Deep graph learning of inter-protein contacts

Molecular surfaces modeling: Advancements in deep learning for molecular interactions and predictions

Molecular surface analysis can provide a high-dimensional, rich representation of molecular properties and interactions, which is crucial for enabling powerful predictive modeling and rational molecular design across diverse scientific and technological domains. With remarkable successes achieved by artificial intelligence (AI) in different fields such as computer vision and natural language processing, there is a growing imperative to harness AI's potential in accelerating molecular discovery and innovation. The integration of AI techniques with molecular surface analysis has opened up new frontiers, allowing researchers to uncover hidden patterns, relationships, and design principles that were previously elusive. By leveraging the complementary strengths of molecular surface representations and advanced AI algorithms, scientists can now explore chemical space more efficiently, optimize molecular properties with greater precision, and drive transformative advancements in areas like drug development, materials engineering, and catalysis. In this review, we aim to provide an overview of recent advancements in the field of molecular surface analysis and its integration with AI techniques. These AI-driven approaches have led to significant advancements in various downstream tasks, including interface site prediction, protein-protein interaction prediction, surface-centric molecular generation and design.

ToxDL 2.0: Protein toxicity prediction using a pretrained language model and graph neural networks

Motivation

Assessing the potential toxicity of proteins is crucial for both therapeutic and agricultural applications. Traditional experimental methods for protein toxicity evaluation are time-consuming, expensive, and labor-intensive, highlighting the requirement for efficient computational approaches. Recent advancements in language models and deep learning have significantly improved protein toxicity prediction, yet current models often lack the ability to integrate evolutionary and structural information, which is crucial for accurate toxicity assessment of proteins.

Results

In this study, we present ToxDL 2.0, a novel multimodal deep learning model for protein toxicity prediction that integrates both evolutionary and structural information derived from a pretrained language model and AlphaFold2. ToxDL 2.0 consists of three key modules: (1) a Graph Convolutional Network (GCN) module for generating protein graph embeddings based on AlphaFold2-predicted structures, (2) a domain embedding module for capturing protein domain representations, and (3) a dense module that combines these embeddings to predict the toxicity. After constructing a comprehensive toxicity benchmark dataset, we obtained experimental results on both an original non-redundant test set (comprising pre-2022 protein sequences) and an independent non-redundant test set (a holdout set of post-2022 protein sequences), demonstrating that ToxDL 2.0 outperforms existing state-of-the-art methods. Additionally, we utilized Integrated Gradients to discover known toxic motifs associated with protein toxicity. A web server for ToxDL 2.0 is publicly available at www.csbio.sjtu.edu.cn/bioinf/ToxDL2/.

PIPENN-EMB ensemble net and protein embeddings generalise protein interface prediction beyond homology

Protein interactions are crucial for understanding biological functions and disease mechanisms, but predicting these remains a complex task in computational biology. Increasingly, Deep Learning models are having success in interface prediction. This study presents PIPENN-EMB which explores the added value of using embeddings from the ProtT5-XL protein language model. Our results show substantial improvement over the previously published PIPENN model for protein interaction interface prediction, reaching an MCC of 0.313 vs. 0.249, and AUROC 0.800 vs. 0.755 on the BIO_DL_TE test set. We furthermore show that these embeddings cover a broad range of 'hand-crafted' protein features in ablation studies. PIPENN-EMB reaches state-of-the-art performance on the ZK448 dataset for protein-protein interface prediction. We showcase predictions on 25 resistance-related proteins from Mycobacterium tuberculosis. Furthermore, whereas other state-of-the-art sequence-based methods perform worse for proteins that have little recognisable homology in their training data, PIPENN-EMB generalises to remote homologs, yielding stable AUROC across all three test sets with less than 30% sequence identity to the training dataset, and even to proteins with less than 15% sequence identity.

AI-integrated network for RNA complex structure and dynamic prediction

RNA complexes are essential components in many cellular processes. The functions of these complexes are linked to their tertiary structures, which are shaped by detailed interface information, such as binding sites, interface contact, and dynamic conformational changes. Network-based approaches have been widely used to analyze RNA complex structures. With their roots in the graph theory, these methods have a long history of providing insight into the static and dynamic properties of RNA molecules. These approaches have been effective in identifying functional binding sites and analyzing the dynamic behavior of RNA complexes. Recently, the advent of artificial intelligence (AI) has brought transformative changes to the field. These technologies have been increasingly applied to studying RNA complex structures, providing new avenues for understanding the complex interactions within RNA complexes. By integrating AI with traditional network analysis methods, researchers can build more accurate models of RNA complex structures, predict their dynamic behaviors, and even design RNA-based inhibitors. In this review, we introduce the integration of network-based methodologies with AI techniques to enhance the understanding of RNA complex structures. We examine how these advanced computational tools can be used to model and analyze the detailed interface information and dynamic behaviors of RNA molecules. Additionally, we explore the potential future directions of how AI-integrated networks can aid in the modeling and analyzing RNA complex structures.

