key: cord-0333411-lelw59ys
authors: Puentes, Paola Ruiz; Valderrama, Natalia; González, Cristina; Daza, Laura; Muñoz-Camargo, Carolina; Cruz, Juan C.; Arbeláez, Pablo
title: PharmaNet: Pharmaceutical discovery with deep recurrent neural networks
date: 2020-10-22
journal: bioRxiv
DOI: 10.1101/2020.10.21.348441
sha: e297a38b9a136b5251bfe67c68c202eaa7e8ebde
doc_id: 333411
cord_uid: lelw59ys

The discovery and development of novel pharmaceuticals is an area of active research mainly due to the large investments required and long payback times. As of 2016, the development of a novel drug candidate required up to $ USD 2.6 billion in investment for only 10% rate of approval by the FDA. To help decreasing the costs associated with the process, a number of in silico approaches have been developed with relatively low success due to limited predicting performance. Here, we introduced a machine learning-based algorithm as an alternative for a more accurate search of new pharmacological candidates, which takes advantage of Recurrent Neural Networks (RNN) for active molecule prediction within large databases. Our approach, termed PharmaNet was implemented here to search for ligands against specific cell receptors within 102 targets of the DUD-E database, which contains 22886 active molecules. PharmaNet comprises three main phases. First, a SMILES representation of the molecule is converted into a raw molecular image. Second, a convolutional encoder processes the data to obtain a fingerprint molecular image that is finally analyzed by a Recurrent Neural Network (RNN). This approach enables precise predictions of the molecules’ target on the basis of the feature extraction, the sequence analysis and the relevant information filtered out throughout the process. Molecule Target prediction is a highly unbalanced detection problem and therefore, we propose that an adequate evaluation metric of performance is the area under the Normalized Average Precision (NAP) curve. PharmaNet largely surpasses the previous state-of-the-art method with 95.8% in the Receiver Operating Characteristic curve (ROC-AUC) and 58.9% in the NAP curve. We obtained a perfect performance for human farnesyl pyrophosphate synthase (FPPS), which is a potential target for antimicrobial and anticancer treatments. We decided to test PharmaNet for activity prediction against FPPS by searching in the CHEMBL data set. We obtained [3] potential inhibitors that were further validated through both molecular docking and in silico toxicity prediction. Most importantly, one of this candidates, CHEMBL2007613, was predicted as a potential antiviral due to its involvement on the PCDH17 pathway, which has been reported to be related to viral infections.

