disease in in vivo models and, ultimately, in clinical trials. Therefore, the ultimate success of a drug development project depends on early identification of promising drug targets.

T A B L E 1 AI-related learning techniques used in drug discovery

Supervised learning • A predictive model trained on data points with known outcomes ("labeled data") • Two types of problems:

Regression: Model finds outputs that are real variables Classification: The model divides inputs into classes or groups

Naïve Bayes Classification • A "probabilistic classifier" that determines the probability of the features occurring in each class by treating every feature independently to return the most likely class based on the Bayes rule. • Particularly suited when the dimensionality of the inputs is high.

• A discriminative classifier that outputs an optimal hyperplane to categorize new examples. The vectors that define the hyperplane are the support vectors.

• An ensemble of simple tree predictors that vote for the most popular class for classification problems. In the regression problems, the tree responses are averaged to obtain an estimate of the dependent variable. • Overfitting is less likely to occur as more decision trees are added to the forest.

• A nonparametric algorithm based on feature similarity by assuming that similar things exist in close proximity. • Useful for a classification study when there is little or no prior knowledge about the distribution data.

• A method that learns from input data based on layers of connected neurons consisting of input layers, hidden layers, and output layers.

• A collection of neurons organized in a sequence of multiple layers. • Type of artificial neural network with several advantages (i.e., shared weights [parameter sharing), spatial relations, and local receptive fields • Learning can be supervised, unsupervised, or semisupervised. • End-to-end learning and transfer learning are the major approaches performed by the deep neural network. • Autoencoders and generative adversarial networks are the two specific forms of deep neural networks.

• A statistical approach to find relationships between dependent variables and one or more independent variables.

Unsupervised learning • A self-organized model that organizes the data in some way or describe its structure to learn underlying patterns of features directly from unlabeled data.

K-means clustering Clustering • A classification method that divides data into k groups by minimizing within-group distances to the centroid (Continues) VATANSEVER ET AL.

| 1431 T A B L E 1 (Continued)

Fuzzy clustering Clustering • A form of clustering (Fuzzy C-means clustering) in which each data point can belong to more than one cluster. • It computes the coefficients of being in the clusters for each data point.

Hierarchical clustering Clustering • A classification method that builds a hierarchy of clusters by merging two close clusters into the same cluster. This algorithm ends when there is only one cluster left.

Dimensionality reduction

• A nonparametric statistical technique that uses an orthogonal procedure to transform a set of correlated features to new independent variables called principal components Independent component analysis Dimensionality reduction

• A statistical method that separates a multivariable output into statistical independent additive components Autoencoders Dimensionality reduction

• A deep neural network trained with backpropagation to reconstruct its original input

Deep belief nets Dimensionality reduction

• Probabilistic generative models with many layers of stochastic, latent variables. Each layer is a Restricted Boltzmann machine.

Anomaly detection • Deep generative models that use two neural networks, pitting one against the other (thus the "adversarial") to generate new synthetic but realistic instances of data.

Self-organizing map Dimensionality reduction

• A competitive learning network that reduces the input dimensionality to represent its distribution as a map.

Semisupervised learning • A combination of supervised and unsupervised learning methods that uses a small amount of labeled data and also a large amount of unlabeled data during training to gain more understanding of the sample population.

Active learning • A particular case of semisupervised learning, where the algorithm is allowed to query the user for the label of a subset of training instances • Used to construct a high-performance classifier while keeping the size of the training data set to a minimum by actively selecting the valuable data points Reinforcement learning • Dynamic programming that trains algorithms using a system of reward and punishment to maximize the performance.

Transfer learning • A deep learning technique enables developers to harness a neural network used for one task and apply it to another domain. • It allows the reuse of a pretrained deep neural network on a new task with only a small amount of data. • Useful when the data is insufficient for a new domain to be handled by a neural network, and there is a big preexisting data pool that can be transferred

• An approach to inductive transfer that improves generalization performance of multiple related tasks by leveraging useful information among them. • Useful when there are multiple related tasks, each of which has limited training samples

Multiple kernel learning • A flexible learning method that use a predefined set of kernels and learn convex combinations of kernels over potentially different domains. • Used when there are heterogeneous sources of data for the task at hand A good drug target need be relevant to the disease phenotype as well as be suitable for therapeutic modulation ("druggable"). Biological and technological advances have continuously driven the generation of high-throughput biomedical data, which present new opportunities for early identification of potential drug targets. However, the analysis of such large-scale multidimensional biological data requires effective techniques that can produce accurate predictions for target identification. AI/ML has emerged as a powerful technology for analyzing the rapidly increasing multiomics data in the identification of potential therapeutic targets.

In literature, the "target identification" term is often used in two different contexts: Target discovery and target deconvolution. 26 The first is the discovery of a new disease target whose modulation would have therapeutic effects. The second is the identification of a target with a known active compound, which is also called "target fishing." To avoid confusion, we will use context-specific terms of target discovery and deconvolution rather than generic target identification.

Ensemble learning • A meta-algorithm that combines decisions from multiple models into one predictive model to decrease variance (bagging), bias (boosting), or improve predictions (stacking).

End-to-end learning • A deep learning process in which all of the parameters are trained jointly, rather than step by step. It allows the training of a deep neural network based on raw data without descriptors. Since the pipeline is replaced with a single learning algorithm, it goes directly from the input to the desired output and thereby overcome limitations of the traditional approach.

Note: The rows with gray backgrounds show the basic learning categories and their definition, while the rows following supervised and unsupervised learning parts display the different algorithms used in these categories.

F I G U R E 2 The basic steps of building an artificial intelligence (AI) platform for drug discovery. The process for developing an AI model as follows: (1) Define the problem appropriately (objective, desired outputs, etc.), (2) prepare the data (collection, exploration and profiling, formatting, and improving the quality), (3) transform raw data into features and select meaningful features (a.k.a. feature engineering), (4) split data into training and validation sets, (5) develop a model, (6) train the model with a fraction of the data, test its performance (crossvalidation) and tune its parameters with the validation set (7) evaluate model performance on the validation set and refine the model, and (8) Drug discovery begins with the identification of a novel target candidate that is followed by a target evaluation consisting of experimental target validation and theoretical assessment of its ability to bind small molecule drugs (druggability). 27 The target discovery process includes identification of targets that play a role in the disease pathophysiology, 28 assessment of druggability, and prioritization of candidate targets. However, because of the complex nature of human diseases, this process often requires more comprehensive approaches that integrate available heterogeneous data and information to understand the molecular mechanisms underlying disease phenotypes and identifying the patient-specific changes. 29 To overcome such difficulties, researchers have applied AI/ML methods to predict "reliable" drug targets. The following sections demonstrate the AI/ML applications in different stages of the target discovery process ( Figure 3 ).

In complex heterogeneous diseases, classifying patients into clinically and biologically homogenous subtypes is critical for understanding disease pathophysiology and developing appropriate subtype specific therapies. 30 Researchers have developed AI/ML algorithms that can integrate multiscale data to identify different etiological subtypes of complex diseases. For example, Shen et al. 31 developed iCluster, a joint latent variable model for integrative clustering analysis, which was applied to breast cancer and lung cancer and identified subtypes characterized by concordant DNA copy number changes and gene expression. 31 Yuan et al. 32 also integrated copy number variation and gene expression data by using a nonparametric Bayesian model and discovered prognostic subtypes in prostate cancer and breast cancer. 32 Zhang et al. 33 revealed the prognostic subtypes in neuroblastoma using DL-based integration of multi-Omics data and K-means clustering analysis. Recently, Gao et al. 34 described a cancer classification method, deep cancer subtype classification (DeepCC), based on DL of functional spectra, which is a vector of gene set enrichment scores associating with biological functions for each patient sample.

Overall, in recent years, AI/ML methods have been employed to analyze large-scale genomic and other molecular profiling data in cancer for the identification of distinct, molecular disease subtypes. However, such AI-based subtyping analysis have not been widely applied to other complex diseases. Implementation of robust and scalable AI/ML techniques for discovery of disease subtypes paves the way for developing more efficacious therapeutic strategies.

One of the most challenging tasks in target discovery is the prediction of disease-causing genes from huge amount of genetic and functional genomic data. To predict these disease-associated genes from multiomics data, researchers have employed various ML classifiers, [35] [36] [37] [38] including Random Forest (RF)-, 39,40 support vector machines (SVM)-, 41, 42 and decision tree (DT)-based classifiers. 43 More detailed information about those applications can be found in the Supporting Information. Besides the ML-methods using multiomics data, DriverML, 44 a supervised F I G U R E 3 AI-guided target discovery. AI/ML methods can efficiently analyze all available information to speed up the discovery of disease-related drug targets. Specifically, AI/ML methods are utilized for disease subtyping, identification of disease driver genes and microRNAs, alternative splicing prediction, triaging of novel drug targets, modeling of three-dimensional target structures, and druggability assessment. AI, artificial intelligence; ML, machine learning [Color figure can be viewed at wileyonlinelibrary.com] learning tool, identified cancer driver genes based on DNA sequence alterations from The cancer Genome Atlas (TCGA) data with superior performance over the other tools such as DriverDBv2 database. 45 In addition to ML classifiers, DL-based methods have been implemented in more recently developed tools. For example, deepDriver 46 trained similarity networks and a convolutional neural network (CNN) on mutation data simultaneously to predict driver genes with better performance than the competing approaches when applied in breast cancer and colorectal cancer. In another example, Peng et al. 47 used deep neural network (DNN) to reduce the dimensionality of transcriptomics data to predict Parkinson's disease genes. This DNN-based tool, namely, N2A-SVM, consists of three steps, including extraction of vector representation of each gene in the protein-protein interaction (PPI) network, dimension reduction for the obtained vector with autoencoder, and prediction of the genes associated with Parkinson's disease using SVM.

Multitask learning has also been employed for the prediction of cancer driver genes. LOTUS, an ML-based algorithm, predicts cancer driver genes in a pan-cancer setting, as well as for specific cancer types, using a multitask learning strategy sharing information across cancer types. 48 For the readers who want to learn more about opportunities and challenges in predictive modeling for multiomics data sets, we suggest the review paper of Kim and Tagkopoulos. 49 Different from the tools using omics data sets, BeFree 50 was developed to extract relations between genes and diseases from text mining. This supervised learning approach utilized natural language processing (NLP) Kernel methods to identify gene-disease associations from the abstracts collected by Medline.

The challenges in targeting disease proteins have shifted the focus in target selection to disease microRNAs (miR-NAs), which are small noncoding RNAs that regulate gene expression by targeting messenger RNAs. 51 miRNAs are regarded as high-potential drug targets due to their involvement in various diseases. 52 Therefore, considerable effort has been devoted in identifying relationships between miRNAs and diseases using ML-based methods, such as the network based approach by Xu et al. 53, 54 and RLSMDA. New strategies in miRNA target discovery have utilized neural networks (NN). Zeng et al. 55 developed a NN method, NNMDA to predict miRNA-disease associations with the best performance among the existing algorithms. Application of NNMDA to lung neoplasm and breast neoplasm predicted novel disease-related miRNAs. Very soon after that, Zheng et al. 56 published a new ML-based method, MLMDA, which predicts miRNA-disease associations by integrating miRNA sequence, disease semantics, miRNA-disease association, and miRNA function but with slightly worse performance than NNMDA.

Alternative splicing (AS) plays a fundamental role in gene expression regulation and protein diversity by causing the generation of different transcripts from single genes. 57 Understanding the genetic variation in splicing signals is within the scope for AI/ML-based models to discover therapeutic opportunities through novel targets. For splicing prediction and analysis, a web tool, AVISPA, 58 has been developed. For a given exon and its proximal sequence, AVISPA predicts if the exon is alternatively spliced and if it has associated regulatory elements by using a Bayesian NN classifier. However, the method by Leung et al. 59 

While increasing effort has been devoted to nominating novel drug targets involved in diseases, experimental validation of identified target candidates is an expensive and time-consuming task. 63 68 predicted and prioritized over 3,000 candidate age-related human genes using three positive unlabeled learning algorithms, Naïve Bayes, Spy, and Rocchio-SVM. They ranked the human genes according to their implication in aging based on binary gene features from 11 human biology databases. 68 Target protein structure prediction AI/ML architectures have been applied in protein structure prediction over 30 years, and several groups have comprehensively reviewed those strategies. [69] [70] [71] [72] [73] Therefore, we will focus on recent applications in this field. Also, we provide a background of conventional protein structure prediction methods (i.e., template-based and templatefree) for those who want to learn more about this field in the Supporting Information.