Large language models and their applications in bioinformatics

Recent advancements in Natural Language Processing (NLP) have been significantly driven by the development of Large Language Models (LLMs), representing a substantial leap in language-based technology capabilities. These models, built on sophisticated deep learning architectures, typically transformers, are characterized by billions of parameters and extensive training data, enabling them to achieve high accuracy across various tasks. The transformer architecture of LLMs allows them to effectively handle context and sequential information, which is crucial for understanding and generating human language. Beyond traditional NLP applications, LLMs have shown significant promise in bioinformatics, transforming the field by addressing challenges associated with large and complex biological datasets. In genomics, proteomics, and personalized medicine, LLMs facilitate identifying patterns, predicting protein structures, or understanding genetic variations. This capability is crucial, e.g., for advancing drug discovery, where accurate prediction of molecular interactions is essential. This review discusses the current trends in LLMs research and their potential to revolutionize the field of bioinformatics and accelerate novel discoveries in the life sciences.

Computational tools to predict context-specific protein complexes

Interactions between thousands of proteins define cells' protein-protein interaction (PPI) network. Some of these interactions lead to the formation of protein complexes. It is challenging to identify a protein complex in a haystack of protein-protein interactions, and it is even more difficult to predict all protein complexes of the complexome. Simulations and machine learning approaches try to crack these problems by looking at the PPI network or predicted protein structures. Clustering of PPI networks led to the first protein complex predictions, while most recently, atomistic models of protein complexes and deep-learning-based structure prediction methods have also emerged. The simulation of PPI level interactions even enables the quantitative prediction of protein complexes. These methods, the required data sources, and their potential future developments are discussed in this review.

Pairing interacting protein sequences using masked language modeling

Predicting which proteins interact together from amino acid sequences is an important task. We develop a method to pair interacting protein sequences which leverages the power of protein language models trained on multiple sequence alignments (MSAs), such as MSA Transformer and the EvoFormer module of AlphaFold. We formulate the problem of pairing interacting partners among the paralogs of two protein families in a differentiable way. We introduce a method called Differentiable Pairing using Alignment-based Language Models (DiffPALM) that solves it by exploiting the ability of MSA Transformer to fill in masked amino acids in multiple sequence alignments using the surrounding context. MSA Transformer encodes coevolution between functionally or structurally coupled amino acids within protein chains. It also captures inter-chain coevolution, despite being trained on single-chain data. Relying on MSA Transformer without fine-tuning, DiffPALM outperforms existing coevolution-based pairing methods on difficult benchmarks of shallow multiple sequence alignments extracted from ubiquitous prokaryotic protein datasets. It also outperforms an alternative method based on a state-of-the-art protein language model trained on single sequences. Paired alignments of interacting protein sequences are a crucial ingredient of supervised deep learning methods to predict the three-dimensional structure of protein complexes. Starting from sequences paired by DiffPALM substantially improves the structure prediction of some eukaryotic protein complexes by AlphaFold-Multimer. It also achieves competitive performance with using orthology-based pairing.

Review and Comparative Analysis of Methods and Advancements in Predicting Protein Complex Structure

Protein complexes perform diverse biological functions, and obtaining their three-dimensional structure is critical to understanding and grasping their functions. In many cases, it's not just two proteins interacting to form a dimer; instead, multiple proteins interact to form a multimer. Experimentally resolving protein complex structures can be quite challenging. Recently, there have been efforts and methods that build upon prior predictions of dimer structures to attempt to predict multimer structures. However, in comparison to monomeric protein structure prediction, the accuracy of protein complex structure prediction remains relatively low. This paper provides an overview of recent advancements in efficient computational models for predicting protein complex structures. We introduce protein-protein docking methods in detail and summarize their main ideas, applicable modes, and related information. To enhance prediction accuracy, other critical protein-related information is also integrated, such as predicting interchain residue contact, utilizing experimental data like cryo-EM experiments, and considering protein interactions and non-interactions. In addition, we comprehensively review computational approaches for end-to-end prediction of protein complex structures based on artificial intelligence (AI) technology and describe commonly used datasets and representative evaluation metrics in protein complexes. Finally, we analyze the formidable challenges faced in current protein complex structure prediction tasks, including the structure prediction of heteromeric complex, disordered regions in complex, antibody-antigen complex, and RNA-related complex, as well as the evaluation metrics for complex assessment. We hope that this work will provide comprehensive knowledge of complex structure predictions to contribute to future advanced predictions.

A variational expectation-maximization framework for balanced multi-scale learning of protein and drug interactions

Protein functions are characterized by interactions with proteins, drugs, and other biomolecules. Understanding these interactions is essential for deciphering the molecular mechanisms underlying biological processes and developing new therapeutic strategies. Current computational methods mostly predict interactions based on either molecular network or structural information, without integrating them within a unified multi-scale framework. While a few multi-view learning methods are devoted to fusing the multi-scale information, these methods tend to rely intensively on a single scale and under-fitting the others, likely attributed to the imbalanced nature and inherent greediness of multi-scale learning. To alleviate the optimization imbalance, we present MUSE, a multi-scale representation learning framework based on a variant expectation maximization to optimize different scales in an alternating procedure over multiple iterations. This strategy efficiently fuses multi-scale information between atomic structure and molecular network scale through mutual supervision and iterative optimization. MUSE outperforms the current state-of-the-art models not only in molecular interaction (protein-protein, drug-protein, and drug-drug) tasks but also in protein interface prediction at the atomic structure scale. More importantly, the multi-scale learning framework shows potential for extension to other scales of computational drug discovery.