The development and subsequent market penetration of new pharmaceuticals is a 2 critical yet time consuming and expensive process that has increased in cost by nearly 3 150% over the last decade. In 2016, the development of just one medicine was estimated 4 at around $ USD 2.6 billion [1] . This is mainly attributed to the costs of pre-clinical 5 and clinical trials where ethical issues and complications are encountered very often. As 6 a result, only 10% of the pharmaceuticals that reach trials finally obtain FDA 7 approval [1, 2] . For these reasons, such large investments have often limited the 8 development of drugs for medical conditions where the niche market is not sufficient for 9 a payback in a reasonable time frame. Even for some molecules of urgent need such as 10 the antibiotics, where resistance is increasingly worrisome worldwide, there has been an 11 stagnation in the discovery of alternative candidate molecules for over a decade. As a 12 result, these issues in the discovery and production of pharmaceuticals have been seen 13 as an opportunity to explore new approaches that combine both experimental and 14 computational routes to accelerate the development. In this regard, some of the most 15 successful experimental approaches include soil-dwelling, Rule of 5 (Ro5), genomics, 16 proteomics, phenotypic screening, binding assays to identify relevant target interaction, 17 turbidimetric solubility measurements and high throughput solubility 18 measurements [3] [4] [5] [6] [7] . Despite the progress, such approaches still rely on large 19 investments in sophisticated infrastructure for automated manipulation of samples and 20 data collection and processing [8, 9] . Alternatively, in silico approaches are more 21 cost-effective and consequently, have attracted significant attention over the past few 22 years [8, 9] . Examples include virtual library screening, signature matching, molecular 23 docking, genetic association, pathway mapping, among others [6, 8, 9] . In this case, 24 however, the developed algorithms still lag behind in precision and effectiveness and the 25 obtained candidates might require considerable experimental testing [9, 10] . This 26 combined approach is therefore leading to the repurposing of known molecules for new 27 and more potent treatments, which is attractive for both companies and the 28 patients [11] . To reduce the time for screening and implementation of new therapeutic 29 candidates even further, recent advances in artificial intelligence (AI) have provided 30 more effective search algorithms that rely on the capacity to model relationships 31 between the variables, which can also be trained to discover patterns in significantly 32 large data sets simultaneously [12] . 33 Machine Learning-based algorithms have been particularly useful for improving drug 34 discovery because they can analyze large data sets and learn the optimal representation 35 for specific tasks rather than using hand-craft fingerprints, which are difficult to achieve 36 otherwise [2] . Moreover, computational techniques such as Support Vector Machines 37 (SVM) and Random Forests (RF) have been successfully applied for the design of 38 pharmaceuticals with high specificity and selectivity, and improved physiological 39 behavior in terms of important parameters such as circulation times, bioavailability and 40 biological activity [2] , toxicity [13] and potential side effects [14, 15] . These 41 developments have been enabled by the availability of large public databases with 42 information about the physicochemical and biological properties of pharmaceuticals [16] . 43 With this information it is possible to train deep learning models, which allow virtual 44 screening over large data sets by means of efficient optimization algorithms and new 45 computational capabilities [17] . A recent example of the application of such models was 46 the screening conducted by Stokes et al. [11] over a data set of more than 107 million 47 candidate molecules. The main result was the identification of the antibiotic potential 48 of halicin, which for the first time allowed the successful re-purpose of this molecule 49 fully in silico. Halicin was originally researched for the treatment of diabetes due to the 50 inhibition potential of the enzyme c-Jun N-terminal kinase but was abandoned because 51 of low performance [18] . This finding provides remarkable evidence for the notion that 52

AI is a suitable route for the screening and eventual development of new drugs. 53 Moreover, it offers the opportunity of a reduction in both the required investment for 54 development and the potential risks to be undertaken in pre-clinical and clinical trials. 55 Finally, it is possible to assure that from the beginning of the development, candidate 56 molecules comply with requirements imposed by regulatory frameworks in terms 57 safeness and reliability. 58 The current global COVID-19 pandemic is a compelling example of the urgent need 59 for automating drug discovery, as this situation is the result of a novel coronavirus 60 (SARS-CoV-2) capable of infecting humans at an extremely fast pace [19] . To respond 61 to this contingency, novel antiviral treatments and vaccines need to be developed in an 62 extraordinarily short time. In this regard, according to the experts and even with the 63 unprecedented resources allocated by governments, the shortest possible period for 64 developing and deploying a COVID-19 vaccine is of about eighteen months [20, 21] . The 65 European Union has raised $ 8 billion for collaborative development and universal 66 deployment of diagnostics, treatments and vaccines against SARS-CoV-2 [22] . This is 67 also the case of the U.S. and German governments, which are planning to invest in 68 vaccine and treatments development and distribution over $ 2 billion and $ 812 million, 69 respectively [23, 24] . 70 Here, we applied recent developments in the field of computer vision to the critical 71 task of active molecule prediction, which mainly involves the estimation of whether a 72 molecule is able to bind to particular membrane receptors. Starting from the publicly 73 available AD Dataset [25] , we formalized active molecule prediction as a detection 74 problem for which we designed an experimental framework that allowed us to evaluate 75 results with the aid of normalized Precision-Recall curves [26] . According to our newly 76 proposed framework, the state-of-the-art technique only performs with a 1% efficiency, 77 however, it was reported to show an AUC score of 52% [25] . In search for a superior 78 performance, we developed an algorithm based on deep learning for active molecule 79 prediction, which we called PharmaNet. Our approach elevated the prediction 80 performance (i.e., the area under the Normalized Average Precision (NAP) curve) to 81 the unprecedented level of 58.9%.