Since 1994, the Critical Assessment of protein Structure Prediction (CASP) competitions have been organized biannually for blind evaluation of the state-of-the-art methods that predict three-dimensional (3D) protein structures from protein sequences. There, each group submits structure predictions for each of the given protein sequences for which experimentally determined structures were sequestered. In December 2018, Google's AI firm

DeepMind won the CASP13 competition with its latest AI system, AlphaFold. DeepMind's success generated significant interest in the protein folding community, where the researchers published several articles discussing the method. [74] [75] [76] [77] AlphaFold determines the 3D shape of a protein from its amino acid sequence by merging two approaches: (i) Inferring physical contact in protein structure from residue covariation in protein sequence based on coevolution analysis of a multiple sequence alignment and (ii) identifying coevolutionary patterns in protein sequences as contact distributions by using DNNs and convert them into protein-specific statistical energy potentials. AlphaFold system has achieved an unprecedented prediction accuracy among the ab initio methods.

Although AlphaFold's performance represents a big leap in protein structure prediction, its accuracy still needs to be improved.

Inspired by AlphaFold as well as previous successful applications of DL to residue contact predictions, 78 researchers have developed different strategies to improve the protein structure prediction, including a deep residual network model, 79 a fragment library that is built using deep contextual learning techniques called DeepFragLib 80 and a community-built, open-source implementation of Alphafold (i.e., ProSPr). 81 The emergence of DL has suggested the rethinking of how to address the problem of protein structure and thereby, encourages the new approaches. RGN (recurrent geometric network) is an end-to-end differentiable model that takes a sequence of amino acids and position-specific scoring matrices (a summary of residue propensities for mutation) as inputs and outputs a 3D structure. In contrast to the complexity of conventional structure prediction models, a trained RGN model is a single mathematical function that is evaluated once per prediction. Hence, a trained RGN makes predictions six to seven orders of magnitude faster than other methods. The same lab developed the RGN also published a data set to provide a standardized resource for training and assessing ML frameworks for predicting protein structures. The data set called ProteinNet integrates sequence, structure, and evolutionary information into preformatted input/output records. ProteinNet is available in a public repository, https://github.com/ aqlaboratory/proteinnet.

Going beyond the structure prediction, researchers have employed the ML for the prediction of protein dynamics since target proteins are dynamic and sample multiple states. Ung et al. 82 used RF to classify pharmacologically relevant conformations of protein kinases. Using a 3D-CNN, Okuno et al. 83 developed DEFMap, which extracts the dynamics information hidden in a given cryo-EM density map. This approach allows us to grasp the dynamic changes associated with molecular recognition and the accompanying conformational selections from the cryo-EM structure, which derive insights into the protein function as well.

The studies discussed above clearly demonstrate the utility of the AI/ML frameworks to make predictions of protein structural features from sequence alone. Rost et al. 84 comprehensively discussed how ML algorithms help to understand the effects of protein sequence variants on protein function and pathways. AI/ML algorithms are readily available for structural biologists to quickly estimate protein structures. Of course, the accuracy and speed of a framework will depend on the creativity in problem formulation, network design, and data storage. We can look forward to a rapid growth in the number of AI/ML applications in the prediction of protein structures.

In target discovery, another crucial step is the evaluation of the target's druggability, "the likelihood of being able to modulate a target with a small-molecule drug". 85 In drug design, a selected target must have the biophysical properties that allow it to bind small molecules with drug-like properties. ML-based models usually estimate a 

Target deconvolution (a.k.a. target fishing) is an important step following the discovery of compounds that cause a desirable change in phenotype. Understanding the binding targets of phenotypic screen-derived compounds can help design better analogs, find potential off-targets, and thereby explain observed adverse events. However, existing experimental approaches for target deconvolution are labor, resource, and time-intensive. Researchers have adapted computational approaches to target deconvolution problems to reduce the required sources for the experiments. Several studies implemented AI/ML algorithms into computational target deconvolution tools for higher predictive power. For example, Schneider and colleagues have widely applied self-organizing maps (SOMs)

to predict the macromolecular targets of compounds. [94] [95] [96] [97] They preferred to use "fuzzy" molecular VATANSEVER ET AL.

| 1437

representations, such as pharmacophoric feature descriptors, since such fuzzy molecular representations demonstrated greater scaffold-hopping potential than atomistic approaches in similarity searches. On the basis of the similarity of pharmacophoric features, their unsupervised SOM algorithm clustered the query molecules with unknown targets as well as drug-like molecules with known targets. Hence, the trained SOM was able to transfer the knowledge of annotated drug targets to query molecules that are the nearest neighbors to known drugs. 94 They have applied this SOM approach to identify the macromolecular targets of de novo-designed molecules, 95 complex natural products, 94 fragment-like natural products, 96 and a natural anticancer compound. 97 Besides the SOM models, a multiple-category Naïve Bayesian model was developed for the rapid identification of potential targets for compounds based on only chemical structure information, which is the connectivity fingerprints of compounds from 964 target classes in the WOMBAT (World Of Molecular BioAcTivity) chemogenomics database. 98 Moreover, a target-fishing server named RF-QSAR was built based on target SAR models that were created using an RF algorithm to rank candidate targets for a query compound. 99 A recent target identification tool, BANDIT, 100 uses a Bayesian approach to integrates six distinct data types-drug efficacies, posttreatment transcriptional responses, chemical structures, reported side effects, bioassay results, and known targets.

In the identification of the novel targets of drugs, there has been increasing interest in predicting drug-target interaction (DTI), given its relevance for side effect prediction and drug-repositioning attempts. 101 The availability of heterogeneous biological data on known DTI has enabled the development of various AI/ML-based strategies to exploit unknown DTI, 102 including ensemble learning, 103-106 tree-ensemble learning, 107 active learning, 108 DL, 109 end-to-end DL, 110 and kernel-based learning. [111] [112] [113] [114] [115] Such AI/ML-enabled data integration strategies outperform the traditional methods in classifying both positive and negative interactions, 110 improved the quality of the predicted interactions, and expedited the identification of new DTI. 115 

To identify new compounds with potential interactions to target proteins, researchers commonly use HTS, an in vitro method that automatically tests large compound libraries towards a specific target. However, high cost and low hit rate of HTS have expedited the development of virtual screening (VS) alternatives, which enable cheaper and faster screening of larger compound libraries. 116, 117 VS predicts the compounds that most likely to bind to a protein of interest using various approaches. Two broad categories of VS are structure-based VS (SBVS) and

ligand-based VS (LBVS)-the former takes the structures of target proteins as input, 118, 119 and the latter uses information on known inhibitors. 120 LBVS is basically "analoging" to some extent based on that similar molecules tend to exhibit similar properties, 121 and it also helps to build better pharmacophore models. SBVS and LBVS are often used synergistically: Leads from SBVS can be improved with LBVS, and data from improved yields can be used to refine models for SBVS. 122 For achieving better performance in VS workflows, AI/ML-based methods have been utilized for both SBVS and LBVS. We will begin with the application of AI/ML methods in SBVS and continue with their applications in LBVS in the next section.

SBVS requires the 3D structure of a target protein to predict whether a compound is likely to bind the target. One widely used method to do this is molecular docking, which models the protein-ligand complex based on the estimated interaction energy. In recent years, ML methods have been employed in SBVS workflow to increase the robustness and accuracy of scoring functions (SFs), conformational sampling and ranking. Researchers have developed SFs using RF-, 123-126 SVM-, 127, 128 and NN-129-134 based learning algorithms and they outperformed the conventional SF predictions. 135 However, no ML-based SF is superior to all the other approaches in all respects. 136 Indeed, the performance of an SF differs from target to target. 137 Therefore, researchers have developed ML-based, target-specific SFs to improve the efficiency of existing SFs for kinases, [138] [139] [140] [141] histone methyltransferases, 142 cyclin-dependent kinases and G protein-coupled receptors (GPCRs), 137 and cytochrome P450 aromatase. 143 Moreover, such ML-based models have been applied to post-docking processes to improve the accuracy of molecular docking. For example, ML algorithms 142,144-148 improve pose/compound selection by automating the evaluation of docked ligands, which was done manually before. 149 Details about ML-based scoring functions and AI/ML applications in the post-docking stage can be found in the Supporting Information.

When the 3D structure of a given target is available, SBVS approaches (i.e., molecular docking) can be employed.

However, LBVS methods are the only option if the 3D structure of the target protein is not known. In contrast to the molecular docking that predicts the binding pose of ligands to the target protein using the protein structure, LBVS is based on the principle that ligands structurally similar to an active compound tend to have similar activity. 150 Hence, LBVS requires the information of known active compounds rather than the target protein structure. In drug discovery efforts, researchers often have a set of active compounds generated from testing molecules in biochemical or functional assays without knowing the target protein structure. In such cases, the LBVS approach can be utilized to find new ligands by assessing the structural similarity of candidate ligands to the known active compounds. The challenge is thereby to find an appropriate model for similarity that relates compound features to assay outcomes. In recent years, ML has emerged as an attractive approach to boost the predictive power of LBVS models. The specific aims of ML approaches include prediction of the active compounds against a particular target using models trained on input data sets, discrimination of drug modules from nondrug ones, and prioritization of compounds based on the probability of activity. For these purposes, researchers have used SVMs, Bayesian architectures, and artificial neural networks (ANNs) ( Table S2 ). Further information regarding AI/ML applications in LBVS is available in some comprehensive review papers. 136, 151, 152 On the contrary, one of the most recent advances in AI/ML-based LBSV was made by Stokes et al. 153 They successfully discovered new antibiotics by employing graph convolutional networks (GCN), whose outstanding performance over conventional ML models in predicting molecular properties was confirmed by two studies. 154, 155 Using their GCN model, the authors performed a large-scale screening and identified a promising new antibiotic, halicin. 153 In conclusion, the advances in selection and design of AI/ML algorithms for LBVS and the availability of large bioactivity data sets have enabled more accurate and faster selection of compounds that are predicted to be active against a particular target and will undergo further experimental assays eventually. Although traditional ML classifiers had been widely used in LBVS, recent successful applications have shown GCN's potential to become a popular approach for LBVS. 151 

QSAR models are developed to identify a mathematical relationship between the physicochemical properties, which are represented by molecular descriptors, and biological activity of chemicals. These models play a prominent role in drug optimization, providing a preliminary in silico evaluation of essential attributes related to the activity, selectivity, and toxicity of candidate compounds. 156 | 1439 SVM, [161] [162] [163] Naïve Bayesian, [164] [165] [166] [167] [168] [169] [170] [171] [172] [173] and ANN 143, [174] [175] [176] [177] [178] [179] [180] [181] [182] [183] [184] ) have been extensively employed in QSAR modeling (For the detailed discussion of the applications, see the Supporting Information). Notably, the RF algorithm is commonly used as a classification and regression tool 159 and considered to be the golden standard in QSAR studies. 185 Hence, the performance of new QSAR prediction tools often is compared with that of RF. Many RF-based QSAR models have been developed, such as pQSAR, 186 a method for the soluble epoxide hydrolase, 187 and a model for Janus kinase 2. 188 When the predictive performance and interpretability of RF-based QSAR models are compared to those of two widely used linear modeling approaches-SVMs and partial least-squares, RF not only yields better predictive performance but also enables an amenable chemical and biological interpretation. 189 In the applications of NN to QSAR prediction, researchers use the data from a single assay using molecular descriptors as input to train an NN and record activities as training labels. However, the efficiency of those simple single-task NN models depends on having sufficient training data in a single assay. To benefit from the data obtained from multiple assays, researchers aim to develop multitask QSAR models. Several groups constructed the multitask learning structures based on plain feed-forward NN to avoid overfitting by learning multiple bioassays simultaneously. [190] [191] [192] [193] [194] [195] [196] Moreover, multitask QSAR models were also utilized for predicting the activity against multiple targets. [197] [198] [199] In 2012, a data science competition (www.kaggle.com/c/MerckActivity) was organized to find state-of-the-art methods for QSAR. Using multitask DNNs, the winning team improved the prediction accuracy by 15% over the baseline RF method. 200 Since its introduction into the QSAR modeling, 159 RF has served as a "golden standard" and no QSAR methods other than DNNs outperform it. On the contrary, in the following DREAM challenges on predicting kinase-drug-binding, 201 the models based on DL algorithms did not perform better than the other learning algorithms. 202 In the next study, using the DNNs, Ma et al. 185 showed that DNNs could make better prospective predictions than RF, on large and diverse QSAR data sets. However, they could not propose a clear strategy for choosing between multitask and single-task DNNs. Xu et al. 203 Ensemble-based ML approaches combining several basic models have also been used to overcome the weaknesses of individual learning models and thereby improve the overall performance of the QSAR predictors.

