Deep Learning for Protein-Ligand Docking: Are We There Yet?

The effects of ligand binding on protein structures and their in vivo functions carry numerous implications for modern biomedical research and biotechnology development efforts such as drug discovery. Although several deep learning (DL) methods and benchmarks designed for protein-ligand docking have recently been introduced, to date no prior works have systematically studied the behavior of docking methods within the practical context of (1) predicted (apo) protein structures, (2) multiple ligands concurrently binding to a given target protein, and (3) having no prior knowledge of binding pockets. To enable a deeper understanding of docking methods' real-world utility, we introduce PoseBench, the first comprehensive benchmark for practical protein-ligand docking. PoseBench enables researchers to rigorously and systematically evaluate DL docking methods for apo-to-holo protein-ligand docking and protein-ligand structure generation using both single and multi-ligand benchmark datasets, the latter of which we introduce for the first time to the DL community. Empirically, using PoseBench, we find that all recent DL docking methods but one fail to generalize to multi-ligand protein targets and also that template-based docking algorithms perform equally well or better for multi-ligand docking as recent single-ligand DL docking methods, suggesting areas of improvement for future work. Code, data, tutorials, and benchmark results are available at https://github.com/BioinfoMachineLearning/PoseBench.

Cryo2StructData: A Large Labeled Cryo-EM Density Map Dataset for AI-based Modeling of Protein Structures.

The advent of single-particle cryo-electron microscopy (cryo-EM) has brought forth a new era of structural biology, enabling the routine determination of large biological molecules and their complexes at atomic resolution. The high-resolution structures of biological macromolecules and their complexes significantly expedite biomedical research and drug discovery. However, automatically and accurately building atomic models from high-resolution cryo-EM density maps is still time-consuming and challenging when template-based models are unavailable. Artificial intelligence (AI) methods such as deep learning trained on limited amount of labeled cryo-EM density maps generate inaccurate atomic models. To address this issue, we created a dataset called Cryo2StructData consisting of 7,600 preprocessed cryo-EM density maps whose voxels are labelled according to their corresponding known atomic structures for training and testing AI methods to build atomic models from cryo-EM density maps. Cryo2StructData is larger than existing, publicly available datasets for training AI methods to build atomic protein structures from cryo-EM density maps. We trained and tested deep learning models on Cryo2StructData to validate its quality showing that it is ready for being used to train and test AI methods for building atomic models.

Outcomes of the EMDataResource Cryo-EM Ligand Modeling Challenge.

The EMDataResource Ligand Model Challenge aimed to assess the reliability and reproducibility of modeling ligands bound to protein and protein/nucleic-acid complexes in cryogenic electron microscopy (cryo-EM) maps determined at near-atomic (1.9-2.5 Å) resolution. Three published maps were selected as targets: E. coli beta-galactosidase with inhibitor, SARS-CoV-2 RNA-dependent RNA polymerase with covalently bound nucleotide analog, and SARS-CoV-2 ion channel ORF3a with bound lipid. Sixty-one models were submitted from 17 independent research groups, each with supporting workflow details. We found that (1) the quality of submitted ligand models and surrounding atoms varied, as judged by visual inspection and quantification of local map quality, model-to-map fit, geometry, energetics, and contact scores, and (2) a composite rather than a single score was needed to assess macromolecule+ligand model quality. These observations lead us to recommend best practices for assessing cryo-EM structures of liganded macromolecules reported at near-atomic resolution.

De Novo Atomic Protein Structure Modeling for Cryo-EM Density Maps Using 3D Transformer and Hidden Markov Model.

Accurately building three-dimensional (3D) atomic structures from 3D cryo-electron microscopy (cryo-EM) density maps is a crucial step in the cryo-EM-based determination of the structures of protein complexes. Despite improvements in the resolution of 3D cryo-EM density maps, the de novo conversion of density maps into 3D atomic structures for protein complexes that do not have accurate homologous or predicted structures to be used as templates remains a significant challenge. Here, we introduce Cryo2Struct, a fully automated ab initio cryo-EM structure modeling method that utilizes a 3D transformer to identify atoms and amino acid types in cryo-EM density maps first, and then employs a novel Hidden Markov Model (HMM) to connect predicted atoms to build backbone structures of proteins. Tested on a standard test dataset of 128 cryo-EM density maps with varying resolutions (2.1 - 5.6 °A) and different numbers of residues (730 - 8,416), Cryo2Struct built substantially more accurate and complete protein structural models than the widely used ab initio method - Phenix in terms of multiple evaluation metrics. Moreover, on a new test dataset of 500 recently released density maps with varying resolutions (1.9 - 4.0 °A) and different numbers of residues (234 - 8,828), it built more accurate models than on the standard dataset. And its performance is rather robust against the change of the resolution of density maps and the size of protein structures.

Cryo2StructData: A Large Labeled Cryo-EM Density Map Dataset for AI-based Modeling of Protein Structures.