Protein language model-embedded geometric graphs power inter-protein contact prediction

Accurate prediction of contacting residue pairs between interacting proteins is very useful for structural characterization of protein-protein interactions. Although significant improvement has been made in inter-protein contact prediction recently, there is still a large room for improving the prediction accuracy. Here we present a new deep learning method referred to as PLMGraph-Inter for inter-protein contact prediction. Specifically, we employ rotationally and translationally invariant geometric graphs obtained from structures of interacting proteins to integrate multiple protein language models, which are successively transformed by graph encoders formed by geometric vector perceptrons and residual networks formed by dimensional hybrid residual blocks to predict inter-protein contacts. Extensive evaluation on multiple test sets illustrates that PLMGraph-Inter outperforms five top inter-protein contact prediction methods, including DeepHomo, GLINTER, CDPred, DeepHomo2, and DRN-1D2D_Inter, by large margins. In addition, we also show that the prediction of PLMGraph-Inter can complement the result of AlphaFold-Multimer. Finally, we show leveraging the contacts predicted by PLMGraph-Inter as constraints for protein-protein docking can dramatically improve its performance for protein complex structure prediction.

Deep learning in modeling protein complex structures: From contact prediction to end-to-end approaches

Protein-protein interactions play crucial roles in many biological processes. Traditionally, protein complex structures are normally built by protein-protein docking. With the rapid development of artificial intelligence and its great success in monomer protein structure prediction, deep learning has widely been applied to modeling protein-protein complex structures through inter-protein contact prediction and end-to-end approaches in the past few years. This article reviews the recent advances of deep-learning-based approaches in modeling protein-protein complex structures as well as their advantages and limitations. Challenges and possible future directions are also briefly discussed in applying deep learning for the prediction of protein complex structures.

Flexible protein-protein docking with a multitrack iterative transformer

Conventional protein-protein docking algorithms usually rely on heavy candidate sampling and reranking, but these steps are time-consuming and hinder applications that require high-throughput complex structure prediction, for example, structure-based virtual screening. Existing deep learning methods for protein-protein docking, despite being much faster, suffer from low docking success rates. In addition, they simplify the problem to assume no conformational changes within any protein upon binding (rigid docking). This assumption precludes applications when binding-induced conformational changes play a role, such as allosteric inhibition or docking from uncertain unbound model structures. To address these limitations, we present GeoDock, a multitrack iterative transformer network to predict a docked structure from separate docking partners. Unlike deep learning models for protein structure prediction that input multiple sequence alignments, GeoDock inputs just the sequences and structures of the docking partners, which suits the tasks when the individual structures are given. GeoDock is flexible at the protein residue level, allowing the prediction of conformational changes upon binding. On the Database of Interacting Protein Structures (DIPS) test set, GeoDock achieves a 43% top-1 success rate, outperforming all other tested methods. However, in the standard DIPS train/test splits, we discovered contamination of close homologs in the training set. After decontaminating the training set, the success rate is 31%. On the DB5.5 test set and a benchmark dataset of antibody-antigen complexes, GeoDock outperforms the deep learning models trained using the same dataset but falls behind most of the conventional methods and AlphaFold-Multimer. GeoDock attains an average inference speed of under 1 s on a single GPU, enabling its application in large-scale structure screening. Although binding-induced conformational changes are still a challenge owing to limited training and evaluation data, our architecture sets up the foundation to capture this backbone flexibility. Code and a demonstration Jupyter notebook are available at https://github.com/Graylab/GeoDock.

Combining pairwise structural similarity and deep learning interface contact prediction to estimate protein complex model accuracy in CASP15

Estimating the accuracy of quaternary structural models of protein complexes and assemblies (EMA) is important for predicting quaternary structures and applying them to studying protein function and interaction. The pairwise similarity between structural models is proven useful for estimating the quality of protein tertiary structural models, but it has been rarely applied to predicting the quality of quaternary structural models. Moreover, the pairwise similarity approach often fails when many structural models are of low quality and similar to each other. To address the gap, we developed a hybrid method (MULTICOM_qa) combining a pairwise similarity score (PSS) and an interface contact probability score (ICPS) based on the deep learning inter-chain contact prediction for estimating protein complex model accuracy. It blindly participated in the 15th Critical Assessment of Techniques for Protein Structure Prediction (CASP15) in 2022 and performed very well in estimating the global structure accuracy of assembly models. The average per-target correlation coefficient between the model quality scores predicted by MULTICOM_qa and the true quality scores of the models of CASP15 assembly targets is 0.66. The average per-target ranking loss in using the predicted quality scores to rank the models is 0.14. It was able to select good models for most targets. Moreover, several key factors (i.e., target difficulty, model sampling difficulty, skewness of model quality, and similarity between good/bad models) for EMA are identified and analyzed. The results demonstrate that combining the multi-model method (PSS) with the complementary single-model method (ICPS) is a promising approach to EMA.

ZeroBind: a protein-specific zero-shot predictor with subgraph matching for drug-target interactions

Existing drug-target interaction (DTI) prediction methods generally fail to generalize well to novel (unseen) proteins and drugs. In this study, we propose a protein-specific meta-learning framework ZeroBind with subgraph matching for predicting protein-drug interactions from their structures. During the meta-training process, ZeroBind formulates training a protein-specific model, which is also considered a learning task, and each task uses graph neural networks (GNNs) to learn the protein graph embedding and the molecular graph embedding. Inspired by the fact that molecules bind to a binding pocket in proteins instead of the whole protein, ZeroBind introduces a weakly supervised subgraph information bottleneck (SIB) module to recognize the maximally informative and compressive subgraphs in protein graphs as potential binding pockets. In addition, ZeroBind trains the models of individual proteins as multiple tasks, whose importance is automatically learned with a task adaptive self-attention module to make final predictions. The results show that ZeroBind achieves superior performance on DTI prediction over existing methods, especially for those unseen proteins and drugs, and performs well after fine-tuning for those proteins or drugs with a few known binding partners.