There are various ensemble learning applications in QSAR predictions, including data sampling ensembles, method ensembles, and representation ensembles. Recently, Kwon et al. 205 proposed a model that is a combined ensemble of sampling, method, and representation with an end-to-end NN-based individual classifier. Their ensemble model achieved better performance than the individual models in QSAR prediction.

2.5 | AI/ML applications in prediction of physicochemical properties and ADME-T

Physicochemical properties indicate all aspects of drug action and profoundly affect the clinical success rates of drug candidates. A small molecule drug candidate must be sufficiently soluble and permeable to access its site of action and thereby engage its targets, with optimal safety profiles. Therefore, accurate prediction of the physicochemical characteristics can be beneficial for designing a new chemical entity with suitable pharmacokinetic and pharmacodynamic profiles. Researchers have adopted ML-driven approaches to predict some key physicochemical properties, such as water solubility, membrane permeability, and lipophilicity. We provide a detailed description of each property and discuss the ML-based techniques that specifically predict the water solubility, 206 155 achieved an unprecedentedly high accuracy in predicting molecular physicochemical properties.

A successful drug development pathway must include the evaluation and optimization of pharmacokinetics, pharmacodynamics, and safety profiles of a candidate molecule. In early drug discovery, evaluation of the ADME-T properties help researchers select good drug candidates for further development. ADME-T properties are estimated to be responsible for half of all clinical failures. 220 In this context, in silico ADME-T prediction models have received considerable progress over the past 40 years due to the availability of many compounds with known pharmacokinetic properties. 23, 221 Prediction models usually try to build a direct relationship between a set of molecular descriptors and a given ADME-T property. 222 These methods represent a compound by chemical descriptors as input features such as atom counts, surface areas, weight, van der Waals volume, partial charge information, and the presence or absence of a predefined substructure. The key substructures responsible for certain toxicity are structural alerts, of which detection in given small molecules could be used for toxicity prediction. 223 On the contrary, in these models, the toxicity properties of input compounds are HTS assay measurements of toxic effects that are highly relevant to human health, including nuclear receptor pathway assays (i.e., aryl hydrocarbon receptor, aromatase, androgen and estrogen receptor, PPAR-gamma) and stress response pathway assays (i.e., ATAD5, antioxidant responsive element, heat shock factor response element, mitochondrial membrane potential, p53). 224 While the conventional approaches have yielded physiologically based pharmacokinetic and pharmacokinetic-pharmacodynamic/quantitative systems pharmacology models, researchers have applied AI/ML algorithms to produce high-quality models with improved accuracy and thus provide meaningful predictions of ADME-T responses using chemical structure information. For predicting regulators of drug ADME-T properties, the classification models-DT, K-nearest-neighbor (KNN), SVM, RF, and NN have been extensively used. Even beyond that, the introduction of DL models has led to further developments in this area. As a good example of recent advancements in AI. ML-aided ADME-T prediction, Alchemite 225 -a DL model-predicts ADME-T properties by imputing heterogeneous drug discovery data, including multitarget biochemical activities, phenotypic activities in cell-based assays, and ADME-T endpoints.

Moreover, the introduction of capsule networks, a new class of DNN architectures, has remarkably improved the ADME-T prediction. To predict the cardiotoxicity of drugs, Wang et al. 226 Traditional methods for generating novel chemical structures depend on the previously defined reaction or transformation rules, which bias the chemical space towards prior chemical knowledge. AI/ML-based generative models are entirely data-driven without relying on any explicit rules and can generate new molecules that are not present in a training set. Briefly, these generative models first learn from data, then create an abstract representation of the data, and finally use this representation to generate new data instances. 235 Thus, these generative models demonstrate all aspects of an artificially intelligent system (i.e., problem-solving, learning from experience, and coping with new situations). 235 Recent de novo molecule-generative models with an ML structure include adversarial autoencoders Another commonly used drug design approach is to generate new analogs/similar drugs of a given set of drugs.

In such cases, the transfer learning models have been integrated into NN architectures to increase the prediction accuracy by taking knowledge acquired from training on a previous problem and applying them to a new but related problem. 249, 250 In the generative drug design models above, many ML architectures use the SMILES as molecular representation. SMILES provides a linear representation, referred to as a SMILES string that can be translated into a graph and enables a straightforward application. However, it has one or more limitations: Generated SMILES may not represent a chemically feasible structure, and even a single character alteration in a SMILES representation can change the underlying molecular structure significantly. 251 To overcome its limitations, researchers proposed several solutions like converting SMILES strings into a new SMILES-like syntax 252 or utilizing grammatical evaluation of the SMILES syntax. 253 Besides the SMILES string representation, molecular graphs have also been used to train ML-based molecule generation algorithms. 254 In molecular graph generators, structures are directly represented as graphs in every step and substructures are inferred from the partially generated molecular graphs. 255 Examples of such ML models to design de novo molecules based on graph representation includes GANs 256,257 and

VAEs. 258, 259 In addition to the models mentioned above, some AI/ML-driven de novo molecule design tools are distinguished by introducing novel approaches. An automated de novo molecular design tool, DINGOS, 260 has been developed to emulate the approach of a synthetic chemist. It assembles drug-like new compounds through modular and synthetically feasible design schemes, considering the synthetic feasibility of each step. In brief, the DINGOS algorithm combines a rule-based approach with an ML model trained on known successful synthetic routes, while the former ensures the synthesizability and the later provides a directed approach to limiting the output molecules to compounds with desirable similarity to the template. Another remarkable ML-based generative approach is proposed by Méndez-Lucio et al., 261 which bridges systems biology and molecular design. To our knowledge, it is the first AI/ML-based drug design tool that combines transcriptomic and structural data. Conditioning a GAN architecture with compound-induced transcriptomic data (i.e., L1000 data set), they can automatically design molecules that potentially produce the desired transcriptomic outcome. Their model allows the design of activelike molecules for a desired target using just gene expression signature of target perturbation. However, the current version is not capable of generating compounds that can reverse disease-related gene expression signatures. Also, its performance has not been evaluated in a real drug-discovery setting yet.

Among all the studies of AI/ML-based generative molecular design, maybe the most-mentioned 262 Second, the seemingly novel compound is highly similar to the widely used cancer drug ponatinib, indicating the limitation of the approach 265 in assessing truly novel scaffolds. Therefore, there is still room for improvement of AI/ML-inferred small molecules to obtain a clinical candidate.

Personalized drug response prediction aims to improve the targeted therapy response in complex diseases like cancer. 266 However, the limited application of candidate drugs in clinical settings and the heterogeneity among cancer patients make it difficult to tailor therapy for each individual cancer patient. Personalized treatment design requires predictive methods that are capable of exploiting large, heterogeneous, and sparsely sampled data sets.

Accurate AI/ML-based models employing in vitro and in vivo data sets have the potential to improve the prediction of response of cancer cells to a given compound. There are various AI/ML models to predict drug sensitivity and anticancer drug response. In such efforts, elastic net regression, 267 

In the treatment of complex diseases such as neurological disorders, diabetes, cancer, or cardiovascular disease, drug combinations are highly utilized for medical intervention. Coadministration of drugs in the treatment aims to enhance efficacy, reduced toxicity, and prevent the emergence of resistance. Drug combinations are classified as synergistic, antagonistic, or additive. Drug synergy is the interaction of two or more drugs, causing the total effect of drugs to be greater than sum of individual effects of each drug. 279 If drugs act synergistically, lower doses of each drug could potentially be enough to provide the desired outcome allowing for less adverse effects. Opposite to synergism, the antagonistic combination means that the combined activity of the drugs is lower than the response of the individual agents. 280 Finally, a drug combination is considered to be additive when the response of each drug neither masks nor enhances the efficacy of others. 281 Although combinatorial therapy has advantages VATANSEVER ET AL.

| 1443 over monotherapy, developing a new drug combination regimen that can be transferred to the clinic is still challenging. So far, the effective drug combinations have been suggested based on either clinical experience or HTS of drug pairs at different concentrations on cell lines. However, the former involves the risk of harm to patients, and the latter is unfeasible to test the complete combinatorial space. 282 To accelerate conventional combinatorial therapy efforts, AI/ML algorithms have begun to be utilized for prioritizing the drug pairs and exploring the larger combinatorial space. Tonekaboni et al. 283 introduced some examples of various ML-based prediction frameworks for drug-drug interactions. To avoid duplication, we overview the AI/ML applications in combinatorial therapy after that time, including the applications in cancer [284] [285] [286] [287] [288] and depression treatment, 289 antimalarial, 290 and antibiotic 291 discovery, along with the available AI/ML-based tools to predict the synergistic effects of drug combinations [292] [293] [294] in the Supporting Information.

In addition to the synergistic effects, drug-drug interactions can induce unexpected adverse drug reactions.

Such adverse reactions caused by drug-drug interactions could lead to death in some extreme cases. 295 Therefore, AI/ML-based models have been developed to predict the risk of side effects due to drug-drug interactions.

Applications of GCN, 296 Shankar et al. 300 predicted the adverse drug reactions of coadministered drug pairs using an ANN trained on transcriptomic data, compound chemical fingerprint, and Gene Ontologies. 300 

Drug development and trials in animals and humans is a time-consuming and expensive process. In general, the whole process for developing a new FDA-approved drug requires 10-17 years of period and the tremendous cost of $2.6 billion. 301 However, high expenditures for drug development has not been able to increase the rate of approved drugs. 302 Among the reasons for this limited approval rate, a key factor is the continued adherence to the classical "one gene, one drug, one disease" paradigm in the traditional drug development. 303 Since drug targets do not operate in isolation from the biochemical system, each DTI must be studied in a broader integrative context. 304 This approach provides new insights into "off-target" effects (i.e., side effects), resistance to precision therapy, and drug mechanism of action that can inform drug-repurposing efforts.

Drug repurposing, also known as drug repositioning, denotes the new indications of existing drugs and is an alternative over the de novo drug development. Although the unknown underlying complex biology and pharmacology has challenged the drug-repurposing attempts, intelligent computer algorithms offer a strategy for detecting potential drug indications by integrating large-scale heterogeneous data (i.e., genomic, transcriptomic, phenotypic, chemical, and bioactivity) from hundreds of approved drugs. Various specially designed AI/ML models have been proposed for detecting novel drug indications. Here, we classify the ML applications for drug repositioning into the following three categories: (i) Similarity-based methods that employ different types of classifiers like logistic regression, 305 random walk, 326 and RF 310 ). We provide an in-depth discussion of these three classes of AI-based drug repositioning applications in the Supporting Information. Particularly, in early 2020, researchers at MIT published a milestone paper using a DL approach to antibiotic discovery. 153 They trained the deep GCN model based on molecular features and predicted halicin as an antibacterial molecule from the Drug-Repurposing Hub. Halicin showed a broad-spectrum activity against drug-resistant strains in mice. This is the first time an AI/ML-assisted tool was used to identify thoroughly new types of antibiotic from scratch, without the need for any previous human assumptions.

CNS diseases are a group of neurological disorders that impose a significant economic and social impact. Development of new drugs for CNS diseases poses unique challenges compared to other diseases, including the complexity of brain anatomy and function, incomplete understanding of the biology of the complex nature of CNS diseases and the presence of BBB. In this section, we present an overview of AI/ML-based approaches to meet challenges such as BBB permeability in CNS drug discovery (Figure 4 ).

Despite significant progress in our understanding of CNS diseases, the development of novel therapies for CNS diseases faces some great challenges. In addition to the difficulties in CNS target identification, designing new molecules with the ability to penetrate the BBB is also a major obstacle. The role of the BBB is to protect the brain from variations in blood composition (e.g., hormones, amino acids, and potassium) and circulating pathogens. It consists of capillary endothelial cells that are lined by the basal lamina made from structural proteins (i.e., extracellular matrix proteins collagen and laminin), pericytes, astrocytic endfeet, and microglial cells. 327 This biologic membrane allows the uptake of water, glucose, and essential amino acids, the efflux of small molecules and nonessential amino acids from the brain to the blood and the passage of some molecules by passive diffusion. 328 While negligible penetration is desirable to minimize the brain side effects for peripheral drugs, high penetration is needed for CNS-active drugs. To improve success rates in CNS drug discovery, the BBB permeability of drug candidates needs to be addressed early in the drug discovery process.

In recent years, AI-based predictive models have been proposed to minimize the number of laborious, ex- 340 and ANN. [341] [342] [343] All of these methods were developed to process physical and chemical features, which mainly include molecular weight, hydrophilicity (ClogP), lipophilicity (ClogD), topological polar surface area, acidic and basic atoms numbers, hydrogen bond donors and acceptors, wateraccessible volume, flexibility (rotatable bonds), van der Waals volume, and ionization potential.