PharmaNet was designed on the basis of natural language processing (NLP) 83 techniques given that in this case the most important information lies in the sequence of 84 each of the elements. Consequently, we implemented recurrent neural networks (RNNs) 85 as the baseline for PharmaNet due to their demonstrated performance in problems 86 involving language [27] [28] [29] . Specifically, we considered a Gated Recurrent Unit (GRU) 87 cell as it enables the analysis of atom sequences in an information flow direction that 88 finalizes in the current element by analyzing the ones before it. These architectures 89 have been used previously explored for similar tasks such as those required for property 90 prediction and the generation of molecules according to properties of interest. For 91 instance, Marwin et al. [30] trained Long Short-Term Memory (LSTM) cells to learn a 92 statistical chemical language model for the generation of large sets of novel molecules 93 with similar physicochemical properties to those in the training set. The LSTM network 94 receives as input a canonical Simplified Molecular Input Line Entry System (SMILES) 95 representation of the molecules. In the same way, Goh et al. [31] used SMILES as the 96 input of a GRU to predict different chemical properties of the pharmaceuticals. In 97 consequence, due to the versatility of the SMILES format, we implemented it for data 98 representation in PharmaNet. This approach allowed us to build and map a 2D 99 representation of the molecules as simple sequences of characters with varying positions 100 in such 2D space [32] . Furthermore, our method enabled us to classify a molecule against multiple targets with 117 a single trained model, which turned out to be much more efficient than most of the 118 previously reported ones [25, 33, 34] . This is because as opposed to Pharmanet, such 119 approaches are trained to classify between active and decoys for individual targets.

Finally, our experiments demonstrated that all existing methods for this task provide a 121 performance that is nominally zero, which, to our knowledge, positions PharmaNet as 122 the most robust AI algorithm for target molecule prediction.

According to our measures, PharmaNet obtains a maximum performance for human 124 farnesyl pyrophosphate synthase (FPPS) as a target. This protein is a key enzyme on 125 the mevalonate (MVA) pathway that is responsible for the isoprenoid biosynthesis 126 where it catalyzes the formation of farnesyl diphosphate (FDP). This is a precursor for 127 several classes of essential metabolites including sterols, dolichols, carotenoids, and 128 ubiquinones [35] . Overexpression of FPPS has been reported for multiple types of 129 cancer, including prostate, glioblastoma, breast, and bone metastases from breast 130 cancer. FPPS is therefore a potential target for anticancer treatments [36] [37] [38] [39] [40] [41] . Also, 131 silencing of FPPS via siRNA has slowed down viral influenza A replication and release 132 in infected cells. This antiviral activity has been attributed to altered plasma membrane 133 fluidity and consequently to a limited formation of the lipid rafts required for the 134 survival of the virus [42] . On this basis, FPPS might potentially exhibit antiviral 135 activity against enveloped viruses [42] . Given the pharmacological application for FPPS 136 inhibitors and the high performance of our approach for FPPS, we evaluated a subset of 137 the CHEMBL dataset to search for candidates towards this target. Our top prediction 138 was the CHEMBL2007613 member of the data set, which corresponded to human cells, as well as to support hazard and risk assessment activities [43] . Taken 144 together, our results strongly suggest that CHEMBL2007613 is a potential candidate for 145 antiviral treatments.

The DUD-E database contains 22886 active molecules against 102 targets and 50 decoys 149 per target. Each decoy is a chemical compound with physicochemical properties similar 150 to those of the active ones but different structure. Both groups of molecules (i.e., active 151 compounds and decoys) have therefore different data distribution, thereby making the 152 binary classification problem more amenable for a neural network. This approach has 153 been proved successful previously by Gonczarek A. et al. [34] and Chen et al. [25] .

The Active Decoys (AD) data set was proposed by [25] as a strategy to eliminate the 155 bias introduced by the decoys. This data set is based on DUD-E but changes the decoys 156 of each target by those contained within the 101 receptors with the highest affinity 157 towards the target as estimated by molecular docking. This approach leads to a rather 158 challenging binary classification problem because all the molecules show the same data 159 distribution.