The advent of single-particle cryo-electron microscopy (cryo-EM) has brought forth a new era of structural biology, enabling the routine determination of large biological molecules and their complexes at atomic resolution. The high-resolution structures of biological macromolecules and their complexes significantly expedite biomedical research and drug discovery. However, automatically and accurately building atomic models from high-resolution cryo-EM density maps is still time-consuming and challenging when template-based models are unavailable. Artificial intelligence (AI) methods such as deep learning trained on limited amount of labeled cryo-EM density maps generate inaccurate atomic models. To address this issue, we created a dataset called Cryo2StructData consisting of 7,600 preprocessed cryo-EM density maps whose voxels are labelled according to their corresponding known atomic structures for training and testing AI methods to build atomic models from cryo-EM density maps. It is larger and of higher quality than any existing, publicly available dataset. We trained and tested deep learning models on Cryo2StructData to make sure it is ready for the large-scale development of AI methods for building atomic models from cryo-EM density maps.

Impact of AlphaFold on structure prediction of protein complexes: The CASP15-CAPRI experiment.

We present the results for CAPRI Round 54, the 5th joint CASP-CAPRI protein assembly prediction challenge. The Round offered 37 targets, including 14 homodimers, 3 homo-trimers, 13 heterodimers including 3 antibody-antigen complexes, and 7 large assemblies. On average ~70 CASP and CAPRI predictor groups, including more than 20 automatics servers, submitted models for each target. A total of 21 941 models submitted by these groups and by 15 CAPRI scorer groups were evaluated using the CAPRI model quality measures and the DockQ score consolidating these measures. The prediction performance was quantified by a weighted score based on the number of models of acceptable quality or higher submitted by each group among their five best models. Results show substantial progress achieved across a significant fraction of the 60+ participating groups. High-quality models were produced for about 40% of the targets compared to 8% two years earlier. This remarkable improvement is due to the wide use of the AlphaFold2 and AlphaFold2-Multimer software and the confidence metrics they provide. Notably, expanded sampling of candidate solutions by manipulating these deep learning inference engines, enriching multiple sequence alignments, or integration of advanced modeling tools, enabled top performing groups to exceed the performance of a standard AlphaFold2-Multimer version used as a yard stick. This notwithstanding, performance remained poor for complexes with antibodies and nanobodies, where evolutionary relationships between the binding partners are lacking, and for complexes featuring conformational flexibility, clearly indicating that the prediction of protein complexes remains a challenging problem.

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.

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 .

Deep learning for reconstructing protein structures from cryo-EM density maps: Recent advances and future directions.

Cryo-Electron Microscopy (cryo-EM) has emerged as a key technology to determine the structure of proteins, particularly large protein complexes and assemblies in recent years. A key challenge in cryo-EM data analysis is to automatically reconstruct accurate protein structures from cryo-EM density maps. In this review, we briefly overview various deep learning methods for building protein structures from cryo-EM density maps, analyze their impact, and discuss the challenges of preparing high-quality data sets for training deep learning models. Looking into the future, more advanced deep learning models of effectively integrating cryo-EM data with other sources of complementary data such as protein sequences and AlphaFold-predicted structures need to be developed to further advance the field.

Improving Protein-Ligand Interaction Modeling with cryo-EM Data, Templates, and Deep Learning in 2021 Ligand Model Challenge.

Elucidating protein-ligand interaction is crucial for studying the function of proteins and compounds in an organism and critical for drug discovery and design. The problem of protein-ligand interaction is traditionally tackled by molecular docking and simulation, which is based on physical forces and statistical potentials and cannot effectively leverage cryo-EM data and existing protein structural information in the protein-ligand modeling process. In this work, we developed a deep learning bioinformatics pipeline (DeepProLigand) to predict protein-ligand interactions from cryo-EM density maps of proteins and ligands. DeepProLigand first uses a deep learning method to predict the structure of proteins from cryo-EM maps, which is averaged with a reference (template) structure of the proteins to produce a combined structure to add ligands. The ligands are then identified and added into the structure to generate a protein-ligand complex structure, which is further refined. The method based on the deep learning prediction and template-based modeling was blindly tested in the 2021 EMDataResource Ligand Challenge and was ranked first in fitting ligands to cryo-EM density maps. These results demonstrate that the deep learning bioinformatics approach is a promising direction for modeling protein-ligand interactions on cryo-EM data using prior structural information.

High-Performance Deep Learning Toolbox for Genome-Scale Prediction of Protein Structure and Function.

Computational biology is one of many scientific disciplines ripe for innovation and acceleration with the advent of high-performance computing (HPC). In recent years, the field of machine learning has also seen significant benefits from adopting HPC practices. In this work, we present a novel HPC pipeline that incorporates various machine-learning approaches for structure-based functional annotation of proteins on the scale of whole genomes. Our pipeline makes extensive use of deep learning and provides computational insights into best practices for training advanced deep-learning models for high-throughput data such as proteomics data. We showcase methodologies our pipeline currently supports and detail future tasks for our pipeline to envelop, including large-scale sequence comparison using SAdLSA and prediction of protein tertiary structures using AlphaFold2.