The predictive capability of all the methods mentioned above is limited to passive diffusional uptake and predominantly relies on few molecular descriptors. However, many molecules, for example, glucose and insulin, pass BBB via complex mechanisms that involve specific drug-transporter/drug-receptor interactions. 344, 345 Hence, such mechanisms are hard to be described by simple physicochemical features of compounds. Moreover, achieving therapeutic drug concentrations in CNS may be limited by membrane transporters such as the ATP-binding cassette and efflux transporter P-glycoprotein (P-gp), 346 which mediates efflux of drugs from the BBB. Although the primary role of these efflux transporters is limiting the brain entry of neurotoxins, they also limit the entry of many therapeutics and may contribute to CNS pharmacoresistance. 347, 348 Therefore, prediction methods need to both overcome the limitations of physicochemical features and address the multiple mechanisms associated with the drugs that pass the barrier and sustain in the brain. For this purpose, Yuan et al. 333 

Schizophrenia is arguably the most puzzling of psychiatric disorders. 351 As a neurodevelopmental disorder, 352 schizophrenia shows a lifetime prevalence of 0.30%-0.66%, 353 generally beginning before age 25 years and persisting throughout life, making it one of the leading factors of global disease burden. 354 Despite more than a century of research, its complex pathophysiology remains unknown, 355 and currently, there is no effective drug for schizophrenia. Therefore, there is a need for alternative strategies to develop innovative drug treatments for schizophrenia. 356 In recent years, AI/ML has seen as a promising technology to inform schizophrenia diagnosis, 355, 357 detecting heterogeneity, [358] [359] [360] subtyping, 361, 362 and treatment.

In drug discovery studies for schizophrenia, researchers have utilized AI/ML methods with various purposes, including drug target identification, 363, 364 developing QSAR models, 365 predicting monitoring dosing compliance, 366 predicting GPCRs targeting compounds, 364 and drug repositioning. 367 Specifically, schizophrenia target genes were identified based on publicly available microarray data sets using an SVM-RFE (recursive feature elimination)-based feature selection, where the genes initially ranked by an SVM classifier and the signature was then identified by discarding the genes that were not differentially expressed. To detect optimal biomarkers of presynaptic dopamine overactivity, which may cause schizophrenia, an SVM classifier was used. 363 SVM classifiers were also used to predict QSAR models of the GABA (gamma aminobutyric acid) uptake inhibitor drugs, which can be beneficial in the treatment of schizophrenia. 365 Moreover, SVM outperformed the other ML methods in predicting the repositioning drugs for schizophrenia when trained on drug expression profiles. 367 On the contrary, for schizophrenia subtyping, an unsupervised learning approach, multi-view clustering, was employed by combining transcriptomic data with clinical phenotypes. 368 Setting a good example of the beneficiary of AI/ML in clinical drug trials, a novel AI platform AiCure 366 on mobile devices was used to assess the dosing compliance in Phase 2 clinical trial in schizophrenia patients. It, simply, confirms the medication ingestion visually by using facial recognition and computer vision.

One of the major obstacles in developing AI/ML methods for schizophrenia drug discovery is data availability. 369 Publicly available, large-scale, well-structured information on neural phenotypes, genomics, and clinical stages are greatly lacking, which arouses questions for the generalizability of AI/ML algorithms across different data sets without performance loss. However, the availability of such integrative databases can encourage the development of AI/ML-based methods to investigate personalized therapies by solving the disease heterogeneity.

Another neurodevelopmental disorder is autism spectrum disorder (ASD), which is characterized by deficits in social communication and social interaction and the presence of restricted, repetitive patterns in behaviors or interests. 370 ML methods have been utilized in ASD research for improving the diagnosis 371 and prognosis prediction. 371 Also, there are few ML applications in drug discovery for ASD. For example, ML-based cluster analysis (i.e., affinity propagation and k-medoids) of clinical data (i.e., signs and biomarkers) exhibited a good performance in drug response prediction of ASD patients. 372 Moreover, Bayesian ML models trained on HTS data revealed the potential repurposing of nicardipine or other dihydropyridine calcium channel antagonists for the treatment of Pitt Hopkins Syndrome, a rare genetic disorder that exhibits features of autistic spectrum disorders. 373 Recently, ML algorithms have been employed to predict the functional effects of variants in voltage-gated sodium and calcium ion channels, which have been associated with ASD, schizophrenia and developmental encephalopathy. 374 Being trained on sequence-and structure-based features, the ML model predicted the gain or loss of function effects of likely pathogenic missense variants in ion channels and the results were validated in exome-wide data. On the contrary, the toxic compounds may trigger the recent increases in neurodevelopmental disorders among children. 375 To identify developmental neurotoxicants, researchers developed ML algorithms to predict the neurodevelopmental toxicity of compounds. 376,377

AI/ML-based methods have been utilized in psychiatric drug discovery, especially for pharmacological decision support. 367, 378, 379 In a depression study, researchers have developed a gradient boosting machine using the predictors identified by the elastic net to predict whether a patient will achieve symptomatic remission using an antidepressant, citalopram. 380 This model was also successfully applied to an escitalopram treatment group of an independent clinical trial. 378 In the next study of Chekroud et al., 381 they clustered the symptoms using an unsupervised learning approach (hierarchical clustering) and predict the responsiveness of each cluster to the treatment of different antidepressant drugs using the same model in the previous study. To provide decision support for clinicians to select the best drugs for a given cluster of symptoms, a web-based application was VATANSEVER ET AL.

designed. This AI-based service is prospectively tested in hospital settings and thereby serve as a promising model for direct research translation. 382 On the contrary, the model of Chekroud et al. 380 has some limitations. The model only predicts whether a patient responds to a specific antidepressant without measuring the degree of antidepressant response. Since it was designed for only one antidepressant, the model is not capable of selecting the most effective drugs among various antidepressant candidates for patients. 383 To address these limitations, Chang et al. 383 developed an Antidepressant Response Prediction Network (ARPNet) model based on an NN architecture. Through the literature-based and data-driven feature selection process, ARPNet predicts the degree of antidepressant response, whether the patient will reach clinical remission from depression, and a patient's response to a combination of one or more antidepressants.

Electroencephalography (EEG) and functional magnetic resonance imaging (fMRI) data also have been employed in predicting drug responses to treatments of depression. Zhdanov et al. 384 used an SVM classifier to accurately predict the outcome of escitalopram treatment using patients' EEG data at the baseline and after the first 2 weeks of treatment. To identify a robust signature from resting-state EEG that would predict response to antidepressants, Wu et al. 385 designed an end-to-end prediction algorithm with a latent space model. They applied their algorithm, Sparse EEG Latent SpacE Regression (SELSER), to data from an imaging-coupled, placebocontrolled antidepressant study and identified an EEG signature of patient's response to antidepressant treatment (i.e., sertraline). Ichikawa et al. 386 aimed to develop a melancholic depressive disorder biomarker to extract critically important functional connections (FCs) from fMRI data. By combining two ML algorithms (i.e., L1-regularized sparse canonical correlation analysis and sparse logistic regression), they developed a classifier for melancholic depressive disorder and found out that antidepressants had a heterogeneous effect on the identified FCs of melancholic depressive disorder.

Although some of the recent AI/ML-aided tools have been rapidly translated into the clinical trials, the AI/ML methods still are not used widely in clinical practice, while AI has been employed in psychiatric research over 20 years. 387 To close the gap between research and clinic, we need to improve the validity of diagnostic and prognostic labels, representability of the features, and generalizability of models. 388 As scientists continue to work to bridge the gap between research and clinic, it will be possible to provide efficient, personalized treatments based on a patient's unique characteristics. 389 

Parkinson's disease (PD) is the second most common age-related neurodegenerative disorder, affecting over 1% of the population above the age of 60, increasing to 5% in individuals above 85 years of age. 390 PD is a prime example of a multifaceted disease, including a broad range of motor and non-motor symptoms and possible contribution of genetic and environmental risk factors. 391 Currently, there is no treatment to prevent the progressive depletion of dopaminergic neurons in the substantia nigra that underlies the movement control and cognitive loss, which is manifested with tremors and memory loss. 392, 393 Available drug treatments are based on the administration of levodopa (L-dopa) and catechol-O-methyltransferase or monoamine oxidase B inhibitors, offering only symptomatic relief to the patients. 392 In PD research, previous AI applications have focused on diagnostic biomarker discovery in cerebrospinal fluid (CSF) and blood [394] [395] [396] [397] and remote monitoring of treatment response by using electronic wearables. [398] [399] [400] [401] [402] On the contrary, recently, AI/ML has received little attention in PD drug discovery. Particularly, Shao et al. 403 initially built SVM models to quickly select the compounds containing indole-piperazine-pyrimidine scaffold among large chemical databases and subsequently identified novel compounds that simultaneously bind the two receptorsadenosine A2A receptor and dopamine D2 receptor-implicated in the PD pathophysiology. In another study, Sebastián-Pérez 404 utilized several ML techniques to infer QSAR models for the identification of putative inhibitors of LRRK2 protein, a key genetic risk factor for familiar and sporadic PD. Moreover, AI-based technologies have helped overcome the drug side effects in PD treatment. While L-dopa has remained the cornerstone of PD therapy for reducing the symptoms associated with dopamine deficiency, almost half of PD patients treated with it eventually develop levodopa-induced dyskinesia (LID), a side effect that causes abnormal involuntary movements. In a review paper, Johnston et al. 405 discussed the use of AI platforms to identify repurposing candidates for LID treatment and highlighted the potential of AI approaches by designing a drug repositioning case study. To identify novel repurposing candidates that may reduce LID, they utilized a literature mining approach based on an IBM Watson engine, where the semantic similarity and a "graph diffusion" algorithm were applied to score and rank each candidate drug.

Along with the identification of novel and repurposing candidates, AI/ML techniques have been applied to the development of in vitro and in vivo PD models for drug screening. Monzel et al. 406 

Increasing life expectancy has produced a dramatic rise in the prevalence, and thus impact, of aging-related diseases. The most prevalent neurodegenerative disease in older adults is Alzheimer's disease (AD), characterized by insidious and progressive impairment of behavioral and cognitive functions, including memory. 408 The cause of AD is still unclear; however, generally accepted neuropathological hallmarks of AD include extracellular A-beta plaques and intracellular neurofibrillary tangles, along with neuronal and synaptic loss and/or dysfunction. 409 Current drugs for AD target cholinergic and glutamatergic neurotransmission, thus improving symptoms, although they show limited benefits to most AD patients. 410 Therefore, new treatments are urgently needed to prevent or delay disease onset, slow its progression, or improve patients' symptoms. 411 However, drug development for AD has been extraordinarily difficult, with a failure rate of over 99% and no new drug approved since 2003. 411 AD drug failures are likely due to the lack of sufficient target engagement and toxicity, while drug discovery efforts mainly challenged by an incomplete understanding of AD pathogenesis, multifactorial etiology, and complex pathophysiology.

In recent years, AI/ML-based models have become popular in AD research, mostly utilizing for AD diagnosis and prognosis in dealing with electronic health records and images. 412 On the contrary, AI/ML techniques have not been widely employed in AD drug discovery. However, there have been a few studies that show the potential benefits of AI/ML applications for the discovery of AD drugs. ML approaches have assisted the target identification and characterization in AD, which is the initial phase of drug discovery. For example, Cordax 413 (https:// cordax.switchlab.org) is a novel structure-based amyloid core sequence prediction method that implements ML to detect aggregation-prone regions in proteins as well as to predict the structural topology, orientation and overall architecture of the resulting amyloid core. As an aggregation predictor, it uses structural information on amyloid cores currently available in the protein databank and translates structural compatibility and interaction energies into sequence aggregation propensity using logistic regression. Along with the characterization of amyloid fibrils, ML approaches have been utilized for identifying potential drug targets. HENA, 414 a heterogeneous network-based data set for AD, integrates distinct data types (i.e., PPI, gene coexpression, epistasis, genome-wide association study, gene expression in different brain regions, and positive selection data) through GCN to predict AD-associated genes.

VATANSEVER ET AL.

