Antibody interface prediction with 3D Zernike descriptors and SVM

Abstract

Motivation Antibodies are a class of proteins capable of specifically recognizing and binding to a virtually infinite number of antigens. This binding malleability makes them the most valuable category of biopharmaceuticals for both diagnostic and therapeutic applications. The correct identification of the antigen-binding residues in the antibody is crucial for all antibody design and engineering techniques and could also help to understand the complex antigen binding mechanisms. However, the antibody-binding interface prediction field appears to be still rather underdeveloped. Results We present a novel method for antibody interface prediction from their experimentally solved structures based on 3D Zernike Descriptors. Roto-translationally invariant descriptors are computed from circular patches of the antibody surface enriched with a chosen subset of physico-chemical properties from the AAindex1 amino acid index set, and are used as samples for a binary classification problem. An SVM classifier is used to distinguish interface surface patches from non-interface ones. The proposed method was shown to outperform other antigen-binding interface prediction software. Availability and implementation Linux binaries and Python scripts are available at https://github.com/sebastiandaberdaku/AntibodyInterfacePrediction. The datasets generated and/or analyzed during the current study are available at https://doi.org/10.6084/m9.figshare.5442229. Supplementary information Supplementary data are available at Bioinformatics online.

In-Silico evidence for two receptors based strategy of SARS-CoV-2

“…In the last decade the 3D Zernike formalism has been widely applied for the characterization of molecular interactions [29,[34][35][36]: in this work we adopted a new representation, based on the 2D Zernike polynomials, which allows the quantitative characterization of protein surface regions. As shown in Fig.1, our computational protocol associates to each molecular patch an ordered set of numbers (the expansion coefficients) that describes its shapes.…”

Learning Context-aware Structural Representations to Predict Antigen and Antibody Binding Interfaces

“…Furthermore, the improvements in performance of all models after transfer learning illustrates the benefits of leveraging data from general protein-protein interactions to establish a base model that can be fine-tuned with antibody-antigen data. Table 3 summarizes the paratope test set prediction performance of our different neural networks and state-of-the-art structure-based methods Daberdaku et al [9] and Antibody i-Patch [23]. To Daberdaku et al [9] 0.658 0.950 --Antibody i-Patch [23] 0.376 0.840 -enable a direct comparison to previous studies, we predict for the entire structure of the antibody Fv region instead of just the CDR clouds as described in our methods.…”

“…Table 3 summarizes the paratope test set prediction performance of our different neural networks and state-of-the-art structure-based methods Daberdaku et al [9] and Antibody i-Patch [23]. To Daberdaku et al [9] 0.658 0.950 --Antibody i-Patch [23] 0.376 0.840 -enable a direct comparison to previous studies, we predict for the entire structure of the antibody Fv region instead of just the CDR clouds as described in our methods. Our networks perform better than the other methods on both AUC-PR and AUC-ROC, establishing the superior performance of learned features over pre-defined features as used by Daberdaku et al The network with a single layer of convolution and attention achieves the best performance, but the attention layer provides only a small performance improvement over convolution.…”

“…This yielded resulted in 103 complexes for training, 38 for validation, and 30 for testing. Paratope prediction: The dataset from Daberdaku et al [9] consists of 471 antibody-antigen complexes, with 213 complexes for training, 106 for validation, and 152 for testing. Since our framework accepts only proteins, we discarded complexes with non-protein antigens (e.g., DNA), resulting in 205 complexes for training, 103 for validation, and 152 for testing.…”

Deciphering interaction fingerprints from protein molecular surfaces

“…Its usefulness for many tasks involving protein interactions has long been known 7–9 , and has been the preferred structural description to study protein:solvent electrostatic interactions 10 . More recently, some efforts have captured molecular surface patterns with functional relevance, using techniques such as 3D Zernike descriptors 11–14 and geometric invariant fingerprint descriptors (GIF) 15 . These approaches proposed ‘handcrafted’ descriptors, manually-optimized vectors which describe protein surface features.…”