Enhancing alphafold-multimer-based protein complex structure prediction with MULTICOM in CASP15

To enhance the AlphaFold-Multimer-based protein complex structure prediction, we developed a quaternary structure prediction system (MULTICOM) to improve the input fed to AlphaFold-Multimer and evaluate and refine its outputs. MULTICOM samples diverse multiple sequence alignments (MSAs) and templates for AlphaFold-Multimer to generate structural predictions by using both traditional sequence alignments and Foldseek-based structure alignments, ranks structural predictions through multiple complementary metrics, and refines the structural predictions via a Foldseek structure alignment-based refinement method. The MULTICOM system with different implementations was blindly tested in the assembly structure prediction in the 15th Critical Assessment of Techniques for Protein Structure Prediction (CASP15) in 2022 as both server and human predictors. MULTICOM_qa ranked 3rd among 26 CASP15 server predictors and MULTICOM_human ranked 7th among 87 CASP15 server and human predictors. The average TM-score of the first predictions submitted by MULTICOM_qa for CASP15 assembly targets is ~0.76, 5.3% higher than ~0.72 of the standard AlphaFold-Multimer. The average TM-score of the best of top 5 predictions submitted by MULTICOM_qa is ~0.80, about 8% higher than ~0.74 of the standard AlphaFold-Multimer. Moreover, the Foldseek Structure Alignment-based Multimer structure Generation (FSAMG) method outperforms the widely used sequence alignment-based multimer structure generation.

Integrating unsupervised language model with multi-view multiple sequence alignments for high-accuracy inter-chain contact prediction

Accurate identification of inter-chain contacts in the protein complex is critical to determine the corresponding 3D structures and understand the biological functions. We proposed a new deep learning method, ICCPred, to deduce the inter-chain contacts from the amino acid sequences of the protein complex. This pipeline was built on the designed deep residual network architecture, integrating the pre-trained language model with three multiple sequence alignments (MSAs) from different biological views. Experimental results on 709 non-redundant benchmarking protein complexes showed that the proposed ICCPred significantly increased inter-chain contact prediction accuracy compared to the state-of-the-art approaches. Detailed data analyses showed that the significant advantage of ICCPred lies in the utilization of pre-trained transformer language models which can effectively extract the complementary co-evolution diversity from three MSAs. Meanwhile, the designed deep residual network enhances the correlation between the co-evolution diversity and the patterns of inter-chain contacts. These results demonstrated a new avenue for high-accuracy deep-learning inter-chain contact prediction that is applicable to large-scale protein-protein interaction annotations from sequence alone.

Co-evolution at protein-protein interfaces guides inference of stoichiometry of oligomeric protein complexes by de novo structure prediction

The quaternary structure with specific stoichiometry is pivotal to the specific function of protein complexes. However, determining the structure of many protein complexes experimentally remains a major bottleneck. Structural bioinformatics approaches, such as the deep learning algorithm Alphafold2-multimer (AF2-multimer), leverage the co-evolution of amino acids and sequence-structure relationships for accurate de novo structure and contact prediction. Pseudo-likelihood maximization direct coupling analysis (plmDCA) has been used to detect co-evolving residue pairs by statistical modeling. Here, we provide evidence that combining both methods can be used for de novo prediction of the quaternary structure and stoichiometry of a protein complex. We achieve this by augmenting the existing AF2-multimer confidence metrics with an interpretable score to identify the complex with an optimal fraction of native contacts of co-evolving residue pairs at intermolecular interfaces. We use this strategy to predict the quaternary structure and non-trivial stoichiometries of Bacillus subtilis spore germination protein complexes with unknown structures. Co-evolution at intermolecular interfaces may therefore synergize with AI-based de novo quaternary structure prediction of structurally uncharacterized bacterial protein complexes.

Improving AlphaFold2-based protein tertiary structure prediction with MULTICOM in CASP15

Since the 14th Critical Assessment of Techniques for Protein Structure Prediction (CASP14), AlphaFold2 has become the standard method for protein tertiary structure prediction. One remaining challenge is to further improve its prediction. We developed a new version of the MULTICOM system to sample diverse multiple sequence alignments (MSAs) and structural templates to improve the input for AlphaFold2 to generate structural models. The models are then ranked by both the pairwise model similarity and AlphaFold2 self-reported model quality score. The top ranked models are refined by a novel structure alignment-based refinement method powered by Foldseek. Moreover, for a monomer target that is a subunit of a protein assembly (complex), MULTICOM integrates tertiary and quaternary structure predictions to account for tertiary structural changes induced by protein-protein interaction. The system participated in the tertiary structure prediction in 2022 CASP15 experiment. Our server predictor MULTICOM_refine ranked 3rd among 47 CASP15 server predictors and our human predictor MULTICOM ranked 7th among all 132 human and server predictors. The average GDT-TS score and TM-score of the first structural models that MULTICOM_refine predicted for 94 CASP15 domains are ~0.80 and ~0.92, 9.6% and 8.2% higher than ~0.73 and 0.85 of the standard AlphaFold2 predictor respectively.