Researchers have built ML models-SVM, ANN, and RF-to predict the inhibitory effect of compounds against AD-related proteins-histone deacetylase (HDAC), 415 acetylcholinesterase (AChE), 416 and S100 calcium-binding protein A9 (S100A9), 417 respectively. Although these target-specific models were successful for predicting the bioactive compounds, a high level of reliability is necessary for prioritizing compounds that are ultimately translated into assays. To generate hyper-predictive ML models, Jamal et al. 418 have included dynamic properties of compounds and protein-ligand interactions. Extracting the dynamic descriptors from molecular dynamics simulations of caspase-8 ligand complexes to train ANN and RF models, they predicted the active compounds against caspase-8, which plays a key role in causing AD. The major challenge in developing such predictive models of inhibitor activity is the lack of data on true-negative compound-protein interactions. To address this challenge, Miyazaki et al. 419 constructed a graph CNN model to explore compounds specifically targeting proteins without using the information on the true-negative interaction and applied the model to identify inhibitors of BACE1 enzyme, a major target for AD.

Although these ML applications have advanced the discovery of single-target inhibitors, the complex nature of AD requires the discovery of multitarget drugs to address the multiple pathways contributing to disease pathogenesis. Therefore, researchers have developed ML algorithms for predicting multitarget-directed compounds against AD. Kleandrova et al. 420 however, nonspecificity and nonselectivity are the major problems of current HDAC inhibitors. Therefore, Gupta et al. 422 combined VS and ML to classify the HDAC inhibitors and identified a novel compound that potentially inhibits all isoforms of class I and class IIb HDAC for AD therapy. In addition to these, Fang et al. 423 built 100 binary classifiers based on the naive Bayesian and RP algorithms to predict active small molecules against 25 key targets toward AD. Experimental validation of the predicted molecules yielded a compound that is a dual cholinesterase inhibitor and H3R antagonist. In their following study, 424 the system has been updated by assembling 204 binary classifiers towards 54 critical targets related to AD and the information of the classifiers was shared in a web server named AlzhCPI. Utilizing this classifier system, another group of researchers 425 has identified multiple targets of a traditional Chinese herbal medicine formula, Naodesheng, for application to AD. Natural products has continued to generate an increased interest as a mean of discovering novel bioactive compounds against AD. Grisoni et al. 426 proposed a VS protocol based on ML models to explore the bioactive synthetic mimetics of the natural product galantamine, which is the first natural product-based AD drug approved by the FDA in 2001. 427 Using an ML-based selection and target profiling program, they identified galantamine-mimetic small molecules with multitarget activity on enzymes and receptors related to AD.

Besides the predictions of multitarget compounds based on their bioactivity against known drug targets in AD, Jamal et al. 428 predicted small molecules that show a high binding affinity for ML-inferred possible therapeutic targets. Unlike previous studies that target known AD-related proteins, they initially predicted the probable AD-associated genes using ML classifiers that are trained on network, sequence and functional features. Then, they used a conventional VS tool to select the compounds that have high affinity for the majority of the predicted targets.

In addition to applications for identifying small molecules towards therapeutic targets for AD, ML techniques also have been utilized in drug repositioning efforts. For example, telmisartan has been associated with AD by a network-based classification model. 310 AI/ML approaches have also been applied to drug response studies to treat AD patients in a more precise, personalized way. Hampel et al. 429 has built an AI/ML-based precision medicine framework for identifying the genomic biomarkers of response to AD therapy. Specifically, they studied blarcamesine (ANAVEX2-73), a selective sigma-1 receptor agonist, in a Phase 2a trial, where they obtained the patients' whole-exome and transcriptome data and recorded the measures of safety, clinical features, pharmacokinetics, and efficacy. They analyzed the relationship between the patient data and efficacy outcome measures using unsupervised formal concept analysis, which ultimately identified the biomarkers of drug response. On the contrary, Lu et al. 430 evaluated the therapeutic effects of Dengzhan Shengmai formula, a traditional Chinese medicine, on AD patients by analyzing the diffusion tensor imaging data with ML. Their ML classifier revealed significant white-matter network alterations after treatment.

The CNC drugs include general anesthetics and the analgesics, as well. In the past few years, we have witnessed the widespread use of autonomous and AI-based recommender systems in therapeutic decision making in anesthesia and pain management. Especially, pharmacological robots have become an integral part of the anesthesia field, offering a personalized anesthetic drug dosage for maintaining patient homeostasis during general anesthesia and sedation. 431 These robots use complex ML algorithms based on patient data (e.g., EEG monitor, blood pressure, heart rate, etc.) and pharmacokinetic features of drugs to provide the optimal drug dosage. The role of pharmacological robots and even more intelligent autonomous systems (i.e., cognitive robot, which can recognize crucial clinical state that requires human intervention) in the anesthesia field has been comprehensively overviewed by Cédrick et al. 432 Besides the robotic systems, ML applications assisted the clinicians 433 to monitor the drug-specific anesthetic states [434] [435] [436] and predict the adverse outcomes in anesthesia patients. [437] [438] [439] Similar to the anesthesia field, AI models have mainly utilized for clinical decision support in pain management.

With the increasing amount of data collected by state-of-the-art monitoring sensors and the Internet of Things, the AI-assisted patient-controlled analgesia has a great potential for personalized pain therapy. 440 The other clinical applications of AI systems in pain management include prediction of pain severity/modality and analgesic requirements, [441] [442] [443] individualized medicine decision support in analgesic treatment, 444, 445 prediction of the effectiveness of the analgesics, 446, 447 and prediction of medication overuse. [448] [449] [450] Besides the clinical applications, researchers have employed ML methods at the early stages of analgesic discovery, such as identifying novel genes and pathways associated with acute and chronic pain 451 and predicting inhibitors of a drug target for pain (i.e., NaV1.7 sodium channel). 452 To facilitate the prediction of novel multi-target analgesics or drug combinations for pain treatment, researchers have established a comprehensive pain-domain-specific chemogenomics knowledgebase that includes the analgesics in current use, pain-related targets with all available 3D structures, and the compounds reported for these target proteins. 453 

Given the complexity of neurological disorders, CNS drug development is still a long, expensive, inefficient, and challenging process with a low rate of new successful therapeutic discovery. To overcome the challenges of CNS drug discovery, researchers have utilized AI/ML-based methods, which have played a promising role in all stages of drug discovery for a variety of diseases (Table 2 ). In general, AI/ML practices in pharmaceutical development have aroused great interest among researchers working in academia and industry. The number of start-ups in this area has grown rapidly and reached 230 by June 2020. 454 Also, many pharmaceutical companies have invested in internal AI-based research programs as well as in collaboration with AI start-ups and academic institutions. 455 of therapeutic development, the use of AI technologies to improve CNS drug discovery is still at an early stage. Below, we discuss the limitations as well as the future directions to guide further advancement in this evolving field.

The main bottleneck in applying AI/ML into CNS drug discovery is the lack of high-quality, well-annotated data sets to train effective algorithms. The data collected in the public databases are generally generated by different biological assays, methods, or conditions, which are not comparable. Also, multiple data sets on the same subject may contradict each other. Therefore, filtering the raw inputs to obtain high-quality data is a crucial step before performing specific AI/ML tasks.

The "black box" nature of most next-generation AI architectures an additional challenge in CNS drug discovery. 

Abbreviations: AD, Alzheimer's disease; AI, artificial intelligence; CNS, central nervous system; ML, machine learning; PD, Parkinson's disease.

learns from one task and applies it to the other task, can offer a solution. However, in the long term, the most promising solution to overcome data scarcity would be for the scientific community to share their data. Such largescale sharing of data would make significant improvement in the CNS drug discovery process, with advances in hardware that lead to faster machines such as quantum computers in the near future.

A particular limitation for the AI/ML applications in CNS drug discovery is the unknown pathophysiology for many nervous system disorders, which makes target identification very challenging. To explore the complex disease mechanisms and define the right biological targets, we need better AI/ML tools that can pull information out of the data sets generated across the different biological layers (e.g., transcriptomics, proteomics, and metabolomics). Here, capsule networks, 457 a next-generation AI architecture where CNNs are encapsulated in an interconnected module, can provide a solution. As the first application of capsule networks to drug discovery, capsule networks showed excellent performance to predict the cardiotoxicity of compounds, which highlights their unique potential in drug discovery efforts. 226 Because of the modular representation of the CNNs, capsule networks can learn from heterogeneous data sets by preserving the hierarchical aspects of the data itself. Considering the highly modular nature of CNS data sets with specified layers of genes, proteins, metabolites, capsule networks can analyze the changes in the functional organization and interplay of these layers upon the diseases.

Another critical issue in the application of AI/ML models into CNS drug discovery is the integration of different data types, including genotypic data from patients, multiomics data from drug treatments, and chemical data from bioactivity and toxicity assays. Considering the availability of various databases that include biological, structural, and chemical information, how to integrate these data to generate AI/ML models becomes a critical question in CNS drug discovery applications. Multitask learning, learning of different tasks jointly, can be suitable for these types of applications. Multitask NNs are capable of integrating data from many distinct sources. For example, a multitask architecture can predict the effects of a drug and its BBB permeability at the same time by learning from multiomics data sets, physicochemical properties, HTS, and bioactivity assays.

In recent years, we have seen the emergence of novel neuroimaging techniques such as pharmacological functional magnetic resonance imaging (pharmacoMRI) and pharmacologically induced functional ultrasound (pharmaco-fUS), which provide in vivo functional data of specific effects of drugs on the brain. Although phar-macoMRI continues to play a useful role in neuropharmacology studies as a well-established technique, 458-461 a variety of challenges (i.e., low sensitivity, the requirement for anesthesia, and blood oxygenation-level dependent imaging) limit the preclinical use of it. A newer tool, pharmaco-fUS enables brain activity imaging through the local monitoring of cerebral blood volume dynamics at an unprecedented spatiotemporal resolution without the bias of anesthesia. 462, 463 Recent studies demonstrated fUS imaging's potential to characterize dynamic profiles of CNS drugs, including a drug combination of donepezil plus mefloquine for AD 464 and atomoxetine for attention-deficit/hyperactivity disorder. 465 Moreover, Rabut et al. 466 adapted ML to analyze the rich data content provided by fUS connectivity imaging. Their ML model identified the "fingerprint" of drug-induced brain connectivity changes in awake mice for scopolamine, a major preclinical drug to model AD. As evident from the previous applications, AI/ML methods hold the promise of characterization of treatment effects from novel neuroimaging data sets and thereby improving our understanding of the mechanism of action of drugs in the brain. Getting drugs across the BBB is an essential step to developing successful therapies to treat CNS disorders. However, it is often overlooked that BBB is not only a physical barrier for drug delivery to the CNS but also a complex, dynamic interface that might be affected by diseases. CNS disorders may result in dysfunction of BBB, such as its disruption or dysfunctions related to BBB transporters. To date, AI/ML-based predictive algorithms have assumed that BBB is a static entity by neglecting the effects of CNS pathologies on it. Therefore, a prediction model for BBB penetrance that trained on data from non-CNS diseases may not work for a CNS disease. To develop better prediction models for BBB permeability, we need to take into account disease-related changes in the barrier. This also provides many unique opportunities for developing disease-specific AI/ML tools in CNS drug discovery.

It is important to highlight that CNS drug discovery has a nondeterministic nature, where the neurological targets involve different pathways and their biological consequences are not the sums of the single functions, most drugs have diverse activities through multiple biological targets, and drug response is dependent on a range of factors (i.e., patient's genetic profile and drug's membrane permeability). Moreover, physiologic events are highly context-specific: A receptor interaction may take place in the liver but not in the brain. AI/ML systems often fail to pick up such context-specific nonlinear relationships and many other unknown contributing factors. As a result of incomplete domain representation, partial predictability in CNS drug discovery is inevitable. For example, an AI/ML algorithm may predict drug targets that neuroscientists know will likely have significant side effects in the brain or generate unsynthesizable molecules. Here, we need the human refinement process and hypothesis-driven approach 467 to address many of these challenges to achieve better performance. Knowledge acquisition from the human experts to the AI systems can help the AI/ML system learn and thereby guarantee the best scientific results.

In consequence, this mixture of machine and mind 468 will improve decision making as an essential component of the CNS drug discovery process.

Although AI/ML algorithms have already revolutionized other fields, the adoption of them to drug discovery is still at an eraly stage. Initially, AI/ML algorithms have been developed and practically used for certain areas such as image recognition, gaming, and internet search. Inspired by the successful applications in other disciplines, scientists have applied AI/ML algorithms to pharmaceutical research. And yet, we do not have any AI/ML algorithm that is developed specifically for a drug discovery problem. But this means that there should be many opportunities to develop innovative and novel algorithms in the field of therapeutic discovery. In this way, AI/ML methods will play an increasingly important role in not just the field of general pharmaceutical research but also CNS drug discovery.