The output of our model for a molecule is the binding probability distribution to 162 individual targets within the set of target classes. This was accomplished by defining a 163 multiclass classification problem instead of a binary one. For this reason, the ligand 164 sequence is labeled with the corresponding target protein prior to be input into our 165 model. 166 We randomly split the complete set of binding ligand-protein SMILES sequences 167 from the AD data set into two main subsets where 90% of the available active molecules 168 for a target helped training the model, while the remaining 10% were employed for 169 testing purposes. Then, we split the training subset into four folders by making sure 170 that all the folders had the same distribution of active molecules per target.

Subsequently, we conducted a four-fold cross-validation approach where data in three of 172 the folders were used to train each model while the remaining one was only for 173 validation. This multi-step validation approach allowed us to train the model very 174 robustly and finally report the mean performance of the models as an overall metric. 175 Lastly, we converted the SMILES representation of the active molecules in the data set 176 into Raw molecular images. This process is described into detail below. denominate a raw molecular image. In the second phase, that representation is the 181 input of a CNN to obtain a fingerprint molecular image that incorporates information 182 of each atom and its neighbors [44] . This stage allows the model to extract fingerprints 183 of functional groups in the molecule, which are essential in defining their functionality. 184 After this stage, the CNN's output is processed by an RNN architecture that enables a 185 sequence-based analysis of the molecules [27] . This approach enables precise predictions 186 of the molecules' target on the basis of the feature extraction, the sequence analysis and 187 the relevant information filtered in the last two stages. The extracted information allows a different representation of the molecule in which the model can learn the PharmaNet's input is a SMILE sequence embedded into a one hot vector as performed 194 in [31] . We then generated a matrix of size n x m, where n = 36 is the total of unique 195 chars (atoms and type of bonds) in the data set, and m = 116 is the longest SMILES 196 sequence over all the molecules. This representation can be seen as a two dimensional 197 image of the molecule. This type of data arrays are typically encountered in computer 198 vision algorithms, which is attractive as we have extensive expertise with such 199 approaches [45] [46] [47] . As delineated below, we indeed applied our most recent information between them, namely, the reset gate (r ) and the update gate (z ). (r ) 217 decides whether to keep, in the current cell state (h' t ), information of the previous cell 218 or to change it by information from the input (x t ). The recurrence of this process is 219 described by the following set of equations:

In parallel, (z ) controls which type of information of the previous and current cell will 222 go to the next one. The process is described by the following set of equations

where W and U are weight matrices, b is a non-linearity, h' is the current cell state and 225 h is the output of the memory block [27, 49] . Finally, we considered the last hidden which has been also known as Average Precision (AP).

Nevertheless, as pointed out by Hoiem et al. [26] , AP strongly depends on the true 254 positive samples in each class t (N t ). As a result, the best performance is for classes 255 with the largest numbers of true positives. To create a normalized precision P n , [26] 256 replaces (N t ) with a constant N that corresponds to the average of positive samples over 257 all the classes. Eq. 5 corresponds to the definition of Normalized Precision (P N ), where 258 R is the fraction of objects detected while F is the number of those incorrectly detected. 259

As not all targets have the same number of active molecules, we prefer to use AP n 260 for evaluating the prediction on each target. We tested the four trained models in the 261 test subset and computed the AP n for each target. With the estimates of mean and 262 standard deviation for the four models, we then computed the Normalized Averaged 263 Precision (NAP) curve based on the frequency of targets that achieve each AP n score. 264 We estimated the overall performance of PharmaNet as the area under the NAP curve 265 normalized by the total number of targets. In order to verify the complexity of the binding ligand-receptor task in the AD data set, 315 we applied molecular docking for the best five hundred (500) molecules predicted by 

