PIsToN: Evaluating Protein Binding Interfaces with Transformer Networks

Abstract

The computational studies of protein binding are widely used to investigate fundamental biological processes and facilitate the development of modern drugs, vaccines, and therapeutics. Scoring functions aim to predict complexes that would be formed by the binding of two biomolecules and to assess and rank the strength of the binding at the interface. Despite past efforts, the accurate prediction and scoring of protein binding interfaces remain a challenge. The physics-based methods are computationally intensive and often have to trade accuracy for computational cost. The possible limitations of current machine learning (ML) methods are ineffective data representation, network architectures, and limited training data. Here, we propose a novel approach called PIsToN (evaluating Protein binding Interfaces with Transformer Networks) that aim to distinguish native-like protein complexes from decoys. Each protein interface is transformed into a collection of 2D images (interface maps), where each im- age corresponds to a geometric or biochemical property in which pixel intensity represents the feature values. Such a data representation provides atomic-level resolution of relevant protein characteristics. To build hybrid machine learning models, additional empirical-based energy terms are computed and provided as inputs to the neural network. The model is trained on thousands of native and computationally-predicted protein complexes that contain challenging examples. The multi-attention transformer network is also endowed with explainability by highlighting the specific features and binding sites that were the most important for the classification decision. The developed PIsToN model significantly outperforms existing state-of-the-art scoring functions on well-known datasets.

A Comprehensive Survey of Scoring Functions for Protein Docking Models

“…Labeling the data: The CAPRI criterion was used to label all the docking models [109]. As described in [56], a docking is labeled acceptable (or correct) if:…”

“…PDB-2023 was used in the PIsToN paper [56] and contains complexes that are deposited in the RCSB Protein Data Bank (PDB) [1] in the year 2023. All docking models were generated using HDOCK [106].…”

“…MaSIF dataset was used for training and testing the MaSIF-search tool [54]. Later, the PIsToN work used 678 complexes from this dataset as a testing set, and for each complex, randomly used one non-native and one near-native conformation [56]. We used exactly this test dataset for our experiments here.…”

Molecular interactions between perlecan LG3 and the SARS‐CoV‐2 spike protein receptor binding domain

“…The binding interfaces were evaluated with PIsToN (protein binding interfaces with transformer networks), a recently developed novel deep learning-based approach (Stebliankin et al, 2023) for distinguishing native-like complexes from decoys. In this approach, each protein interface is transformed into a collection of 2D images (interface maps), where each image corresponds to a geometric or biochemical property in which pixel intensity represents the feature values (Figure S1).…”

ProInterVal: Validation of Protein-Protein Interfaces through Learned Interface Representations

“…14 Furthermore, the potential of the use of transformers to investigate protein-protein interfaces has been demonstrated by a recent approach, called PIsToN. 15 This approach differentiates nativelike protein complexes from incorrect conformations by transforming protein interfaces into twodimensional interface maps. It combines Vision Transformer 16 architecture with hybrid and multiattention networks.…”