Molecular docking, a technique for predicting ligand binding poses, is crucial in structure-based drug design for understanding protein-ligand interactions. Recent advancements in docking methods, particularly those leveraging geometric deep learning (GDL), have demonstrated significant efficiency and accuracy advantages over traditional sampling methods. Despite these advancements, current methods are often tailored for specific docking settings, and limitations such as the neglect of protein side-chain structures, difficulties in handling large binding pockets, and challenges in predicting physically valid structures exist. To accommodate various docking settings and achieve accurate, efficient, and physically reliable docking, we propose a novel two-stage docking framework, DeltaDock, consisting of pocket prediction and site-specific docking. We innovatively reframe the pocket prediction task as a pocket-ligand alignment problem rather than direct prediction in the first stage. Then we follow a bi-level coarse-to-fine iterative refinement process to perform site-specific docking. Comprehensive experiments demonstrate the superior performance of DeltaDock. Notably, in the blind docking setting, DeltaDock achieves a 31\% relative improvement over the docking success rate compared with the previous state-of-the-art GDL model DiffDock. With the consideration of physical validity, this improvement increases to about 300\%.

BACKGROUND: The liquid-liquid phase separation (LLPS) of biomolecules in cell underpins the formation of membraneless organelles, which are the condensates of protein, nucleic acid, or both, and play critical roles in cellular function. Dysregulation of LLPS is implicated in a number of diseases. Although the LLPS of biomolecules has been investigated intensively in recent years, the knowledge of the prevalence and distribution of phase separation proteins (PSPs) is still lag behind. Development of computational methods to predict PSPs is therefore of great importance for comprehensive understanding of the biological function of LLPS. RESULTS: Based on the PSPs collected in LLPSDB, we developed a sequence-based prediction tool for LLPS proteins (PSPredictor), which is an attempt at general purpose of PSP prediction that does not depend on specific protein types. Our method combines the componential and sequential information during the protein embedding stage, and, adopts the machine learning algorithm for final predicting. The proposed method achieves a tenfold cross-validation accuracy of 94.71%, and outperforms previously reported PSPs prediction tools. For further applications, we built a user-friendly PSPredictor web server ( http://www.pkumdl.cn/PSPredictor ), which is accessible for prediction of potential PSPs. CONCLUSIONS: PSPredictor could identifie novel scaffold proteins for stress granules and predict PSPs candidates in the human genome for further study. For further applications, we built a user-friendly PSPredictor web server ( http://www.pkumdl.cn/PSPredictor ), which provides valuable information for potential PSPs recognition.

The COVID-19 pandemic calls for rapid development of effective treatments. Although various drug repurpose approaches have been used to screen the FDA-approved drugs and drug candidates in clinical phases against SARS-CoV-2, the coronavirus that causes this disease, no magic bullets have been found until now. In this study, we used directed message passing neural network to first build a broad-spectrum anti-beta-coronavirus compound prediction model, which gave satisfactory predictions on newly reported active compounds against SARS-CoV-2. Then, we applied transfer learning to fine-tune the model with the recently reported anti-SARS-CoV-2 compounds and derived a SARS-CoV-2 specific prediction model COVIDVS-3. We used COVIDVS-3 to screen a large compound library with 4.9 million drug-like molecules from ZINC15 database and recommended a list of potential anti-SARS-CoV-2 compounds for further experimental testing. As a proof-of-concept, we experimentally tested seven high-scored compounds that also demonstrated good binding strength in docking studies against the 3C-like protease of SARS-CoV-2 and found one novel compound that can inhibit the enzyme. Our model is highly efficient and can be used to screen large compound databases with millions or more compounds to accelerate the drug discovery process for the treatment of COVID-19.

BACKGROUND: Protein-protein interactions (PPIs) are critical for many biological processes. It is therefore important to develop accurate high-throughput methods for identifying PPI to better understand protein function, disease occurrence, and therapy design. Though various computational methods for predicting PPI have been developed, their robustness for prediction with external datasets is unknown. Deep-learning algorithms have achieved successful results in diverse areas, but their effectiveness for PPI prediction has not been tested. RESULTS: We used a stacked autoencoder, a type of deep-learning algorithm, to study the sequence-based PPI prediction. The best model achieved an average accuracy of 97.19% with 10-fold cross-validation. The prediction accuracies for various external datasets ranged from 87.99% to 99.21%, which are superior to those achieved with previous methods. CONCLUSIONS: To our knowledge, this research is the first to apply a deep-learning algorithm to sequence-based PPI prediction, and the results demonstrate its potential in this field.