Comparison with the state-of-the-art AI algorithms 346 We compared the performance of PharmaNet against the prediction made by the 347 state-of-the-art method recently published by Chen et al. [25] for each target in the AD 348 data set. PharmaNet is 60.8% (Fig. 3B) . We obtained consistent results when we evaluated both 359 methods with the area under the NAP curve, in which [25] achieved 1% while ours 360 approached to 58.9% (Fig. 3C) . main phases of the architecture and their performance in the prediction task. Fig. 4A 365 shows the configuration of PharmaNet for our best result, the same architecture without 366 the convolutional encoder and without the RNN. It is evident that the RNN is the most 367 important component for superior performance but their combination is still beneficial 368 for improving performance somewhat further. Regarding the specific RNN implemented 369 herein, Fig. 4B shows that training with an LSTM decreases significantly PharmaNet's 370 performance. Considering the amount of parameters that an LSTM has to learn and the 371 time to train this network, we decided to keep the GRU as our main RNN architecture. 372 Regarding the Convolutional Encoder, experiments show that the best performance 373 was achieved with two convolutional layers and a residual connection (Fig. 5A, and 5D ). 374 However, we obtained similar results after eliminating one layer. We also found that 375 batch normalization produced the best results (Fig. 5C ) and that the optimal kernel 376 size was 5x5 (Fig. 5B) . 377 We also studied directionality, depth and hidden state size to establish the best 378 configuration for the RNN. Fig. 6A shows that an bidirectional configuration leads to 379 an slight improvement in performance compared with unidirectional GRU cells. This antibiotics [59] . In this regard, having a method to predict with high accuracy new 392 active molecules towards the target proteins might propel the drug discovery process 393 unprecedentedly. Also, we identified multiple possible targets for the development of 394 novel pharmacological anticancer therapies. This is particularly important for certain 395 types of cancer that are resistant to conventional chemotherapy such as doxorubicin [60] , 396 imatinib [61] , nilotinib [62] , cisplatin [63] , tamoxifen [64] , paclitaxel [65] , 397 temozolomide [66] , and docetaxel [67] . Finally, one of the receptors could be a potential 398 target for Parkinson's disease and consequently, a route to improve the palliative 399 treatments for the disease.

A closer inspection of each class separately allowed us to identify certain functional 401 groups that enabled a better classification, as we illustrate in Table 2 . Nitrogen atoms, which have been reported previously to be essential in the interaction 414 with the IPP active site of FPPS [68] . However, it can also be seen that none of the candidates has a phosphonic acid group, which we detected previously as a common 416 moiety of active molecules towards FPPS. This unexpected outcome supports the idea 417 that neural networks learn different fingerprints that those proposed and implemented 418 by chemists for drug design for decades [11] . This shift on the molecule and fingerprint 419 learning will enable the discovery of pharmaceuticals within a further representation 420 space that consequently leads to significant structural differences. This, in turn, is likely 421 to lead to revealing new and different mechanisms of action for known molecules. This is 422 a very important characteristic for drug discovery since drugs with similar mechanisms 423 of action are very likely to have the same drawbacks of current pharmaceuticals.

To further corroborate the performance of our method, the top-10 candidates were 425 analyzed via molecular docking and compared with Zolonadrate, a molecule reported to 426 induce inhibition of FPPS (Table 3 ) [68] . All of our candidates showed binding energies 427 lower than Zolonadrate's and consequently, they should exhibit better affinity towards 428 FPPS than Zolonadrate. This suggests a potential pharmacological use of the identified 429 molecules and therefore the need for further toxicity studies. The interactions and The obtained binding energy for Zolonadrate agrees well with that reported 433 elsewhere [55] . In this case (Fig. 8A) , the prevalent interaction is by means of hydrogen 434 bonding between the hydroxyl groups of Zolonadrate and the amine groups of the 435 residues Gln254 and Arg126 [69] . Additional hydrogen bonding takes place between the 436 protonated nitrogen atom of the heterocyclic ring in the side chain of Zolonadrate and 437 the conserved main-chain carbonyl oxygen of Lys214 and the hydroxyl group of the 438 Thr215. This stabilization mechanism resembles that of a carbocation intermediate [68] . Finally, CHEMBL222102 (2-morpholin-4-yl-6-thianthren-1-ylpyran-4-one) (Fig. 8D ) has 450 one heterocyclic ring with only one Nitrogen atom capable of forming hydrogen bonds 451 primarily with Arg126 residues, which explains the lowest biding energy of the analyzed 452 set of molecules.

