Hui Yu

As a critical step in drug discovery, molecule optimization aims to improve specific properties of lead compounds. Inspired by isosteres (i.e. molecular substructures with similar reactive electron shells), existing AI-based methods have been proposed for molecule optimization. However, they often rely on available matched molecular pairs (MMPs), thus facing two essential challenges. First, the limited diversity of available MMPs leads to biased training and a lack of novelty in the optimized molecules, in both single-objective (SO) and multi-objective (MO) optimization. Second, the low number of available MMPs leads to limited generalization in MO optimization. To overcome the limitations of MMPs, we propose DFRL-Mol, a novel dual-stage reinforcement learning framework. It consists of three main modules: the Structure–Property Analyst (SPA), the Decorator (DCR), and Curriculum-guided Dual-Stage Policy Optimization (CDSPO). SPA and DCR are two LLM-based models, which are responsible for locating the substructures to be optimized and inferring their appropriate isosteric replacement, respectively. CDSPO leverages curriculum-guided reinforcement learning to facilitate collaboration between the two models. DFRL-Mol ensures the generation of diverse molecular representations while effectively overcoming the issues of bias and data sparsity associated with MMPs. The comparisons with state-of-the-art methods demonstrate the enhanced performance of DFRL-Mol in multiple scenarios of molecule optimization, including SO optimization, substructure-constrained (SC) SO optimization, MO optimization, and MO molecule generation. More importantly, detailed analysis of experimental results demonstrates that the models’ behavior is consistent with chemical intuition.

In molecular representation learning (MRL), tokens (e.g., atoms, motifs, and fingerprints) are the basic elements to represent molecules. It is a common practice by using various tokens to enhance the expressive power of Graph Neural Networks (GNNs) on molecular graphs. Although prior GNNs-based methods employing tokens achieve promising performances in drug-drug interaction (DDI) prediction, the influence of the token on the expressiveness of molecular embedding models remains underexplored. To bridge the gap, we provide an axiomatic definition of MRL from a frequency domain perspective, revealing that the model's performance is closely related to the number of tokens and deriving a theoretical upper bound of likelihood-based model convergency. Building on these insights, we propose SimMotifPro, a simple yet efficient motif-based method, for DDI prediction. Specifically, SimMotifPro uses a variant of DeeperGCN encoder and builds a motif-motif knowledge graph to capture motif interconnections. A Motif Ranker module is also introduced to decouple learned representations and differentiate the contributions of selected motifs. Empirically, we demonstrate that SimMotifPro adheres to the properties demonstrated in our theoretical upper bound and validate the general applicability of our theory across different methods. Furthermore, our approach achieves state-of-the-art performance on various benchmarks for DDI prediction. Our codes and checkpoints are available at https://github.com/siriusong/sim_motif_pro.

Artificial Intelligence (AI), particularly Generative AI (GenAI), has revolutionized molecule design by enabling the creation of novel and complex molecular structures, advancing drug discovery, materials science, and synthetic biology. This survey provides a comprehensive exploration of AI-driven molecule design, categorized within and beyond the central dogma of molecular biology. We first introduce the foundational concepts of GenAI, focusing on architectures such as Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), and Diffusion Models (DMs), which underpin molecular design applications. Then, we examine how GenAI facilitates molecule design within the central dogma, addressing DNA, RNA, and protein design. Beyond the central dogma, we explore applications in designing small molecules, lipids, glycans, and other biomaterials. Finally, we conclude by discussing current challenges, emerging trends, and potential future directions for GenAI in molecule design. Overall, this survey aims to equip researchers with a thorough understanding of this rapidly-evolving field and inspire novel developments and breakthroughs in GenAI for molecule design.