In conclusion, we extensively review the latest AI/ML-assisted drug discovery applications for the therapy of CNS diseases. These applications have been overgrowing in the past couple of years, fueled by the unprecedented success of AI/ML-based approaches in different fields of science and technology. We envision that in the future, AI/ML will play more and more critical roles in CNS drug discovery towards personalized medicine, especially in the following areas: (1) patient subtyping, (2) identification of key disease drivers, (3) prediction of cell type-specific drug response, (4) autonomous design of novel drugs, and (5) disease-specific BBB permeability testing. Today there are structural constraints in data and algorithms that are limiting the role of AI/ML. Nonetheless, in the long run, ongoing and emerging developments in AI/ML approaches to neuropharmacology will enable us to develop more effective drugs for CNS diseases. 

Drug discovery and development: role of basic biological research

Activity, assay and target data curation and quality in the ChEMBL database

The Human Genome Project, and recent advances in personalized genomics

Applications of deep-learning in exploiting large-scale and heterogeneous compound data in industrial pharmaceutical research

medical-devices/software-medical-device-samd/artificial-intelligence-and-machine-learningsoftware-medical-device

Books-handbook of molecular descriptors

ChemDes: an integrated web-based platform for molecular descriptor and fingerprint computation

Large-scale prediction of human kinase-inhibitor interactions using protein sequences and molecular topological structures

Prediction and validation of enzyme and transporter off-targets for metformin

An overview of molecular fingerprint similarity search in virtual screening

Molecular fingerprint similarity search in virtual screening

Similarity-based virtual screening using 2D fingerprints

Fingerprint design and engineering strategies: rationalizing and improving similarity search performance

Predicted biological activity of purchasable chemical space

A simple representation of three-dimensional molecular structure

Current trends in ligand-based virtual screening: molecular representations, data mining methods, new application areas, and performance evaluation

A rotation-translation invariant molecular descriptor of partial charges and its use in ligand-based virtual screening

Electrotopological state atom (E-state) index in drug design, QSAR, property prediction and toxicity assessment

Prediction of aqueous solubility of druglike organic compounds using partial least squares, back-propagation network and support vector machine

Atom type preferences, structural diversity, and property profiles of known drugs, leads, and nondrugs: a comparative assessment

Predictions of the ADMET properties of candidate drug molecules utilizing different QSAR/QSPR modelling approaches

Correction to "admetSAR: a comprehensive source and free tool for assessment of chemical ADMET properties

Applying machine learning techniques for ADME-Tox prediction: a review

Learning molecular representations for medicinal chemistry

Opportunities and challenges in phenotypic drug discovery: an industry perspective

Drug targets, target identification, validation, and screening

What makes a good drug target? Drug Discovery Today

What makes a good drug target? Drug Discovery Today

Target discovery

Gene expression patterns of breast carcinomas distinguish tumor subclasses with clinical implications

Integrative clustering of multiple genomic data types using a joint latent variable model with application to breast and lung cancer subtype analysis

Patient-specific data fusion defines prognostic cancer subtypes

Deep learning-based multi-omics data integration reveals two prognostic subtypes in high-risk neuroblastoma

DeepCC: a novel deep learning-based framework for cancer molecular subtype classification

DIABLO: an integrative approach for identifying key molecular drivers from multi-omics assays

Regularized generalized canonical correlation analysis

Multi-omics factor analysis-a framework for unsupervised integration of multi-omics data sets

Joint and individual variation explained (Jive) for integrated analysis of multiple data types

CHASM and SNVBox: toolkit for detecting biologically important single nucleotide mutations in cancer

Identifying disease genes using machine learning and gene functional similarities, assessed through Gene Ontology

CanDrA: cancer-specific driver missense mutation annotation with optimized features

Machine learning-assisted network inference approach to identify a new class of genes that coordinate the functionality of cancer networks

G2P: using machine learning to understand and predict genes causing rare neurological disorders. bioRxiv

DriverML: a machine learning algorithm for identifying driver genes in cancer sequencing studies

DriverDBv2: a database for human cancer driver gene research

deepDriver: predicting cancer driver genes based on somatic mutations using deep convolutional neural networks

Predicting Parkinson's disease genes based on Node2vec and autoencoder

LOTUS: a single-and multitask machine learning algorithm for the prediction of cancer driver genes

Data integration and predictive modeling methods for multi-omics datasets

Extraction of relations between genes and diseases from text and large-scale data analysis: implications for translational research

MicroRNAs: genomics, biogenesis, mechanism, and function

miRNA targeting drugs: the next blockbusters?

Prioritizing candidate disease miRNAs by topological features in the miRNA targetdysregulated network: case study of prostate cancer

Prediction of microRNAs associated with human diseases based on weighted k most similar neighbors

Prediction of potential disease-associated microRNAs by using neural networks

MLMDA: a machine learning approach to predict and validate microRNA-disease associations by integrating of heterogenous information sources

Mechanisms of alternative pre-messenger RNA splicing

AVISPA: a web tool for the prediction and analysis of alternative splicing

Deep learning of the tissue-regulated splicing code

Integrative deep models for alternative splicing

Predicting alternatively spliced exons using semi-supervised learning

RNA splicing. The human splicing code reveals new insights into the genetic determinants of disease

Recent approaches to the prioritization of candidate disease genes

A systematic approach to identify novel cancer drug targets using machine learning, inhibitor design and high-throughput screening

An extensive analysis of disease-gene associations using network integration and fast kernel-based gene prioritization methods

In silico prediction of novel therapeutic targets using gene-disease association data

DigSee: disease gene search engine with evidence sentences (version cancer)

Genome-wide prediction and prioritization of human aging genes by data fusion: a machine learning approach

Predicting protein structural features with Artificial neural networks

Machine learning methods for protein structure prediction

Recent developments in deep learning applied to protein structure prediction

Neural networks for protein structure and function prediction and dynamic analysis

Deep learning methods in protein structure prediction

AlphaFold at CASP13

Protein structure prediction beyond AlphaFold

Science unfolding in time: a situated protein folding landscape in retrospect. OSF Preprints

Deep-learning contact-map guided protein structure prediction in CASP13

Accurate de novo prediction of protein contact map by ultra-deep learning model

Improved protein structure prediction using predicted interresidue orientations

Improved fragment sampling for ab initio protein structure prediction using deep neural networks

ProSPr: democratized implementation of alphafold protein distance prediction network

Redefining the protein kinase conformational space with machine learning

Extraction of protein dynamics information hidden in Cryo-EM map using deep learning

Protein function in precision medicine: deep understanding with machine learning

Determining druggability

On the nature of cavities on protein surfaces: application to the identification of drug-binding sites

Prediction of potential drug targets based on simple sequence properties

Properties and identification of human protein drug targets

A machine learning approach for genome-wide prediction of morbid and druggable human genes based on systems-level data

A novel framework for the identification of drug target proteins: combining stacked auto-encoders with a biased support vector machine

Druggability assessment in TRAPP using machine learning approaches

Machine learning prediction of oncology drug targets based on protein and network properties

Prediction of druggable proteins using machine learning and systems biology: a mini-review

Revealing the macromolecular targets of complex natural products

Identifying the macromolecular targets of de novo-designed chemical entities through self-organizing map consensus

Revealing the macromolecular targets of fragment-like natural products

Deorphaning the macromolecular targets of the natural anticancer compound doliculide

Prediction of biological targets for compounds using multiple-category Bayesian models trained on chemogenomics databases

Utilizing random Forest QSAR models with optimized parameters for target identification and its application to target-fishing server

A Bayesian machine learning approach for drug target identification using diverse data types

Machine learning models for drug-target interactions: current knowledge and future directions

Drug-target interactions: prediction methods and applications

DrugE-Rank: improving drug-target interaction prediction of new candidate drugs or targets by ensemble learning to rank

Computational prediction of drug-target interactions via ensemble learning

Drug-target interaction prediction using ensemble learning and dimensionality reduction

iDTI-ESBoost: identification of drug target interaction using evolutionary and structural features with boosting

Drug-target interaction prediction with tree-ensemble learning and output space reconstruction

BE-DTI': ensemble framework for drug target interaction prediction using dimensionality reduction and active learning

NeoDTI: neural integration of neighbor information from a heterogeneous network for discovering new drug-target interactions

Drug-target interaction prediction: end-to-end deep learning approach

Combining drug and gene similarity measures for drug-target elucidation

Kernel-based data fusion improves the drug-protein interaction prediction

Drug repositioning by kernel-based integration of molecular structure, molecular activity, and phenotype data

Prediction of drug-target interaction networks from the integration of chemical and genomic spaces

A multiple kernel learning algorithm for drug-target interaction prediction

Quo vadis, virtual screening? A comprehensive survey of prospective applications

What has virtual screening ever done for drug discovery? Expert Opin Drug Discovery

Structure-based virtual screening for drug discovery: principles, applications and recent advances

Structure-based virtual screening of chemical libraries for drug discovery

Recent advances in ligand-based drug design: relevance and utility of the conformationally sampled pharmacophore approach

Molecular similarity and diversity in chemoinformatics: from theory to applications

The chemical basis of pharmacology

A machine learning approach to predicting protein-ligand binding affinity with applications to molecular docking

SFCscore(RF): a random forest-based scoring function for improved affinity prediction of protein-ligand complexes

Binding affinity prediction for protein-ligand complexes based on beta contacts and B factor

istar: a web platform for large-scale protein-ligand docking

ID-Score: a new empirical scoring function based on a comprehensive set of descriptors related to protein-ligand interactions

Machine learning scoring functions based on random forest and support vector regression

NNScore: a neural-network-based scoring function for the characterization of proteinligand complexes

CScore: a simple yet effective scoring function for protein-ligand binding affinity prediction using modified CMAC learning architecture

TopologyNet: topology based deep convolutional and multi-task neural networks for biomolecular property predictions

Protein-ligand scoring with convolutional neural networks

Development and evaluation of a deep learning model for protein-ligand binding affinity prediction

BgN-Score and BsN-Score: bagging and boosting based ensemble neural networks scoring functions for accurate binding affinity prediction of protein-ligand complexes

Machine-learning scoring functions to improve structure-based binding affinity prediction and virtual screening

Concepts of artificial intelligence for computer-assisted drug discovery

Improving the virtual screening ability of target-specific scoring functions using deep learning methods

Target-specific support vector machine scoring in structure-based virtual screening: computational validation, in vitro testing in kinases, and effects on lung cancer cell proliferation

Structure-based target-specific screening leads to smallmolecule CaMKII inhibitors

Constructing and validating high-performance MIEC-SVM models in virtual screening for kinases: a better way for actives discovery

Machine learning classification models to improve the docking-based screening: a case of PI3K-tankyrase inhibitors

Protein-ligand empirical interaction components for virtual screening

Use of artificial intelligence in the design of small peptide antibiotics effective against a broad spectrum of highly antibiotic-resistant superbugs

A machine learning-based method to improve docking scoring functions and its application to drug repurposing

Boosting docking-based virtual screening with deep learning

Prediction of N-methyl-D-aspartate receptor GluN1-ligand binding affinity by a novel SVM-Pose/SVM-Score combinatorial ensemble docking scheme

Improving Autodock Vina using Random Forest: the growing accuracy of binding affinity prediction by the effective exploitation of larger data sets

Improvement of virtual screening results by docking data feature analysis

Towards improving compound selection in structure-based virtual screening. Drug Discovery Today

Concepts and Applications of Molecular Similarity

Deep learning and virtual drug screening

Machine learning in virtual screening

A deep learning approach to antibiotic discovery

MoleculeNet: a benchmark for molecular machine learning

Improvement in ADMET prediction with multitask deep featurization

Quantitative structure-activity relationship: promising advances in drug discovery platforms

An in silico platform for predicting, screening and designing of antihypertensive peptides

QSAR Accelerated discovery of potent ice recrystallization inhibitors

Random forest: a classification and regression tool for compound classification and QSAR modeling

QSAR modeling and prediction of drug-drug interactions

Improving virtual screening predictive accuracy of Human kallikrein 5 inhibitors using machine learning models

Developing an in silico pipeline for faster drug candidate discovery: virtual high throughput screening with the signature molecular descriptor using support vector machine models

Identifying novel factor XIIa inhibitors with PCA-GA-SVM developed vHTS models

Classification of kinase inhibitors using a Bayesian model

Screening for dihydrofolate reductase inhibitors using MOLPRINT 2D, a fast fragment-based method employing the Naive Bayesian classifier: limitations of the descriptor and the importance of balanced chemistry in training and test sets

Global Bayesian models for the prioritization of antitubercular agents

Bayesian models leveraging bioactivity and cytotoxicity information for drug discovery

Combinatorial library enumeration and lead hopping using comparative interaction fingerprint analysis and classical 2D QSAR methods for seeking novel GABA(A) alpha(3) modulators