After corroborating the interaction through molecular docking, the top-3 candidates 454 were analyzed with the aid of the online servers GUSAR, for cytotoxicity, and 455 DIGEP-Pred for genotoxicity. Table 4 shows the toxicity label for the three compounds 456 after administration via four different routes. Toxicity is categorized in a relative scale 457 that goes from 1 to 5, with 1 for absence of toxicity and 5 the highest toxicity [57] . only the most promising candidates are further analyze. We implemented this approach 483 here and put forward PharmaNet, a deep learning architecture for predicting binding of 484 a molecule to possible protein target receptors. PharmaNet´s algorithm represents the 485 2D structural information of a molecule as a molecular image and process it with 486 modern computer vision recognition techniques. Our architecture is trained end-to-end 487 and consists of a convolutional encoder processing phase followed by an RNN. This 488 approach allows multiclass classification with a single model. The conventional metric 489 for this type of task has been the Receiver Operating Characteristic curve (ROC-AUC); 490 however, we propose that a more accurate metric is the area under the Normalized 491 Average Precision (NAP) curve. Under this framework, PharmaNet outperforms the 492 state-of-the-art algorithm by one order of magnitude (from 1% to 58.9%) in the AD 493 dataset, and has a perfect performance in identifying the active molecules for 3 494 receptors of the 102 targets: CXCR4, FPPS and KITH. We selected FPPS as target to 495 apply our model in the search of active molecules within the large database of 496 pharmacological molecules, ChEMBL. We chose the 10 best candidate molecules to 497 investigate into detail interactions with FPSS via molecular docking. We also conducted 498 an in silico evaluation of toxicity with the aid of online servers. We found that the 499 compound identified with the ID CHEMBL2007613, i.e., 500 (5-[(5-Amino-4H-1,2,4-triazol-3-yl)amino]sulfonyl-2-chloro-4-mercaptophenyl acetate) 501 exhibits potential antiviral activity, which needs to be corroborated in vitro. We expect 502 that our algorithm opens new opportunities for the rediscovery and repurpose of 503 pharmacological compounds that otherwise might be disregarded in importance by the 504 pharmaceutical industry. Moreover, we are currently exploring its potential in the 505 reverse problem, i.e., searching for multiple receptor targets for molecules with certain 506 physicochemical properties. 

Computational methods in drug discovery

Artificial intelligence in drug design

Expanding the soil antibiotic resistome: exploring environmental diversity. Current opinion in microbiology

Flexible-Acceptor" General Solubility Equation for beyond Rule of 5 Drugs

Experimental and computational approaches to estimate solubility and permeability in drug discovery and development settings

Drug repurposing: progress, challenges and recommendations

High throughput solubility measurement in drug discovery and development. Advanced drug delivery reviews

Idea2Data: toward a new paradigm for drug discovery

Computational methods in drug discovery

High-throughput and in silico screenings in drug discovery

A deep learning approach to antibiotic discovery

Concepts of artificial intelligence for computer-assisted drug discovery

DeepTox: toxicity prediction using deep learning

Large-scale prediction of drug-target interactions from deep representations

Large-scale prediction of drug-target interactions from deep representations

Pharmaceutical Research: Databases. University of Southern California

Deep Learning

Powerful antibiotic discovered using machine learning for first time. The Guardian

Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) and corona virus disease-2019 (COVID-19): the epidemic and the challenges

Human trials for a coronavirus vaccine could begin 'within a few weeks,' top US health official says

Why a coronavirus vaccine is more than a year away, despite medical researchers' progress. USA Today

Coronavirus Global Response: e7.4 billion raised for universal access to vaccines

Federal Government To Invest Over $2 Billion Into Coronavirus Vaccine Development

