Accurate atomistic biomolecular simulations are vital for understanding disease mechanisms and drug discovery, yet existing methods struggle to balance quantum-mechanical accuracy with computational scalability. Classical force fields often lack precision, while quantum methods are computationally prohibitive for complex biological systems. Here we show that LiTEN, a scalable equivariant neural network, resolves this dilemma by efficiently modeling complex three- and four-body interactions with linear complexity via Linearly Tensorized Quadrangle Attention. We introduce LiTEN-FF, a foundation model pre-trained on extensive datasets to ensure broad chemical generalization across diverse molecular spaces. We demonstrate that LiTEN achieves state-of-the-art accuracy on standard benchmarks, consistently outperforming leading approaches in both precision and speed. Furthermore, LiTEN-FF enables comprehensive modeling tasks, ranging from geometry optimization to free energy surface construction, with high computational efficiency for large biomolecules. This framework provides a physically grounded, versatile foundation for advanced biomolecular modeling and drug design applications.
Qun Su, Kai Zhu, Qiaolin Gou et al.· Nature Communications· 2 citations
Molecular representation learning (MRL) has shown promise in accelerating drug development by predicting chemical properties. However, imperfectly annotation among datasets pose challenges in model design and explainability. In this work, we formulate molecules and corresponding properties as a hypergraph, extracting three key relationships: among properties, molecule-to-property, and among molecules, and developed a unified and explainable multi-task MRL framework, OmniMol. It integrates a task-related meta-information encoder and a task-routed mixture of experts (t-MoE) backbone to capture correlations among properties and produce task-adaptive outputs. To capture underlying physical principles among molecules, we implement an innovative SE(3)-encoder for physical symmetry, applying equilibrium conformation supervision, recursive geometry updates, and scale-invariant message passing to facilitate learning-based conformational relaxation. OmniMol achieves state-of-the-art performance in properties prediction, reaches top performance in chirality-aware tasks, demonstrates explainability for all three relations, and shows effective performance in practical applications. Our code is available in our https://github.com/bowenwang77/OmniMol public repository. AI models for drug discovery often struggle with real-world, incomplete data. Here, the authors present OmniMol, a framework using hypergraphs to improve predictions of molecular properties, addressing challenges of imperfect data annotation and enhancing model explainability.
The study of rare events is one of the major challenges in atomistic simulations, and several enhanced sampling methods toward its solution have been proposed. Recently, it has been suggested that the use of the committor, which provides a precise formal description of rare events, could be of use in this context. We have recently followed up on this suggestion and proposed a committor-based method that promotes frequent transitions between the metastable states of the system and allows extensive sampling of the process transition state ensemble. One of the strengths of our approach is being self-consistent and semiautomatic, exploiting a variational criterion to iteratively optimize a neural-network-based parametrization of the committor, which uses a set of physical descriptors as input. Here, we further automate this procedure by combining our previous method with the expressive power of graph neural networks, which can directly process atomic coordinates rather than descriptors. Besides applications on benchmark systems, we highlight the advantages of a graph-based approach in describing the role of solvent molecules in systems, such as ion pair dissociation or ligand binding.
Peilin Kang, Jintu Zhang, Enrico Trizio et al.· Journal of Chemical Theory a...· 6 citations
Drug repositioning holds promise for discovering new therapeutic applications for existing drugs, accelerating drug development and reducing associated costs. However, current methodologies encounter difficulties in managing diverse network representations, tackling cold start issues, and handling intrinsic attribute representations. Here we introduce a Unified Knowledge-Enhanced deep learning framework for Drug Repositioning (UKEDR), which integrates knowledge graph embedding, pre-training strategies, and recommendation systems to address these challenges. To overcome the cold start issue, UKEDR utilizes a semantic similarity-driven embedding approach. Our evaluations show that UKEDR performs better than various baselines, including classical machine learning, network-based, and deep learning approaches. In cold start scenarios, it demonstrates an improved capability in handling unseen nodes and generalizing to new compounds. The model also demonstrates strong robustness on imbalanced datasets and shows excellent generalization capabilities in specific drug-centric and disease-centric cold-start scenarios, validating its potential for real-world applications. Drug repositioning offers a promising avenue for accelerating drug development, yet existing methods struggle with network diversity, cold start issues, and intrinsic attribute representation. Here, the authors introduce UKEDR, a deep learning framework that integrates knowledge graph embedding and pre-training strategies to overcome the intractable cold start issue, achieving superior performance and interpretability in drug repurposing.
Kun Li, Jiacai Yi, Qing Ye et al.· Communications Chemistry· 1 citation
Reach audiences
Advertise in front of researchers, engineers, and readers.
