Geometric deep learning has recently achieved great success in non-Euclidean domains, and learning on 3D structures of large biomolecules is emerging as a distinct research area. However, its efficacy is largely constrained due to the limited quantity of structural data. Meanwhile, protein language models trained on substantial 1D sequences have shown burgeoning capabilities with scale in a broad range of applications. Several preceding studies consider combining these different protein modalities to promote the representation power of geometric neural networks but fail to present a comprehensive understanding of their benefits. In this work, we integrate the knowledge learned by well-trained protein language models into several state-of-the-art geometric networks and evaluate a variety of protein representation learning benchmarks, including protein-protein interface prediction, model quality assessment, protein-protein rigid-body docking, and binding affinity prediction. Our findings show an overall improvement of 20% over baselines. Strong evidence indicates that the incorporation of protein language models’ knowledge enhances geometric networks’ capacity by a significant margin and can be generalized to complex tasks.

Physisorption relying on crystalline porous materials offers prospective avenues for sustainable separation processes, greenhouse gas capture, and energy storage. However, the lack of end-to-end deep learning model for adsorption prediction confines the rapid and precise screen of crystalline porous materials. Here, we present DeepSorption, a spatial atom interaction learning network that realizes accurate, fast, and direct structure-adsorption prediction with only information of atomic coordinate and chemical element types. The breakthrough in prediction is attributed to the awareness of global structure and local spatial atom interactions endowed by the developed Matformer, which provides the intuitive visualization of atomic-level thinking and executing trajectory in crystalline porous materials prediction. Complete adsorption curves prediction could be performed using DeepSorption with a higher accuracy than Grand canonical Monte Carlo simulation and other machine learning models, a 20-35% decline in the mean absolute error compared to graph neural network CGCNN and machine learning models based on descriptors. Since the established direct associations between raw structure and target functions are based on the understanding of the fundamental chemistry of interatomic interactions, the deep learning network is rationally universal in predicting the different physicochemical properties of various crystalline materials.

In recent years, the application of deep learning models to protein-ligand docking and affinity prediction, both vital for structure-based drug design, has garnered increasing interest. However, many of these models overlook the intricate modeling of interactions between ligand and protein atoms in the complex, consequently limiting their capacity for generalization and interpretability. In this work, we propose Interformer, a unified model built upon the Graph-Transformer architecture. The proposed model is designed to capture non-covalent interactions utilizing an interaction-aware mixture density network. Additionally, we introduce a negative sampling strategy, facilitating an effective correction of interaction distribution for affinity prediction. Experimental results on widely used and our in-house datasets demonstrate the effectiveness and universality of the proposed approach. Extensive analyses confirm our claim that our approach improves performance by accurately modeling specific protein-ligand interactions. Encouragingly, our approach advances docking tasks state-of-the-art (SOTA) performance.

Graph neural networks (GNNs) rely mainly on the message-passing paradigm to propagate node features and build interactions, and different graph learning problems require different ranges of node interactions. In this work, we explore the capacity of GNNs to capture node interactions under contexts of different complexities. We discover that GNNs usually fail to capture the most informative kinds of interaction styles for diverse graph learning tasks , and thus name this phenomenon as GNNs' representation bottleneck. As a response, we demonstrate that the inductive bias introduced by existing graph construction mechanisms can result in this representation bottleneck, i.e. , preventing GNNs from learning interactions of the most appropriate complexity. To address that limitation, we propose a novel graph rewiring approach based on interaction patterns learned by GNNs to adjust each node's receptive fields dynamically. Extensive experiments on both real-world and synthetic datasets prove the effectiveness of our algorithm in alleviating the representation bottleneck and its superiority in enhancing the performance of GNNs over state-of-the-art graph rewiring baselines.

Fang Wu, Vijay Prakash Dwivedi, Jure Leskovec. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Antimicrobial resistance (AMR) is projected to cause up to 10 million deaths annually by 2050, underscoring the urgent need for new antibiotics. Here we present ApexAmphion, a deep-learning framework for de novo design of antibiotics that couples a 6.4-billion-parameter protein language model with reinforcement learning. The model is first fine-tuned on curated peptide data to capture antimicrobial sequence regularities, then optimised with proximal policy optimization against a composite reward that combines predictions from a learned minimum inhibitory concentration (MIC) classifier with differentiable physicochemical objectives. In vitro evaluation of 100 designed peptides showed low MIC values (nanomolar range in some cases) for all candidates (100% hit rate). Moreover, 99 our of 100 compounds exhibited broad-spectrum antimicrobial activity against at least two clinically relevant bacteria. The lead molecules killed bacteria primarily by potently targeting the cytoplasmic membrane. By unifying generation, scoring and multi-objective optimization with deep reinforcement learning in a single pipeline, our approach rapidly produces diverse, potent candidates, offering a scalable route to peptide antibiotics and a platform for iterative steering toward potency and developability within hours.