Novel Bayesian classification models for predicting compounds blocking hERG potassium channels

QSAR classification model for antibacterial compounds and its use in virtual screening

Virtual screening of CB(2) receptor agonists from bayesian network and highthroughput docking: structural insights into agonist-modulated GPCR features

Predicting rat and human pregnane X receptor activators using Bayesian classification models

Absorption, distribution, metabolism, excretion, and toxicity evaluation in drug discovery. 14. Prediction of human pregnane X receptor activators by using naive bayesian classification technique

Artificial neural networks and linear discriminant analysis: a valuable combination in the selection of new antibacterial compounds

Neural networks: accurate nonlinear QSAR model for HEPT derivatives

Computer-aided design of novel antibacterial 3-hydroxypyridine-4-ones: application of QSAR methods based on the MOLMAP approach

Identification of novel antibacterial peptides by chemoinformatics and machine learning

Connecting peptide physicochemical and antimicrobial properties by a rational prediction model

Artificial neural network modeling of antimycobacterial chemical space to introduce efficient descriptors employed for drug design

Diverse classification models for anti-hepatitis C virus activity of thiourea derivatives

A neural networks-based drug discovery approach and its application for designing aldose reductase inhibitors

Artificial neural network-based drug design for diabetes mellitus using flavonoids

Molecular fingerprint-based artificial neural networks QSAR for ligand biological activity predictions

Ligand-based virtual screen for the discovery of novel M5 inhibitor chemotypes

Deep neural nets as a method for quantitative structure-activity relationships

Profile-QSAR 2.0: kinase virtual screening accuracy comparable to fourconcentration IC50s for realistically novel compounds

A random forest model to predict the activity of a large set of soluble epoxide hydrolase inhibitors solely based on a set of simple fragmental descriptors

Construction of quantitative structure activity relationship (QSAR) models to predict potency of structurally diverse Janus kinase 2 inhibitors

Comparison of the predictive performance and interpretability of Random Forest and Linear Models on Benchmark Data Sets

New insights toward the discovery of antibacterial agents: multitasking QSBER model for the simultaneous prediction of anti-tuberculosis activity and toxicological profiles of drugs

ANN multiplexing model of drugs effect on macrophages; theoretical and flow cytometry study on the cytotoxicity of the anti-microbial drug G1 in spleen

Model for high-throughput screening of drug immunotoxicity-study of the anti-microbial G1 over peritoneal macrophages using flow cytometry

Simultaneous modeling of antimycobacterial activities and ADMET profiles: a chemoinformatic approach to medicinal chemistry

Simultaneous virtual prediction of anti-Escherichia coli activities and ADMET profiles: a chemoinformatic complementary approach for high-throughput screening

Enabling the discovery and virtual screening of potent and safe antimicrobial peptides. Simultaneous prediction of antibacterial activity and cytotoxicity

Speeding up early drug discovery in antiviral research: a fragmentbased in silico approach for the design of virtual anti-hepatitis C leads

Alignment-free prediction of a drug-target complex network based on parameters of drug connectivity and protein sequence of receptors

Chemoinformatics in multi-target drug discovery for anticancer therapy: in silico design of potent and versatile anti-brain tumor agents

Fragment-based in silico modeling of multi-target inhibitors against breast cancerrelated proteins

Salakhutdinov R Multi-task neural networks for QSAR predictions

Multi-targeting drug community challenge

Crowdsourced mapping extends the target space of kinase inhibitors

Demystifying multitask deep neural networks for quantitative structureactivity relationships

Novel consensus architecture to improve performance of large-scale multitask deep learning QSAR Models

Comprehensive ensemble in QSAR prediction for drug discovery

Binary classification of aqueous solubility using support vector machines with reduction and recombination feature selection

Support vector machines for the estimation of aqueous solubility

Deep architectures and deep learning in chemoinformatics: the prediction of aqueous solubility for drug-like molecules

Convolutional embedding of attributed molecular graphs for physical property prediction

Can human experts predict solubility better than computers?

Prediction of permeability of drug-like compounds across polydimethylsiloxane membranes by machine learning methods

In silico prediction of PAMPA effective permeability using a two-QSAR approach

Prediction of membrane permeation of drug molecules by combining an implicit membrane model with machine learning

Computation of octanol-water partition coefficients by guiding an additive model with knowledge

Three-class classification models of logS and logP derived by using GA-CG-SVM approach

SVM approach for predicting LogP

Application of ALOGPS 2.1 to predict log D distribution coefficient for Pfizer proprietary compounds

Prediction of n-octanol/water partition coefficients from PHYSPROP database using artificial neural networks and E-state indices

4 modeling using Bayesian regularized neural networks. Assessment and correction of the errors of prediction

Can the pharmaceutical industry reduce attrition rates?

Opportunities and challenges using artificial intelligence in ADME/Tox

Estimation of ADME properties with substructure pattern recognition

In silico prediction of chemical toxicity for drug design using machine learning methods and structural alerts

Tox21Challenge to build predictive models of nuclear receptor and stress response pathways as mediated by exposure to environmental chemicals and drugs

Practical applications of deep learning to impute heterogeneous drug discovery data

Capsule networks showed excellent performance in the classification of hERG blockers/nonblockers

Recent progresses in the exploration of machine learning methods as in-silico ADME prediction tools

Opportunities and challenges using artificial intelligence in ADME/Tox

In silico prediction of physicochemical properties of environmental chemicals using molecular fingerprints and machine learning

In silico ADME-Tox modeling: progress and prospects

Predictive multitask deep neural network models for ADME-Tox properties: learning from large data sets

Bayer's in silico ADMET platform: a journey of machine learning over the past two decades

Future de novo drug design

Current and future roles of artificial intelligence in medicinal chemistry synthesis

Generative models for artificially-intelligent molecular design

The cornucopia of meaningful leads: applying deep adversarial autoencoders for new molecule development in oncology

druGAN: an advanced generative adversarial autoencoder model for de novo generation of new molecules with desired molecular properties in silico

Entangled conditional adversarial autoencoder for de novo drug discovery

Automatic chemical design using a data-driven continuous representation of molecules

Molecular generative model based on conditional variational autoencoder for de novo molecular design

Molecular de-novo design through deep reinforcement learning

Generating focused molecule libraries for drug discovery with recurrent neural networks

Chemical space mimicry for drug discovery

Tuning artificial intelligence on the de novo design of natural-productinspired retinoid X receptor modulators

Deep reinforcement learning for de novo drug design

Deep reinforcement learning that matters

Reinforced adversarial neural computer for de novo molecular design

Optimizing distributions over molecular space. An objective-reinforced generative adversarial network for inverse-design chemistry (ORGANIC

Generative recurrent networks for de novo drug design

Drug analogs from fragment-based long short-term memory generative neural networks

Retrosynthetic reaction prediction using neural sequence-to-sequence models

DeepSMILES: an adaptation of SMILES for use in machine-learning of chemical structures. ChemRxiv

Population-based de novo molecule generation, using grammatical evolution

Discovering molecular functional groups using graph convolutional neural networks. ArXiv

Graph convolutional policy network for goal-directed molecular graph generation

MolGAN: an implicit generative model for small molecular graphs

Mol-CycleGAN: a generative model for molecular optimization

Constrained graph variational autoencoders for molecule design

Shape-based generative modeling for de novo drug design

Automated de novo molecular design by hybrid machine intelligence and rule-driven chemical synthesis

De novo generation of hit-like molecules from gene expression signatures using artificial intelligence

Deep learning enables rapid identification of potent DDR1 kinase inhibitors. Overview of attention for article published in Nature Biotechnology

Deep learning enables rapid identification of potent DDR1 kinase inhibitors

What makes a kinase promiscuous for inhibitors?

Assessing the impact of generative AI on medicinal chemistry

VAE: improving drug response prediction via modeling of drug perturbation effects

Clinical drug response can be predicted using baseline gene expression levels and in vitro drug sensitivity in cell lines

Evaluating the molecule-based prediction of clinical drug responses in cancer

A landscape of pharmacogenomic interactions in cancer

Drug response prediction by ensemble learning and druginduced gene expression signatures

PLATYPUS: a multiple-view learning predictive framework for cancer drug sensitivity prediction

MOLI: multi-omics late integration with deep neural networks for drug response prediction

DeepProfile: deep learning of cancer molecular profiles for precision medicine. bioRxiv

DeepDSC: a deep learning method to predict drug sensitivity of cancer cell lines

Precision oncology beyond targeted therapy: combining omics data with machine learning matches the majority of cancer cells to effective therapeutics

Predicting drug-induced transcriptome responses of a wide range of human cell lines by a novel tensor-train decomposition algorithm

Prediction of anti-cancer drug response by kernelized multi-task learning

Systematic identification of feature combinations for predicting drug response with Bayesian multi-view multi-task linear regression

The search for cytotoxic synergy between anticancer agents: a case of Dorothy and the ruby slippers?

An Introduction to terminology and methodology of chemical synergyperspectives from across disciplines

Detecting the potential pharmacological synergy of drug combination by viability assays in vitro

DeepTox: toxicity prediction using deep learning

Haibe-Kains B. Predictive approaches for drug combination discovery in cancer

DeepSynergy: predicting anti-cancer drug synergy with deep learning

Drug combination sensitivity scoring facilitates the discovery of synergistic and efficacious drug combinations in cancer

Predicting synergism of cancer drug combinations using NCI-ALMANAC data

In-silico prediction of synergistic anti-cancer drug combinations using multi-omics data

Prediction of drug synergy in cancer using ensemble-based machine learning techniques

Antidepressant response in patients with major depression exposed to NSAIDs: a pharmacovigilance study

Using machine learning to predict synergistic antimalarial compound combinations with novel structures

Prediction of antibiotic interactions using descriptors derived from molecular structure

Prediction of drug combination effects with a minimal set of experiments

Ensemble prediction of synergistic drug combinations incorporating biological, chemical, pharmacological, and network knowledge

Synergistic drug combinations prediction by integrating pharmacological data

Drug-drug interaction through molecular structure similarity analysis

Modeling polypharmacy side effects with graph convolutional networks

Deep learning improves prediction of drug-drug and drug-food interactions

Predicting adverse drug reactions of combined medication from heterogeneous pharmacologic databases

Prediction of drug adverse events using deep learning in pharmaceutical discovery

Predicting adverse drug reactions of two-drug combinations using structural and transcriptomic drug representations to train an artificial neural network

The $2.6 billion pill-methodologic and policy considerations

FDA drug approvals

1854-1915): man with the magic bullet

Putting the patient back together-social medicine, network medicine, and the limits of reductionism

PREDICT: a method for inferring novel drug indications with application to personalized medicine

Similarity-based prediction for anatomical therapeutic chemical classification of drugs by integrating multiple data sources

Drug repositioning: a machine-learning approach through data integration

DrugR+: a comprehensive relational database for drug repurposing, combination therapy, and replacement therapy

Chemogenomic analysis of the druggable kinome and its application to repositioning and lead identification studies

A network-based classification model for deriving novel drug-disease associations and assessing their molecular actions

Old drug repositioning and new drug discovery through similarity learning from drug-target joint feature spaces

Computational Drug Repositioning by Ranking and Integrating Multiple Data Sources

Inferring drug-related diseases based on convolutional neural network and gated recurrent unit

Deep learning applications for predicting pharmacological properties of drugs and drug repurposing using transcriptomic data

Prediction of drug-disease associations based on ensemble meta paths and singular value decomposition

Relation path feature embedding based convolutional neural network method for drug discovery

Predicting drug-disease associations via using Gaussian interaction profile and Kernel-based autoencoder

Multi-channel PINN: investigating scalable and transferable neural networks for drug discovery

Predicting associations among drugs, targets and diseases by tensor decomposition for drug repositioning

Drug-Target Interaction Prediction in Drug Repositioning Based on Deep Semi-Supervised Learning

Analysis of defective pathways and drug repositioning in Multiple Sclerosis via machine learning approaches

Semi-supervised drug-protein interaction prediction from heterogeneous biological spaces

A semi-supervised method for drug-target interaction prediction with consistency in networks

Drug repurposing with network reinforcement

Prediction of drug-target interaction by label propagation with mutual interaction information derived from heterogeneous network

deepDR: a network-based deep learning approach to in silico drug repositioning

Toxoplasma gondii and the blood-brain barrier

The blood-brain barrier: an overview: structure, regulation, and clinical implications

Predicting CNS permeability of drug molecules: comparison of neural network and support vector machine algorithms

A Bayesian approach to in silico blood-brain barrier penetration modeling

Finding needles in a haystack: determining key molecular descriptors associated with the blood-brain barrier entry of chemical compounds using machine learning

