key: cord-0301277-wqomdbyf authors: Schneider, Constantin; Buchanan, Andrew; Taddese, Bruck; Deane, Charlotte M. title: DLAB - Deep learning methods for structure-based virtual screening of antibodies date: 2021-08-04 journal: bioRxiv DOI: 10.1101/2021.02.12.430941 sha: 8260afc18b505baade0db486d0d6b020877c952a doc_id: 301277 cord_uid: wqomdbyf Antibodies are one of the most important classes of pharmaceuticals, with over 80 approved molecules currently in use against a wide variety of diseases. The drug discovery process for antibody therapeutic candidates however is time-and cost-intensive and heavily reliant on in-vivo and in-vitro high throughput screens. Here, we introduce a framework for structure-based deep learning for antibodies (DLAB) which can virtually screen putative binding antibodies against antigen targets of interest. DLAB is built to be able to predict antibody-antigen binding for antigens with no known antibody binders. We demonstrate that DLAB can be used both to improve antibody-antigen docking and structure-based virtual screening of antibody drug candidates. DLAB enables improved pose ranking for antibody docking experiments as well as selection of antibody-antigen pairings for which accurate poses are generated and correctly ranked. We also show that DLAB can identify binding antibodies against specific antigens in a case study. Our results demonstrate the promise of deep learning methods for structure-based virtual screening of antibodies. Introduction antibody-antigen pairs by randomly sampling 50 non-cognate antibodies per antigen 118 from the crystal structure/model antibody data set. Our definition of non-cognate 119 required the sampled antibodies to share less than 90% of their CDR sequence with 120 the binding antibody. 121 Docking pose generation 122 For both the crystal structure and model data set, docked antibody-antigen pairs were 123 generated using ZDOCK [23] . For each pair, 500 poses (distinct structures of the 124 antibody-antigen complex) were generated. To aid the docking process, the paratope 125 and epitope were identified and residues not belonging to either the paratope or the 126 epitope were excluded from the interaction site using the standard ZDOCK pipeline as 127 described in the ZDOCK documentation. 128 The antigen epitope was defined by separating the antibody and antigen in the crystal structure and calculating surface exposed residues for each binding partner 130 separately using the PSA algorithm [27] . All atoms belonging to surface residues on 131 the antigen less than 4 Å from a surface exposed residue on the antibody were 132 considered part of the epitope. Further, all atoms belonging to a surface exposed 133 residue on the antigen within 4 Å of the defined epitope were included in the allowed 134 docking interface to model a realistic level of access to the epitope. 135 The paratope of crystal structure antibodies was defined in the same way, using 136 the interacting residues from the crystal structure. 137 For modelled antibody structures, the paratope was defined using the IMGT CDR 138 definition, marking the IMGT defined CDRs and two residues to either side of each 139 CDR as the paratope. 140 Clustering for train-test splits 141 To avoid similarity between binding modes in the train and test sets, CD-Hit [28] was 142 used to cluster antibodies by CDR sequence identity as defined above, using a 143 clustering cutoff of 90% sequence identity. For all learning tasks set out below, 144 train-test splits were performed using clustered cross-validation, assigning all members 145 of a cluster to either the train or the test set. 146 Data input for convolutional neural networks 147 Following the method of Ragoza et al. [21] , the docking poses were prepared for input 148 into CNNs by discretising the atom information into four-dimensional grids, where 149 three dimensions describe the spatial arrangement of the interaction site and the 150 fourth dimension is used to indicate atom types (see Supplementary Figure 3 ). 151 The centre of the interaction site of docking poses was calculated using the PSA 152 algorithm by averaging the coordinates of all surface-exposed atoms within 4 Å of the 153 interaction partner on both the antibody and the antigen and taking the mean of the 154 two center points. Poses which after docking had no interactions under 4 Å were 155 discarded. The grid contained only atoms that were within 24 Å of the interaction centre. intervals, we used a stratified sampling scheme, sampling poses from each interval at 174 the same rate during training (but not during testing). During training, the input data was augmented by random rotation around the 176 interaction centre, followed by random translations along the x, y and z axis between 177 −2 Å and 2 Å. Models were trained for 200,000 parameter update steps using 178 categorical cross-entropy and the rectified Adam optimiser. Since we wanted to use the improved ranking performance of DLAB-Re on the To identify antibody-antigen pairings with low quality docking poses, we 186 determined the highest DLAB-Re score given to any of the top 500 poses generated by 187 ZDOCK for each antibody-antigen pairing. This score (DLAB-Re-max) was used to 188 discard particular pairings by ranking all pairings by their DLAB-Re-max score and 189 discarding the bottom 40%, 60% or 80%. To contrast the performance of ZDOCK on 190 the same task, this score thresholding was also applied to the ZDOCK output score of 191 the top pose as ranked by ZDOCK. Virtual screening with DLAB-VS and ZDOCK The goal of virtual antibody screening is to discern binding antibodies against a given 194 epitope from a pool of candidate antibodies. To generate a classification model able to We created a SARS-CoV2 data set (the SARS-CoV2 variant data set) by 248 extracting all antibodies from the Coronavirus Antibody Database (CoVAbDab) [31] 249 7/28 which were confirmed to bind to the SARS-CoV2 wild-type RBD while also being 250 confirmed not to bind to at least one SARS-CoV2 variant and for which an 251 experimentally determined complex structure was available from which the epitope 252 could be determined as described above. We used ABodyBuilder to model the 253 antibodies as described above. We created structural models of the variant antigens 254 using Foldx5 [32] , using the PDB files listed in Supplementary Table 1 as templates 255 and copied the epitope definition for docking purposes from the templates onto the 256 variant models. We then docked each antibody model, defining the paratope as 257 described above, against its epitope on both the wild-type RBD and the confirmed 258 non-cognate variant RBDs and performed rescoring and binder classification as Figure 1E and 1F). Using this approach, termed DLAB-VS+ZDOCK, the binder was 370 ranked in the top 2% of antibodies for 6.4% of antigen targets, and in the top 10% in 371 19.7% of antigen targets (see Fig. 2 ). In the following, we use this 372 DLAB-VS+ZDOCK approach. Using DLAB-Re to discard antigen targets enables selection of while also being confirmed not to bind against at least one SARS-CoV2 RBD variant. 