Reaction representation learning is of paramount importance for adopting deep-learning-based chemistry modeling to solve real-world tasks such as synthesis planning. Most prevailing models are prestrained by self-supervised objectives that rely solely on the chemical structure information. Since structurally similar reactions could possess entirely distinct properties (e.g., reaction yields) and the synthesis-related tasks are highly heterogeneous, there are inherent limitations in constructing a foundational reaction model within the existing approaches. To tackle this limitation, we propose HiCLR, a knowledge-induced hierarchical contrastive learning framework for chemical reactions, by introducing relational inductive bias to forge chemically meaningful and generally applicable reaction fingerprints. Critically, the pretraining scheme combining both retrosynthesis prediction and contrastive loss enables HiCLR to tackle generation-based and understanding-based tasks simultaneously. Comprehensive experiments demonstrate that HiCLR successfully organizes the reaction space into hierarchical global semantic clusters, aligned well with prior knowledge. Consequently, HiCLR is the first foundation model that can be broadly applied to various synthesis-related tasks, and it achieves state-of-the-art performance in reaction classification, reaction condition recommendation, reaction yield prediction, synthesis planning, and even molecular property prediction. HiCLR demonstrates clear benefits in incorporating domain knowledge to guide the learning of neural networks, expediting AI-driven advancements in chemistry.
Jialu Wu, Yiheng Zhu, Xiaorui Wang et al.· JACS Au· 0 citations
Antibodies are crucial for medical applications, yet traditional methods for designing sequences are inefficient. This study introduces AntiBMPNN, an advanced deep‐learning framework that leverages an antibody‐specific 3D dataset, a fine‐tuned message‐passing neural network (MPNN), a frequency‐based scoring function, and AlphaFold 3 to achieve highly accurate antibody sequence design. AntiBMPNN surpasses ProteinMPNN with a perplexity of 1.5 and over 80% sequence recovery. Its scoring function, combined with AlphaFold 3, effectively prioritizes sequences based on structural recovery, positional stability, and biochemical or complex properties. Experimental validation highlights a 75% success rate in single‐point antibody design. AntiBMPNN consistently outperforms AbMPNN, AntiFold, and ProteinMPNN in designing complementarity determining regions (CDR) 1‐3, yielding stronger binding affinities. For CDR1 of huJ3 (anti‐HIV nanobody), it achieves a half maximal effective concentration (EC₅₀) of 9.2 nM (nanomolar), better than ProteinMPNN (135.2 nM) and AntiFold (59.3 nM), and comparable to AbMPNN (6.6 nM). For CDR2 of the D6 nanobody (targeting CD16), AntiBMPNN reaches 0.3 nM, outperforming AbMPNN (2.3 nM), AntiFold (0.7 nM), and ProteinMPNN (0.7 nM). In CDR3 of huJ3, it achieves 1.7 nM, surpassing AbMPNN (51.2 nM), with no detectable activity from AntiFold or ProteinMPNN. These findings confirm that AntiBMPNN‐designed sequences for J3 and D6 outperform the originals, highlighting its potential to improve therapeutic antibody design.
Ze-Yu Sun, Jiayi Yuan, Divya Jaiswal et al.· Advancement of science· 9 citations
Enzymatic reactions play an emerging role in a broad spectrum of scientific and industrial applications. The inherent complexity of enzymes, such as their substrate specificity, conformational flexibility, and the vast diversity of reactions involved, poses substantial challenges for the advanced computational prediction of enzymatic reactions with desirable accuracy. Moreover, existing approaches are mostly tailored for a specific sub-task, such as substrate prediction or binding site annotation, which limits their applicability. In this study, we introduce ERAM, a task-agnostic multimodal learning framework capable of addressing a broad range of downstream applications with both accuracy and efficiency. ERAM aligns pre-trained molecular representations from Protein Language Model with the knowledge of enzyme catalysis by modeling enzymatic reactions as multi-relational data. In enzyme retrieval tasks, ERAM achieves an improvement of 28.31% in mean average precision compared with the state-of-the-art (SOTA) method, CREEP. In substrate prediction tasks, ERAM outperforms the SOTA method ESP, achieving average improvements of 35.53% and 22.97% in Matthews correlation coefficient across two datasets. Additionally, ERAM exhibits commendable interpretability by assigning higher attention weights to binding sites, resulting in lower false-positive rates (42.36%) and higher overlap scores (70.59%) in the unsupervised binding site prediction task compared to RXNAA Mapper. By learning embeddings of substrates, enzymes, and products within a unified knowledge graph latent space, ERAM demonstrates its potential as a versatile and effective tool for enzyme catalysis research.