Quantitative structure-activity relationship prediction of blood-tobrain partitioning behavior using support vector machine

Improved prediction of blood-brain barrier permeability through machine learning with combined use of molecular property-based descriptors and fingerprints

A recursive-partitioning model for blood-brain barrier permeation

Predicting penetration across the blood-brain barrier from simple descriptors and fragmentation schemes

Gaussian processes: a method for automatic QSAR modeling of ADME properties

CNS permeability of drugs predicted by a decision tree

QSAR modeling of the blood-brain barrier permeability for diverse organic compounds

Prediction of passive blood-brain partitioning: straightforward and effective classification models based on in silico derived physicochemical descriptors

In silico prediction of blood-brain barrier permeability of compounds by machine learning and resampling methods

In silico prediction of blood brain barrier permeability: an Artificial Neural Network model

Role of breast cancer resistance protein (BCRP) as active efflux transporter on bloodbrain barrier (BBB) permeability

The role of multidrug resistance protein (MRP-1) as an active efflux transporter on blood-brain barrier (BBB) permeability

Structure-based ligand discovery for the large-neutral amino acid transporter 1, LAT-1

Predicting efflux ratios and blood-brain barrier penetration from chemical structure: combining passive permeability with active efflux by P-glycoprotein

Evolutionarily conserved roles for blood-brain barrier xenobiotic transporters in endogenous steroid partitioning and behavior

Regulation of ABC transporters blood-brain barrier: the good, the bad, and the ugly

Regulation of ABC efflux transporters at blood-brain barrier in health and neurological disorders

Predict drug permeability to blood-brain-barrier from clinical phenotypes: drug side effects and drug indications

Improved classification of blood-brain-barrier drugs using deep learning

Puzzling over schizophrenia: schizophrenia as a pathway disease

Schizophrenia: a neurodevelopmental disorder-Integrative genomic hypothesis and therapeutic implications from a transgenic mouse model

Schizophrenia: a concise overview of incidence, prevalence, and mortality

The global burden of disease: 2004 update. World Health Organization

Detecting neuroimaging biomarkers for schizophrenia: a metaanalysis of multivariate pattern recognition studies

Revolution stalled

Effective connectivity within a triple network brain system discriminates schizophrenia spectrum disorders from psychotic bipolar disorder at the single-subject level

Improving individual predictions: machine learning approaches for detecting and attacking heterogeneity in schizophrenia (and other psychiatric diseases)

Neuroanatomical heterogeneity of schizophrenia revealed by semi-supervised machine learning methods

Using machine learning to explain the heterogeneity of schizophrenia. Realizing the promise and avoiding the hype

Identifying schizophrenia subgroups using clustering and supervised learning

Machine learning improved classification of psychoses using clinical and biological stratification: update from the bipolar-schizophrenia network for intermediate phenotypes (B-SNIP)

Model-based optimization approaches for precision medicine: a case study in presynaptic dopamine overactivity

Identification of the gene signature reflecting schizophrenia's etiology by constructing artificial intelligence-based method of enhanced reproducibility

Development of MLR and SVM aided QSAR models to identify common SAR of GABA uptake herbal inhibitors used in the treatment of schizophrenia

Use of a novel artificial intelligence platform on mobile devices to assess dosing compliance in a phase 2 clinical trial in subjects with schizophrenia

Drug Repositioning for schizophrenia and depression/anxiety disorders: a machine learning approach leveraging expression. Data

Integrating clinical data and imputed transcriptome from GWAS to uncover complex disease subtypes: applications in psychiatry and cardiology

Guest editorial: Special issue on machine learning in schizophrenia

Aggression in autism spectrum disorder: presentation and treatment options

Toward the autism motor signature: gesture patterns during smart tablet gameplay identify children with autism

Potential identification of vitamin B6 responsiveness in autism spectrum disorder utilizing phenotype variables and machine learning methods

Repurposing approved drugs as inhibitors of Kv7.1 and Nav1.8 to treat Pitt Hopkins Syndrome

Predicting functional effects of missense variants in voltage-gated sodium and calcium channels

Neurobehavioural effects of developmental toxicity. The Lancet Neurology

From the cover: developmental neurotoxicants disrupt activity in cortical networks on microelectrode arrays: results of screening 86 compounds during neural network formation

Bayesian varying coefficient kernel machine regression to assess neurodevelopmental trajectories associated with exposure to complex mixtures

Predicting escitalopram monotherapy response in depression: the role of anterior cingulate cortex

Multisite prediction of 4-week and 52-week treatment outcomes in patients with first-episode psychosis: a machine learning approach

Cross-trial prediction of treatment outcome in depression: a machine learning approach

Reevaluating the efficacy and predictability of antidepressant treatments: a symptom clustering approach

The perilous path from publication to practice

ARPNet: antidepressant response prediction network for major depressive disorder

Use of machine learning for predicting escitalopram treatment outcome from electroencephalography recordings in adult patients with depression

An electroencephalographic signature predicts antidepressant response in major depression

Primary functional brain connections associated with melancholic major depressive disorder and modulation by antidepressants

Machine learning for precision psychiatry: opportunities and challenges

Machine learning approaches for clinical psychology and psychiatry

Artificial intelligence for mental health and mental illnesses: an overview

Ageing and Parkinson's disease: why is advancing age the biggest risk factor?

Computational systems biology approaches for Parkinson's disease

Boosting drug discovery for Parkinson's: enhancement of the delivery of a monoamine oxidase-B inhibitor by brain-targeted PEGylated polycaprolactone-based nanoparticles

Toxic proteins in neurodegenerative disease

Serum N-glycosylation in Parkinson's disease: a novel approach for potential alterations

Cerebrospinal fluid proteomic patterns discriminate Parkinson's disease and multiple system atrophy

Biosignatures for Parkinson's disease and atypical parkinsonian disorders patients

Detection of biomarkers with a multiplex quantitative proteomic platform in cerebrospinal fluid of patients with neurodegenerative disorders

Remote Monitoring of treatment response in Parkinson's disease: the habit of typing on a computer

Automatic assessment of medication states of patients with Parkinson's disease using wearable sensors

Insights into pharmacotherapy management for Parkinson's disease patients using wearables activity data

Assessment of response to medication in individuals with Parkinson's disease

Can the latest computerized technologies revolutionize conventional assessment tools and therapies for a neurological disease? The example of Parkinson's disease

Discovery of indolylpiperazinylpyrimidines with dual-target profiles at adenosine A2A and dopamine D2 receptors for Parkinson's disease treatment

QSAR modelling to identify LRRK2 inhibitors for Parkinson's disease

Repurposing drugs to treat l-DOPA-induced dyskinesia in Parkinson's disease

Machine learning-assisted neurotoxicity prediction in human midbrain organoids

Machine learning discriminates a movement disorder in a zebrafish model of Parkinson's disease

A century of Alzheimer's disease

Alzheimer's disease: clinical trials and drug development

Drug development status for Alzheimer's disease: present scenario

Alzheimer's disease drug-development pipeline: few candidates, frequent failures

Recent advancements in pathogenesis, diagnostics and treatment of Alzheimer's disease

Structure-based machine-guided mapping of amyloid sequence space reveals uncharted sequence clusters with higher solubilities

HENA, heterogeneous network-based data set for Alzheimer's disease. Sci Data

In silico investigation of traditional Chinese medicine compounds to inhibit human histone deacetylase 2 for patients with Alzheimer's disease

Neural network modeling of AChE inhibition by new carbazolebearing oxazolones

Development of predictive models for identifying potential S100A9 inhibitors based on machine learning methods

Machine learning from molecular dynamics trajectories to predict caspase-8 inhibitors against Alzheimer's disease

Comprehensive exploration of target-specific ligands using a graph convolution neural network

PTML modeling for Alzheimer's disease: design and prediction of virtual multitarget inhibitors of GSK3B, HDAC1, and HDAC6

Developing a multi-target model to predict the activity of monoamine oxidase A and B drugs

Identification of novel class I and class IIb histone deacetylase inhibitor for Alzheimer's disease therapeutics

Discovery of multitarget-directed ligands against Alzheimer's disease through systematic prediction of chemical-protein interactions

AlzhCPI: a knowledge base for predicting chemical-protein interactions towards Alzheimer's disease

Network pharmacology-based analysis of Chinese herbal Naodesheng formula for application to Alzheimer's disease

Design of natural-product-inspired multitarget ligands by machine learning

FDA approves galantamine for Alzheimer's disease

Integrating network, sequence and functional features using machine learning approaches towards identification of novel Alzheimer genes

A precision medicine framework using artificial intelligence for the identification and confirmation of genomic biomarkers of response to an Alzheimer's disease therapy: analysis of the blarcamesine (ANAVEX2-73) Phase 2a clinical study

Network topology and machine learning analyses reveal microstructural white matter changes underlying Chinese medicine Dengzhan Shengmai treatment on patients with vascular cognitive impairment

Robotic anesthesia-a vision for the future of anesthesia

Autonomous systems in anesthesia: where do we stand in 2020? A narrative review

Machine-learning implementation in clinical anesthesia: opportunities and challenges

Dexmedetomidine-induced deep sedation mimics nonrapid eye movement stage 3 sleep: large-scale validation using machine learning

Improved tracking of sevoflurane anesthetic states with drug-specific machine learning models

Predicting deep hypnotic state from sleep brain rhythms using deep learning: a data-repurposing approach

Development of a prediction model for hypotension after induction of anesthesia using machine learning

An automated algorithm incorporating poincare analysis can quantify the severity of opioid-induced ataxic breathing

A machine-learning approach to predicting hypotensive events in ICU settings

From patient-controlled analgesia to artificial intelligence-assisted patientcontrolled analgesia: practices and perspectives

Use of mobile health apps and wearable technology to assess changes and predict pain during treatment of acute pain in sickle cell disease: feasibility study

Machine learning approach to predict postoperative opioid requirements in ambulatory surgery patients

Multimodal recognition of pain intensity and pain modality with machine learning

Reducing opioid prescriptions by identifying responders on topical analgesic treatment using an individualized medicine and predictive analytics approach

Deep reinforcement learning for optimal critical care pain management with morphine using dueling double-deep Q networks

Predicting inadequate postoperative pain management in depressed patients: a machine learning approach

EEG beta-band spectral entropy can predict the effect of drug treatment on pain in patients with herpes zoster

Machine learning approach to predict medication overuse in migraine patients

Classification of opioid usage through semi-supervised learning for total joint replacement patients

A predictive-modeling-based screening tool for prolonged opioid use after surgical management of low back and lower extremity pain

Systems biology-based approaches to summarize and identify novel genes and pathways associated with acute and chronic postsurgical pain

Prediction and optimization of NaV1.7 sodium channel inhibitors based on machine learning and simulated annealing

Pain-CKB, a pain-domain-specific chemogenomics knowledgebase for target identification and systems pharmacology research

Startups using artificial intelligence in drug discovery

Advancing drug discovery via artificial intelligence

Artificial intelligence in drug development: present status and future prospects

Transforming Auto-Encoders

Selective serotonin 5-HT1A receptor biased agonists elicit distinct brain activation patterns: a pharmacoMRI study

Comparing the actions of lanicemine and ketamine in depression: key role of the anterior cingulate

Effects of lactate on the early visual cortex of non-human primates, investigated by pharmaco-MRI and neurochemical analysis

The role of fMRI in drug development

Understanding the neurovascular unit at multiple scales: advantages and limitations of multi-photon and functional ultrasound imaging

Functional ultrasound imaging of the brain

Pharmaco-fUS for characterizing drugs for Alzheimer's disease-the case of THN201, a drug combination of donepezil plus mefloquine

Functional ultrasound imaging to study brain dynamics: application of pharmaco-fUS to atomoxetine

Pharmaco-fUS: quantification of pharmacologically-induced dynamic changes in brain perfusion and connectivity by functional ultrasound imaging in awake mice

Adversarial controls for scientific machine learning

Mind and machine in drug design

Sezen is a physician-scientist, having earned her MD from Ankara University School of Medicine, Turkey, and her PhD in Biomedical Sciences and Engineering from Koc University Graduate School of Sciences and Engineering, Turkey. Her graduate studies focused on computer-aided drug discovery approaches for the treatment of infectious diseases and cancer. Currently, she is identifying novel drug targets for Alzheimer's disease and translating them into new, small molecule drug therapies. She is applying artificial intelligence (AI) and other advanced bioinformatics methods to integrate large multiomics, chemical

Avner Schlessinger is an associate professor of Pharmacological Sciences at the Icahn School of Medicine at Mount Sinai in New York City, an associate director of Mt