life sciences

In drug discovery, cross-source bioactivity data (e.g., from ChEMBL) introduce substantial noise due to heterogeneous experimental protocols, degrading molecular activity prediction performance. To address this, we propose a data selection framework tailored to unlabeled test sets with unknown ground-truth labels. First, we quantify each training assay’s contribution to model predictions via data attribution. Second, leveraging assay text descriptions, we fine-tune a language model to jointly encode semantic similarity and biological plausibility, enabling adaptive selection of relevant training assays for unseen test compounds. Evaluated across 12 model–target pairs, our method outperforms strong baselines on 9, yielding significant improvements in predictive accuracy. It effectively filters out detrimental assays, thereby enhancing model robustness and data utilization efficiency without requiring test-set labels.

To address the low coverage (<0.1%) and inability to support dynamic multi-objective optimization with constraints in virtual screening of ultra-large combinatorial synthesis libraries (CSLs, up to 10 billion compounds), this paper proposes a structure-prior-driven approximate exhaustive search framework. Methodologically, it integrates RDKit-based physicochemical property computation with a neural surrogate model specifically designed for CSL topology, enabling minute-scale enumeration and top-k retrieval of millions of compounds on consumer-grade GPUs. This work achieves, for the first time, scalable and task-transferable approximate exhaustive search over ultra-large CSLs, while natively supporting dynamic constraints and joint multi-objective optimization. Experiments on million-compound benchmark libraries demonstrate that our approach significantly outperforms state-of-the-art methods in both retrieval accuracy and speed, consistently identifying highly differentiated compounds.

To address the low efficiency and high manual dependency of the traditional chemical “design–build–test–learn” cycle, this work proposes a novel paradigm of deep collaboration between chemists and AI researchers, enabling an AI-augmented scientific research system for intelligent laboratories. Methodologically, the system integrates machine learning models—employed for experimental data modeling and synthetic route optimization—with large language model (LLM)-driven AI agents that support knowledge retrieval, experimental reasoning, and decision assistance. Its practical utility is validated through three interdisciplinary case studies spanning experimental design, synthesis optimization, and materials characterization. Key contributions include: (i) the first implementation of an LLM-powered closed-loop intelligent experimental decision-making framework; (ii) significant reduction in experimental iteration time and manual data analysis burden; and (iii) advancement of chemical R&D toward a data-driven, human–AI collaborative, digital paradigm.

Efficiently retrieving structurally diverse yet biologically similar (i.e., potency-similar) molecules from ultra-large-scale chemical libraries remains a critical challenge in reverse drug discovery. Method: We propose a target-agnostic, potency-driven small-molecule search engine. Our approach introduces a novel potency-oriented molecular representation paradigm—decoupling similarity assessment from target-specific information. It leverages large-model-pretrained potency embeddings and accelerates similarity search via processor-level SIMD instruction optimization. Further, we design a target-free contrastive learning framework to enhance generalization across diverse bioactivity contexts. Results: Evaluated on the 40-billion-molecule Enamine REAL library, our method achieves millisecond-scale latency with 100% recall—significantly outperforming state-of-the-art baselines. To our knowledge, this is the first work enabling real-time, high-fidelity potency-similarity retrieval over an exascale (10¹⁸) molecular space, establishing a scalable, AI-powered paradigm for target-agnostic reverse drug discovery.

In drug discovery, AI tools are fragmented across isolated platforms with incompatible interfaces and scripting environments, leading to inefficient, redundant workflows. To address this, we propose FROGENT, an end-to-end intelligent agent that—uniquely—integrates, within a unified framework, a dynamic biochemical knowledge base, a heterogeneous toolkit (encompassing molecular generation, virtual screening, synthetic route planning, etc.), and domain-specific AI models. Leveraging large language models (LLMs) and the Model Context Protocol (MCP), FROGENT achieves task understanding, adaptive workflow orchestration, and autonomous decision-making. Evaluated on eight benchmark tasks, FROGENT significantly outperforms existing approaches: it achieves a target–ligand hit-rate retrieval three times higher than the strongest baseline, improves protein–small-molecule interaction prediction accuracy by 100%, and consistently surpasses Qwen3-32B and GPT-4o across all metrics.