Coronavirus: Germany earmarks e750 million on vaccine development. The Brussels Times

Hidden bias in the DUD-E dataset leads to misleading performance of deep learning in structure-based virtual screening

Diagnosing error in object detectors

Learning phrase representations using RNN encoder-decoder for statistical machine translation

Comparative study of cnn and rnn for natural language processing

Deep Learning applied to NLP

Generating focused molecule libraries for drug discovery with recurrent neural networks

Smiles2vec: An interpretable general-purpose deep neural network for predicting chemical properties

SMILES. 2. Algorithm for generation of unique SMILES notation

AtomNet: a deep convolutional neural network for bioactivity prediction in structure-based drug discovery

Learning deep architectures for interaction prediction in structure-based virtual screening

Human isoprenoid synthase enzymes as therapeutic targets

FDPS cooperates with PTEN loss to promote prostate cancer progression through modulation of small GTPases/AKT axis

Deregulated expression and activity of Farnesyl Diphosphate Synthase (FDPS) in Glioblastoma

Inhibition of farnesyl pyrophosphate (FPP) and/or geranylgeranyl pyrophosphate (GGPP) biosynthesis and its implication in the treatment of cancers

Farnesyl diphosphate synthase is important for the maintenance of glioblastoma stemness

Adjuvant bisphosphonate treatment for breast cancer: Where are we heading and can the pre-clinical literature help us get there

Bisphosphonates as anticancer agents in early breast cancer: preclinical and clinical evidence

The Interferon-Inducible Protein Viperin Inhibits Influenza Virus Release by Perturbing Lipid Rafts

Genetic toxicology in silico protocol

Imagenet classification with deep convolutional neural networks. In: Advances in neural information processing systems

Hypercolumns for Object Segmentation and Fine-Grained Localization

Deep Retinal Image Understanding

ISINet: An Instance-Based Approach for Surgical Instrument Segmentation

Deep residual learning for image recognition

Exploring recurrent neural networks for on-line handwritten signature biometrics

Learning to detect natural image boundaries using local brightness, color, and texture cues

Microsoft COCO: Common Objects in Context

The Pascal Visual Object Classes Challenge: A Retrospective

A Database of Human Segmented Natural Images and its Application to Evaluating Segmentation Algorithms and Measuring Ecological Statistics

The Brussels Times

Insilico Docking Analysis of Nitrogen Containing Bisphosphonate with Human Fernasyl Pyrophosphate Synthase

Understanding Chemical-Biological Interactions

QSAR Modelling of Rat Acute Toxicity on the Basis of PASS Prediction

DIGEP-Pred: web service for in silico prediction of drug-induced gene expression profiles based on structural formula

The Antibiotic Resistance Crisis. Part 1: Causes and Threats. Pharmacy and Therapeutics

Role of MicroRNA miR-27a and miR-451 in the regulation of MDR1/P-glycoprotein expression in human cancer cells

Restoration of miR-424 suppresses BCR-ABL activity and sensitizes CML cells to imatinib treatment

The Different Mechanisms of Cancer Drug Resistance: A Brief Review

MicroRNA-221 induces cell survival and cisplatin resistance through PI3K/Akt pathway in human osteosarcoma

MicroRNA-221/222 confers tamoxifen resistance in breast cancer by targeting p27Kip1

Role of microRNAs in drug-resistant ovarian cancer cells

miR-29c contribute to glioma cells temozolomide sensitivity by targeting O6-methylguanine-DNA methyltransferases indirectely

miRNA-34a is associated with docetaxel resistance in human breast cancer cells

The inhibition of human farnesyl pyrophosphate synthase by nitrogen-containing bisphosphonates. Elucidating the role of active site threonine 201 and tyrosine 204 residues using enzyme mutants

Probing the molecular and structural elements of ligands binding to the active site versus an allosteric pocket of the human farnesyl pyrophosphate synthase

Functional Associations: PCDH17 Gene

We sincerely thank authors from [25] for kindly giving us their predictions in order to 509 make a more direct comparison of both methods.