Deep transfer learning for inter-chain contact predictions of transmembrane protein complexes

Membrane proteins are encoded by approximately a quarter of human genes. Inter-chain residue-residue contact information is important for structure prediction of membrane protein complexes and valuable for understanding their molecular mechanism. Although many deep learning methods have been proposed to predict the intra-protein contacts or helix-helix interactions in membrane proteins, it is still challenging to accurately predict their inter-chain contacts due to the limited number of transmembrane proteins. Addressing the challenge, here we develop a deep transfer learning method for predicting inter-chain contacts of transmembrane protein complexes, named DeepTMP, by taking advantage of the knowledge pre-trained from a large data set of non-transmembrane proteins. DeepTMP utilizes a geometric triangle-aware module to capture the correct inter-chain interaction from the coevolution information generated by protein language models. DeepTMP is extensively evaluated on a test set of 52 self-associated transmembrane protein complexes, and compared with state-of-the-art methods including DeepHomo2.0, CDPred, GLINTER, DeepHomo, and DNCON2_Inter. It is shown that DeepTMP considerably improves the precision of inter-chain contact prediction and outperforms the existing approaches in both accuracy and robustness.

Recent Advances in Deep Learning for Protein-Protein Interaction Analysis: A Comprehensive Review

Deep learning, a potent branch of artificial intelligence, is steadily leaving its transformative imprint across multiple disciplines. Within computational biology, it is expediting progress in the understanding of Protein-Protein Interactions (PPIs), key components governing a wide array of biological functionalities. Hence, an in-depth exploration of PPIs is crucial for decoding the intricate biological system dynamics and unveiling potential avenues for therapeutic interventions. As the deployment of deep learning techniques in PPI analysis proliferates at an accelerated pace, there exists an immediate demand for an exhaustive review that encapsulates and critically assesses these novel developments. Addressing this requirement, this review offers a detailed analysis of the literature from 2021 to 2023, highlighting the cutting-edge deep learning methodologies harnessed for PPI analysis. Thus, this review stands as a crucial reference for researchers in the discipline, presenting an overview of the recent studies in the field. This consolidation helps elucidate the dynamic paradigm of PPI analysis, the evolution of deep learning techniques, and their interdependent dynamics. This scrutiny is expected to serve as a vital aid for researchers, both well-established and newcomers, assisting them in maneuvering the rapidly shifting terrain of deep learning applications in PPI analysis.

Flexible Protein-Protein Docking with a Multi-Track Iterative Transformer

Conventional protein-protein docking algorithms usually rely on heavy candidate sampling and re-ranking, but these steps are time-consuming and hinder applications that require high-throughput complex structure prediction, e.g., structure-based virtual screening. Existing deep learning methods for protein-protein docking, despite being much faster, suffer from low docking success rates. In addition, they simplify the problem to assume no conformational changes within any protein upon binding (rigid docking). This assumption precludes applications when binding-induced conformational changes play a role, such as allosteric inhibition or docking from uncertain unbound model structures. To address these limitations, we present GeoDock, a multi-track iterative transformer network to predict a docked structure from separate docking partners. Unlike deep learning models for protein structure prediction that input multiple sequence alignments (MSAs), GeoDock inputs just the sequences and structures of the docking partners, which suits the tasks when the individual structures are given. GeoDock is flexible at the protein residue level, allowing the prediction of conformational changes upon binding. For a benchmark set of rigid targets, GeoDock obtains a 41% success rate, outperforming all the other tested methods. For a more challenging benchmark set of flexible targets, GeoDock achieves a similar number of top-model successes as the traditional method ClusPro [1], but fewer than ReplicaDock2 [2]. GeoDock attains an average inference speed of under one second on a single GPU, enabling its application in large-scale structure screening. Although binding-induced conformational changes are still a challenge owing to limited training and evaluation data, our architecture sets up the foundation to capture this backbone flexibility. Code and a demonstration Jupyter notebook are available at https://github.com/Graylab/GeoDock.

PIQLE: protein-protein interface quality estimation by deep graph learning of multimeric interaction geometries

Motivation

Accurate modeling of protein-protein interaction interface is essential for high-quality protein complex structure prediction. Existing approaches for estimating the quality of a predicted protein complex structural model utilize only the physicochemical properties or energetic contributions of the interacting atoms, ignoring evolutionarily information or inter-atomic multimeric geometries, including interaction distance and orientations.

Results

Here, we present PIQLE, a deep graph learning method for protein-protein interface quality estimation. PIQLE leverages multimeric interaction geometries and evolutionarily information along with sequence- and structure-derived features to estimate the quality of individual interactions between the interfacial residues using a multi-head graph attention network and then probabilistically combines the estimated quality for scoring the overall interface. Experimental results show that PIQLE consistently outperforms existing state-of-the-art methods including DProQA, TRScore, GNN-DOVE and DOVE on multiple independent test datasets across a wide range of evaluation metrics. Our ablation study and comparison with the self-assessment module of AlphaFold-Multimer repurposed for protein complex scoring reveal that the performance gains are connected to the effectiveness of the multi-head graph attention network in leveraging multimeric interaction geometries and evolutionary information along with other sequence- and structure-derived features adopted in PIQLE.

Availability and implementation

An open-source software implementation of PIQLE is freely available at https://github.com/Bhattacharya-Lab/PIQLE.

Supplementary information

Supplementary data are available at Bioinformatics Advances online.