This study investigates whether increasing model scale consistently enhances performance in predicting molecular properties and activities for drug discovery. The authors systematically evaluate classical machine learning methods (e.g., Random Forest, ExtraTrees), graph neural networks (GIN), and pretrained molecular sequence models (Ligandformer, MoLFormer, ChemBERTa2) across 22 benchmark tasks partitioned rigorously by structural similarity, introducing a baseline based on structure–activity relationship (SAR) rules. Results show that classical models outperform others in 10 tasks, GNNs in 9, and large-scale models in only 3, indicating that model scale is not the decisive factor—task-specific suitability matters more. This work is the first to reveal this pattern under a large-scale, multitask setting with strict data splits, while also suggesting that large models may hold promise in zero-shot reasoning and hypothesis generation.

Early drug discovery is often hindered by slow, costly, and difficult-to-automate small-molecule synthesis, which impedes the efficiency of the design–make–test–analyze cycle. This work addresses this bottleneck by constructing onepot CORE, an enumerated chemical space encompassing 3.4 billion molecules, and integrating AI-driven reaction feasibility prediction with Phil, an end-to-end automated synthesis platform. The system supports seven common medicinal chemistry transformations and a vendor-based building block library, enabling fully autonomous execution—from retrosynthetic planning and liquid handling to purification—without human intervention. Experimental validation demonstrates that the platform efficiently produces high-purity compounds with NMR-confirmed structures, achieving short cycle times and high success rates. Its utility is further illustrated by the successful synthesis and biological evaluation of a series of DPP4 inhibitors, substantially expanding the accessible and synthesizable chemical space for drug discovery.

Traditional small-molecule drug discovery suffers from low efficiency, lengthy timelines, and difficulty in simultaneously achieving structural novelty and drug-likeness. This work proposes the first semi-autonomous drug discovery system driven by a multimodal AI agent, enabling three-tier adaptive inverse molecular design. The system employs a graph-native generative model to directly construct novel compounds on molecular graphs and integrates physics-informed scoring, Boltz-2 affinity prediction, and ChEMBL-based calibration for optimization. Applied to the BCL6 and EZH2 targets, the system generated over 2,300 entirely new molecules per target, with 91.9% of their Murcko scaffolds absent from ChEMBL. Affinity predictions achieved Spearman correlation coefficients between −0.53 and −0.64 and ROC AUC values of 0.88–0.93, demonstrating substantially enhanced discovery efficiency and structural innovation.

Accurately modeling biomolecular interactions remains a central challenge in drug discovery. This work proposes OpenDDE, an open-source, all-atom foundation model that unifies the modeling of sequence–structure–function relationships through co-folding as a core mechanism, treating structure prediction not as an isolated endpoint but as a shared inference layer. By integrating an all-atom architecture, atom-level implicit reasoning, optimized inference strategies, and large-scale data processing, OpenDDE achieves breakthroughs across four scalable dimensions: data, model, training, and inference, attaining co-folding accuracy on par with IsoDDE. The project releases a complete open-source toolchain—including training code, inference pipelines, model checkpoints, and benchmarking protocols—to enhance reproducibility and foster community collaboration in biomolecular intelligence.

Current AI agents for drug discovery exhibit limited generalization in peptide therapeutics, in vivo pharmacology, and resource-constrained settings, revealing critical capability gaps. This work systematically evaluates six leading agent frameworks across fifteen task categories, combining task taxonomy with knowledge probing experiments to demonstrate that their performance bottlenecks stem primarily from architectural limitations rather than insufficient knowledge—particularly in peptide handling, integration of in vitro and in vivo data, and adaptability under resource constraints. Building on these insights, the study proposes a design specification and capability matrix for next-generation drug discovery agents, emphasizing the integration of protein language models, support for multi-objective optimization, and efficient computational collaboration under realistic operational constraints.