Accurate prediction of blood-brain barrier permeability (BBBP) is essential for central nervous system drug discovery, yet existing models are often limited by their reliance on predefined physicochemical descriptors, small-molecule-centered training sets, or conformation-dependent representations, which restricts their transferability across chemically diverse modalities especially peptides. In addition, publicly available BBBP datasets remain fragmented, inconsistently standardized, and weakly controlled for molecular redundancy, increasing the risk of data leakage and overestimated model performance. In this study, we propose BBBP-Atlas, a structure-aware BBB permeability prediction model designed for unified modeling of small molecules and peptides with the first cross-modal dataset OmniBBBP. Designed to bypass descriptor and conformation dependencies, our model represents standardized molecular structures as atom-level graphs to capture local atom-bond environments and long-range topological dependencies associated with BBB transport. This design enables direct learning of structure-permeability relationships from molecular topology. For model training and evaluation, we curated a cross-modal, redundancy-filtered database OmniBBBP that seamlessly unifies small molecules and complex peptides, containing 10,218 unique compounds with 9,316 small molecules and 902 peptides. BBBP-Atlas achieved an accuracy of 0.8914 and an MCC of 0.7678 on the independent test set. On a balanced external benchmark of 200 compounds, our model reached an AUC of 0.9108, an accuracy of 0.8500, and an MCC of 0.7000, outperforming LightBBB by an absolute MCC gain of 6%. Case studies further showed that BBBP-Atlas captured clinically meaningful BBB permeability patterns, correctly identifying lorlatinib as BBB-permeable and vancomycin as BBB-impermeable with high confidence. The OmniBBBP-backed BBBP-Atlas offers a versatile and cross-modal approach for single-compound prediction, batch screening, and dataset exploration for CNS drug discovery. BBBP-Atlas is available at https://cadd.drugflow.com/bbbp/.
Xin Shen, Qun Su, Hao Luo et al.· bioRxiv· 0 citations
Graph neural networks (GNNs) have achieved significant success in various graph-related tasks, including node classification, link prediction, and graph classification. However, a critical limitation of many GNN approaches is their dependence on large amounts of labeled data for training. Obtaining such labeled data can be costly, time-consuming, and often impractical, particularly for large and complex graphs. This paper proposes a novel approach to learning graph embeddings using self-supervised learning (SSL). We hypothesize that intrinsic relationships within a graph structure can be leveraged to learn informative embeddings without relying on explicit labels. Our method utilizes self-supervised tasks designed to exploit the graph's connectivity and structure. Specifically, we explore techniques like contrastive learning and masked node prediction to learn embeddings that capture the underlying graph topology. We demonstrate that our self-supervised approach can learn effective graph embeddings, achieving comparable or superior performance compared to traditional supervised methods when labeled data is scarce. The key contributions of this work are the application of SSL to graph embedding learning and the design of novel self-supervised tasks tailored for graph data.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
This paper proposes a novel approach to cybersecurity threat detection leveraging Graph Neural Networks (GNNs) for personalized modeling. Traditional cybersecurity solutions often rely on generic, rule-based systems that struggle to adapt to the diverse and evolving nature of network environments. We introduce a framework where network traffic is represented as a graph, with nodes representing devices and edges representing communication patterns. A GNN is then trained to learn these relationships and identify anomalous patterns specific to each network. The core claim of this work is that by tailoring threat detection models to individual network characteristics, we can significantly improve detection accuracy and reduce false positives. We detail the graph construction process, the GNN architecture, and the training methodology. The resulting system demonstrates the potential to move beyond generic threat detection and provide truly personalized cybersecurity defense.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
This paper introduces Temporal Topology, a novel system for predicting future topology based on event streams. We propose a Temporal Graph Network, a neural network trained to learn geometric patterns from sequences of events, allowing for dynamic analysis and predictive modeling. The system's core claim centers around leveraging real-time data to create a system that goes beyond static geometric analysis, incorporating a dynamic, predictive element. This approach addresses the limitations of traditional geometric modeling by incorporating temporal dynamics, offering a more responsive and insightful approach to topology analysis. We provide a detailed overview of the Temporal Graph Network architecture, training methodology, and preliminary results demonstrating its effectiveness in generating and analyzing geometric patterns.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
This paper proposes a novel approach to system security analysis leveraging Graph Neural Networks (GNNs). Traditional security assessments often rely on static analysis and may fail to capture the dynamic and interconnected nature of complex systems. We argue that system vulnerabilities can be effectively modeled as graph problems, where nodes represent system components and edges represent their relationships. GNNs are then employed to learn representations of these nodes based on their connections, enabling the prediction of potential security risks. This approach offers improved accuracy compared to conventional methods by incorporating contextual information and learning complex dependencies within the system. The core claim is that system security problems can be modeled as graph problems, utilizing GNNs to analyze system vulnerabilities. The key mechanism involves representing system components and relationships as a graph, and employing GNNs to learn node relationships and predict security risks. This work contributes to a more proactive and comprehensive understanding of system security.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
Assistant Professor Pat Pataranutaporn describes a new interface that lets everyday users glimpse inside an AI's neural network before their chatbot ever says a word.
MIT News · Artificial Intelligence· news.mit.eduJul 6, 2026
PhD student Rachel Sava, winner of the Envisioning the Future of Computing Prize, explores transformative improvements and dystopian risks of neural technology.