Enhancing AlphaFold-Multimer-based Protein Complex Structure Prediction with MULTICOM in CASP15

AlphaFold-Multimer has emerged as the state-of-the-art tool for predicting the quaternary structure of protein complexes (assemblies or multimers) since its release in 2021. To further enhance the AlphaFold-Multimer-based complex structure prediction, we developed a new quaternary structure prediction system (MULTICOM) to improve the input fed to AlphaFold-Multimer and evaluate and refine the outputs generated by AlphaFold2-Multimer. Specifically, MULTICOM samples diverse multiple sequence alignments (MSAs) and templates for AlphaFold-Multimer to generate structural models by using both traditional sequence alignments and new Foldseek-based structure alignments, ranks structural models through multiple complementary metrics, and refines the structural models via a Foldseek structure alignment-based refinement method. The MULTICOM system with different implementations was blindly tested in the assembly structure prediction in the 15th Critical Assessment of Techniques for Protein Structure Prediction (CASP15) in 2022 as both server and human predictors. Our server (MULTICOM_qa) ranked 3rd among 26 CASP15 server predictors and our human predictor (MULTICOM_human) ranked 7th among 87 CASP15 server and human predictors. The average TM-score of the first models predicted by MULTICOM_qa for CASP15 assembly targets is ~0.76, 5.3% higher than ~0.72 of the standard AlphaFold-Multimer. The average TM-score of the best of top 5 models predicted by MULTICOM_qa is ~0.80, about 8% higher than ~0.74 of the standard AlphaFold-Multimer. Moreover, the novel Foldseek Structure Alignment-based Model Generation (FSAMG) method based on AlphaFold-Multimer outperforms the widely used sequence alignment-based model generation. The source code of MULTICOM is available at: https://github.com/BioinfoMachineLearning/MULTICOM3.

Improved inter-protein contact prediction using dimensional hybrid residual networks and protein language models

The knowledge of contacting residue pairs between interacting proteins is very useful for the structural characterization of protein-protein interactions (PPIs). However, accurately identifying the tens of contacting ones from hundreds of thousands of inter-protein residue pairs is extremely challenging, and performances of the state-of-the-art inter-protein contact prediction methods are still quite limited. In this study, we developed a deep learning method for inter-protein contact prediction, which is referred to as DRN-1D2D_Inter. Specifically, we employed pretrained protein language models to generate structural information-enriched input features to residual networks formed by dimensional hybrid residual blocks to perform inter-protein contact prediction. Extensively bechmarking DRN-1D2D_Inter on multiple datasets, including both heteromeric PPIs and homomeric PPIs, we show DRN-1D2D_Inter consistently and significantly outperformed two state-of-the-art inter-protein contact prediction methods, including GLINTER and DeepHomo, although both the latter two methods leveraged the native structures of interacting proteins in the prediction, and DRN-1D2D_Inter made the prediction purely from sequences. We further show that applying the predicted contacts as constraints for protein-protein docking can significantly improve its performance for protein complex structure prediction.

Combining pairwise structural similarity and deep learning interface contact prediction to estimate protein complex model accuracy in CASP15

Estimating the accuracy of quaternary structural models of protein complexes and assemblies (EMA) is important for predicting quaternary structures and applying them to studying protein function and interaction. The pairwise similarity between structural models is proven useful for estimating the quality of protein tertiary structural models, but it has been rarely applied to predicting the quality of quaternary structural models. Moreover, the pairwise similarity approach often fails when many structural models are of low quality and similar to each other. To address the gap, we developed a hybrid method (MULTICOM_qa) combining a pairwise similarity score (PSS) and an interface contact probability score (ICPS) based on the deep learning inter-chain contact prediction for estimating protein complex model accuracy. It blindly participated in the 15th Critical Assessment of Techniques for Protein Structure Prediction (CASP15) in 2022 and ranked first out of 24 predictors in estimating the global accuracy of assembly models. The average per-target correlation coefficient between the model quality scores predicted by MULTICOM_qa and the true quality scores of the models of CASP15 assembly targets is 0.66. The average per-target ranking loss in using the predicted quality scores to rank the models is 0.14. It was able to select good models for most targets. Moreover, several key factors (i.e., target difficulty, model sampling difficulty, skewness of model quality, and similarity between good/bad models) for EMA are identified and analayzed. The results demonstrate that combining the multi-model method (PSS) with the complementary single-model method (ICPS) is a promising approach to EMA. The source code of MULTICOM_qa is available at https://github.com/BioinfoMachineLearning/MULTICOM_qa .

PIQLE: protein-protein interface quality estimation by deep graph learning of multimeric interaction geometries

Accurate modeling of protein-protein interaction interface is essential for high-quality protein complex structure prediction. Existing approaches for estimating the quality of a predicted protein complex structural model utilize only the physicochemical properties or energetic contributions of the interacting atoms, ignoring evolutionarily information or inter-atomic multimeric geometries, including interaction distance and orientations. Here we present PIQLE, a deep graph learning method for protein-protein interface quality estimation. PIQLE leverages multimeric interaction geometries and evolutionarily information along with sequence- and structure-derived features to estimate the quality of the individual interactions between the interfacial residues using a multihead graph attention network and then probabilistically combines the estimated quality of the interfacial residues for scoring the overall interface. Experimental results show that PIQLE consistently outperforms existing state-of-the-art methods on multiple independent test datasets across a wide range of evaluation metrics. Our ablation study reveals that the performance gains are connected to the effectiveness of the multihead graph attention network in leveraging multimeric interaction geometries and evolutionary information along with other sequence- and structure-derived features adopted in PIQLE. An open-source software implementation of PIQLE, licensed under the GNU General Public License v3, is freely available at https://github.com/Bhattacharya-Lab/PIQLE .