424 We ran the data set through the DLAB pipeline as described in section 3.10 and Supplementary Figure 8 . Performance of ZDock-based max score thresholding on ranking performance. While higher overall ZDock scores correlate with improved ranking performance, the effect is much less pronounced than using DLAB-Re-max scores. Supplementary Figure 10 . DLAB-VS and ZDock binder classification performance on different data sets. For each approach, the ratio of pairings for which the binding antibody was ranked in the top 2%, top 5%, top 10% and top 20% respectively is shown. Comparison of the performance of ZDock binder classification on the crystal data set to the random expectation ("random") of finding the binder in the top N% and the performance of ZDock, DLAB-VS and the combination of ZDock and DLAB-VS generated as detailed in the Methods section ("DLAB-VS+Zdock") on the model data set, both with and without using the DLAB-Re-max thresholding approach ("+ thresholding"). Supplementary Figure 11 . Dependence of the performance of DLAB-VS+ZDock on antibody model quality (A), measured via RMSD of the CDR regions to the corresponding crystal structure, and docking quality (B), measured via the highest fnat achieved in the top500 docking poses generated by ZDock. Both good antibody models and high quality docking poses correlate with high classification performance. Supplementary Figure 12 . Full overview of the performance of DLAB-VS+ZDock on the post-snapshot model data set. As in the main text figure, the bars show the fraction of antigens within the data set for which the correct binder is ranked by DLAB-VS within the top 2%, top 5%, top 10% or top 20% respectively. We compare the random expectation value ("random") to the whole post-snapshot model set ("all post-snapshot entries"), the antigen targets within the set for which the binding antibody does ("overlap with train set") or does not ("no overlap with train set") cluster with at least one antibody from the model data set at 90% CDR sequence identity. For each of these three options, the improved performance upon using the DLAB-Re-max score to discard 80% of antigen targets as described above is shown as well ("+ thresholding"). Supplementary Figure 13 . Performance of DLAB-VS+ZDOCK on the post-snapshot dataset at different percentage CDR sequence identity cutoffs for overlap with the training set without (A) and with (B) DLAB-Re thresholding. The performance of the DLAB-VS+ZDOCK model only marginally declines when limiting the allowed overlap to training set antibodies from 90% CDR ID to 70% CDR sequence identity, demonstrating that the model performance is generalisable. The Increasingly Human and Profitable Monoclonal Antibody Market The history of monoclonal antibody development -Progress, remaining challenges and future innovations Antibody-antigen complex modelling in the era of immunoglobulin repertoire sequencing In situ production of therapeutic monoclonal antibodies Five computational developability guidelines for therapeutic antibody profiling Progress and challenges in the design and clinical development of antibodies for cancer therapy Engineering antibody therapeutics Computational approaches to therapeutic antibody design: Established methods and emerging trends ABodyBuilder: Automated antibody structure prediction with data-driven accuracy estimation docking of antibody structures with Rosetta Kotai Antibody Builder: Automated high-resolution structural modeling of antibodies RosettaAntibodyDesign (RAbD): A general framework for computational antibody design Second Antibody Modeling Assessment (AMA-II) Prediction of site-specific interactions in antibody-antigen complexes: The proABC method and server Parapred: antibody paratope prediction using convolutional and recurrent neural networks Deep learning enables therapeutic antibody optimization in mammalian cells Prediction of VH-VL domain orientation for antibody variable domain modeling A Review of Deep Learning Methods for Antibodies A compact vocabulary of paratope-epitope interactions enables predictability of antibody-antigen binding Learning context-aware structural representations to predict antigen and antibody binding interfaces Protein-Ligand Scoring with Convolutional Neural Networks Protein Family-specific Models using Deep Neural Networks and Transfer Learning Improve Virtual Screening and Highlight the Need for More Data Accelerating protein docking in ZDOCK using an advanced 3D convolution library SAbDab: The structural antibody database The Protein Data Bank Antibody side chain conformations are position-dependent The interpretation of protein structures: Estimation of static accessibility Cd-hit: A fast program for clustering and comparing large sets of protein or nucleotide sequences Data Set Augmentation Allows Deep Learning-Based Virtual Screening to Better Generalize to Unseen Target Classes and Highlight Important Binding Interactions Protein docking model evaluation by 3D deep convolutional neural networks CoV-AbDab: the Coronavirus Antibody Database The FoldX web server: An online force field SciPy 1.0: Fundamental Algorithms for Scientific Computing in Python Testing the ratio of two poisson rates statsmodels: Econometric and statistical modeling with python Comparative analysis of the cdr loops of antigen receptors Prediction of protein-protein interactions: The CAPRI experiment, its evaluation and implications Structure-based cross-docking analysis of antibody-antigen interactions A) DLAB-Re architecture. A 3-layer CNN followed by a fully connected layer is used to predict membership of one of 11 fnat intervals. To make fnat predictions, the output of the softmax layer is transformed into a single score via weighted averaging A CNN architecture consisting of a single convolutional layer followed by 3 Denseblocks, using the same output layer as (B)