Designing enzyme backbones with substrate-specific functionality is a critical challenge in computational protein engineering. Current generative models excel in protein design but face limitations in binding data, substrate-specific control, and flexibility for de novo enzyme backbone generation. To address this, we introduce EnzyBind, a dataset with 11,100 experimentally validated enzyme-substrate pairs specifically curated from PDBbind. Building on this, we propose EnzyControl, a method that enables functional and substrate-specific control in enzyme backbone generation. Our approach generates enzyme backbones conditioned on MSA-annotated catalytic sites and their corresponding substrates, which are automatically extracted from curated enzyme-substrate data. At the core of EnzyControl is EnzyAdapter, a lightweight, modular component integrated into a pretrained motif-scaffolding model, allowing it to become substrate-aware. A two-stage training paradigm further refines the model's ability to generate accurate and functional enzyme structures. Experiments show that our EnzyControl achieves the best performance across structural and functional metrics on EnzyBind and EnzyBench benchmarks, with particularly notable improvements of 13\% in designability and 13\% in catalytic efficiency compared to the baseline models. The code is released at https://github.com/Vecteur-libre/EnzyControl.

Metabolism in vivo turns small molecules (e.g., drugs) into metabolites (new molecules), which brings unexpected safety issues in drug development. However, it is costly to determine metabolites by biological assays. Recent computational methods provide new promising approaches by predicting possible metabolites. Rule-based methods utilize predefined reaction-derived rules to infer metabolites. However, they are powerless to new metabolic reaction patterns. In contrast, rule-free methods leverage sequence-to-sequence machine translation to generate metabolites. Nevertheless, they are insufficient to characterize molecule structures, and bear weak interpretability. To address these issues in rule-free methods, this manuscript proposes a novel metabolism type-aware graph generative framework (MTGGF) for molecular metabolite prediction. It contains a two-stage learning process, including a pre-training on a large general chemical reaction dataset, and a fine-tuning on three smaller type-specific metabolic reaction datasets. Its core, an elaborate graph-to-graph generative model, treats both atoms and bonds as bipartite vertices, and molecules as bipartite graphs, such that it can embed rich information of molecule structures and ensure the integrity of generated metabolite structures. The comparison with state-of-the-art methods demonstrates its superiority. Furthermore, the ablation study validates the contributions of its two graph encoding components and its reaction-type-specific fine-tuning models. More importantly, based on interactive attention between a molecule and its metabolites, the case studies on five approved drugs reveal that there exist crucial substructures specific to metabolism types. It is anticipated that this framework can boost the risk evaluation of drug metabolites. The codes are available at https://github.com/zpczaizheli/Metabolite. Graphical Abstract

It is a vital step to find cocrystal formers of drugs in drug development. Dual-view learning (DVL) has achieved inspiring progress in predicting cocrystals since compounds can be represented in a dual-source manner (i.e., sequence, 2D structures). Nonetheless, it is still an ongoing issue that the performance of existing DVL-based approaches depends on how appropriate the combination of dual view is. Furthermore, there is a need to elucidate what atoms are crucial to form a cocrystal of two compounds. This work holds an assumption that the orthogonal separation of view representations into view-shared representations and view-specific representations can eliminate the redundancy and irrelevant features among dual view. To address these issues, this work elaborates a novel Dual-View Learning framework for predicting Compound Cocrystal (DVL-CC). The framework includes molecule encoders of dual view, a dual-view combinator, and a binary predictor. Especially, the dual-view combinator orthogonally disentangles view-shared and view-specific molecule representations from raw view representations by both an elaborate GAN-based consistency learner and a set of complementary constraints. The comparison with state-of-the-art DVL-based methods demonstrates the superiority of DVL-CC. Also, the comprehensive ablation studies validate and illustrate how its main components contribute to the cocrystal prediction, including individual-view representations, the dual-view combinator, the consistency learner, and the complementary constraints. Furthermore, a case study illustrates the interpretability of DVL-CC by indicating crucial atoms associated with cocrystal conformation patterns between compounds. It's anticipated that this work can boost drug development.The code and data underlying this article are freely available at https://github.com/savior-22/DVL-CC.