Growing ecosystem of deep learning methods for modeling protein-protein interactions

Numerous cellular functions rely on protein-protein interactions. Efforts to comprehensively characterize them remain challenged however by the diversity of molecular recognition mechanisms employed within the proteome. Deep learning has emerged as a promising approach for tackling this problem by exploiting both experimental data and basic biophysical knowledge about protein interactions. Here, we review the growing ecosystem of deep learning methods for modeling protein interactions, highlighting the diversity of these biophysically informed models and their respective trade-offs. We discuss recent successes in using representation learning to capture complex features pertinent to predicting protein interactions and interaction sites, geometric deep learning to reason over protein structures and predict complex structures, and generative modeling to design de novo protein assemblies. We also outline some of the outstanding challenges and promising new directions. Opportunities abound to discover novel interactions, elucidate their physical mechanisms, and engineer binders to modulate their functions using deep learning and, ultimately, unravel how protein interactions orchestrate complex cellular behaviors.

DeepHomo2.0: improved protein-protein contact prediction of homodimers by transformer-enhanced deep learning

Protein-protein interactions play an important role in many biological processes. However, although structure prediction for monomer proteins has achieved great progress with the advent of advanced deep learning algorithms like AlphaFold, the structure prediction for protein-protein complexes remains an open question. Taking advantage of the Transformer model of ESM-MSA, we have developed a deep learning-based model, named DeepHomo2.0, to predict protein-protein interactions of homodimeric complexes by leveraging the direct-coupling analysis (DCA) and Transformer features of sequences and the structure features of monomers. DeepHomo2.0 was extensively evaluated on diverse test sets and compared with eight state-of-the-art methods including protein language model-based, DCA-based and machine learning-based methods. It was shown that DeepHomo2.0 achieved a high precision of >70% with experimental monomer structures and >60% with predicted monomer structures for the top 10 predicted contacts on the test sets and outperformed the other eight methods. Moreover, even the version without using structure information, named DeepHomoSeq, still achieved a good precision of >55% for the top 10 predicted contacts. Integrating the predicted contacts into protein docking significantly improved the structure prediction of realistic Critical Assessment of Protein Structure Prediction homodimeric complexes. DeepHomo2.0 and DeepHomoSeq are available at http://huanglab.phys.hust.edu.cn/DeepHomo2/.

DockNet: high-throughput protein-protein interface contact prediction

MOTIVATION

Over 300 000 protein-protein interaction (PPI) pairs have been identified in the human proteome and targeting these is fast becoming the next frontier in drug design. Predicting PPI sites, however, is a challenging task that traditionally requires computationally expensive and time-consuming docking simulations. A major weakness of modern protein docking algorithms is the inability to account for protein flexibility, which ultimately leads to relatively poor results.

RESULTS

Here, we propose DockNet, an efficient Siamese graph-based neural network method which predicts contact residues between two interacting proteins. Unlike other methods that only utilize a protein's surface or treat the protein structure as a rigid body, DockNet incorporates the entire protein structure and places no limits on protein flexibility during an interaction. Predictions are modeled at the residue level, based on a diverse set of input node features including residue type, surface accessibility, residue depth, secondary structure, pharmacophore and torsional angles. DockNet is comparable to current state-of-the-art methods, achieving an area under the curve (AUC) value of up to 0.84 on an independent test set (DB5), can be applied to a variety of different protein structures and can be utilized in situations where accurate unbound protein structures cannot be obtained.

AVAILABILITY AND IMPLEMENTATION

DockNet is available at https://github.com/npwilliams09/docknet and an easy-to-use webserver at https://biosig.lab.uq.edu.au/docknet. All other data underlying this article are available in the article and in its online supplementary material.

SUPPLEMENTARY INFORMATION

Supplementary data are available at Bioinformatics online.

Ten quick tips for sequence-based prediction of protein properties using machine learning

The ubiquitous availability of genome sequencing data explains the popularity of machine learning-based methods for the prediction of protein properties from their amino acid sequences. Over the years, while revising our own work, reading submitted manuscripts as well as published papers, we have noticed several recurring issues, which make some reported findings hard to understand and replicate. We suspect this may be due to biologists being unfamiliar with machine learning methodology, or conversely, machine learning experts may miss some of the knowledge needed to correctly apply their methods to proteins. Here, we aim to bridge this gap for developers of such methods. The most striking issues are linked to a lack of clarity: how were annotations of interest obtained; which benchmark metrics were used; how are positives and negatives defined. Others relate to a lack of rigor: If you sneak in structural information, your method is not sequence-based; if you compare your own model to "state-of-the-art," take the best methods; if you want to conclude that some method is better than another, obtain a significance estimate to support this claim. These, and other issues, we will cover in detail. These points may have seemed obvious to the authors during writing; however, they are not always clear-cut to the readers. We also expect many of these tips to hold for other machine learning-based applications in biology. Therefore, many computational biologists who develop methods in this particular subject will benefit from a concise overview of what to avoid and what to do instead.

Prediction of inter-chain distance maps of protein complexes with 2D attention-based deep neural networks

Residue-residue distance information is useful for predicting tertiary structures of protein monomers or quaternary structures of protein complexes. Many deep learning methods have been developed to predict intra-chain residue-residue distances of monomers accurately, but few methods can accurately predict inter-chain residue-residue distances of complexes. We develop a deep learning method CDPred (i.e., Complex Distance Prediction) based on the 2D attention-powered residual network to address the gap. Tested on two homodimer datasets, CDPred achieves the precision of 60.94% and 42.93% for top L/5 inter-chain contact predictions (L: length of the monomer in homodimer), respectively, substantially higher than DeepHomo's 37.40% and 23.08% and GLINTER's 48.09% and 36.74%. Tested on the two heterodimer datasets, the top Ls/5 inter-chain contact prediction precision (Ls: length of the shorter monomer in heterodimer) of CDPred is 47.59% and 22.87% respectively, surpassing GLINTER's 23.24% and 13.49%. Moreover, the prediction of CDPred is complementary with that of AlphaFold2-multimer.

ProteinGLUE multi-task benchmark suite for self-supervised protein modeling

Self-supervised language modeling is a rapidly developing approach for the analysis of protein sequence data. However, work in this area is heterogeneous and diverse, making comparison of models and methods difficult. Moreover, models are often evaluated only on one or two downstream tasks, making it unclear whether the models capture generally useful properties. We introduce the ProteinGLUE benchmark for the evaluation of protein representations: a set of seven per-amino-acid tasks for evaluating learned protein representations. We also offer reference code, and we provide two baseline models with hyperparameters specifically trained for these benchmarks. Pre-training was done on two tasks, masked symbol prediction and next sentence prediction. We show that pre-training yields higher performance on a variety of downstream tasks such as secondary structure and protein interaction interface prediction, compared to no pre-training. However, the larger base model does not outperform the smaller medium model. We expect the ProteinGLUE benchmark dataset introduced here, together with the two baseline pre-trained models and their performance evaluations, to be of great value to the field of protein sequence-based property prediction. Availability: code and datasets from https://github.com/ibivu/protein-glue .

HDIContact: a novel predictor of residue-residue contacts on hetero-dimer interfaces via sequential information and transfer learning strategy

Proteins maintain the functional order of cell in life by interacting with other proteins. Determination of protein complex structural information gives biological insights for the research of diseases and drugs. Recently, a breakthrough has been made in protein monomer structure prediction. However, due to the limited number of the known protein structure and homologous sequences of complexes, the prediction of residue-residue contacts on hetero-dimer interfaces is still a challenge. In this study, we have developed a deep learning framework for inferring inter-protein residue contacts from sequential information, called HDIContact. We utilized transfer learning strategy to produce Multiple Sequence Alignment (MSA) two-dimensional (2D) embedding based on patterns of concatenated MSA, which could reduce the influence of noise on MSA caused by mismatched sequences or less homology. For MSA 2D embedding, HDIContact took advantage of Bi-directional Long Short-Term Memory (BiLSTM) with two-channel to capture 2D context of residue pairs. Our comprehensive assessment on the Escherichia coli (E. coli) test dataset showed that HDIContact outperformed other state-of-the-art methods, with top precision of 65.96%, the Area Under the Receiver Operating Characteristic curve (AUROC) of 83.08% and the Area Under the Precision Recall curve (AUPR) of 25.02%. In addition, we analyzed the potential of HDIContact for human-virus protein-protein complexes, by achieving top five precision of 80% on O75475-P04584 related to Human Immunodeficiency Virus. All experiments indicated that our method was a valuable technical tool for predicting inter-protein residue contacts, which would be helpful for understanding protein-protein interaction mechanisms.

OUP accepted manuscript

Motivation

Deep learning has revolutionized protein tertiary structure prediction recently. The cutting-edge deep learning methods such as AlphaFold can predict high-accuracy tertiary structures for most individual protein chains. However, the accuracy of predicting quaternary structures of protein complexes consisting of multiple chains is still relatively low due to lack of advanced deep learning methods in the field. Because interchain residue–residue contacts can be used as distance restraints to guide quaternary structure modeling, here we develop a deep dilated convolutional residual network method (DRCon) to predict interchain residue–residue contacts in homodimers from residue–residue co-evolutionary signals derived from multiple sequence alignments of monomers, intrachain residue–residue contacts of monomers extracted from true/predicted tertiary structures or predicted by deep learning, and other sequence and structural features.

Results

Tested on three homodimer test datasets (Homo_std dataset, DeepHomo dataset and CASP-CAPRI dataset), the precision of DRCon for top L/5 interchain contact predictions (L: length of monomer in a homodimer) is 43.46%, 47.10% and 33.50% respectively at 6 Å contact threshold, which is substantially better than DeepHomo and DNCON2_inter and similar to Glinter. Moreover, our experiments demonstrate that using predicted tertiary structure or intrachain contacts of monomers in the unbound state as input, DRCon still performs well, even though its accuracy is lower than using true tertiary structures in the bound state are used as input. Finally, our case study shows that good interchain contact predictions can be used to build high-accuracy quaternary structure models of homodimers.

Availability and implementation

The source code of DRCon is available at https://github.com/jianlin-cheng/DRCon. The datasets are available at https://zenodo.org/record/5998532#.YgF70vXMKsB.

Supplementary information

Supplementary data are available at Bioinformatics online.