cheminformatics

A Nature Machine Intelligence paper on guiding large language models to predict molecular edit sequences for optimizing the synthesizability of candidate drug molecules.

论文提出优化的 MCES 距离计算方法，使结构不相交的数据集划分可在大规模分子库上可行，并发现 MassSpecGym 等基准中单一连通分量占库的 95.5%，导致测试集系统性地偏向更大更复杂的分子，混淆了结构新颖性与外推难度。

该预印本系统比较了在不同推理信息假设（无化学式、已知化学式、峰值带子式注释）下的最近邻基线方法，用于从MS/MS光谱预测分子指纹，评估了五种谱相似度在NPLIB1与MassSpecGym数据集上的表现。

The paper introduces R-GroundBench, a diagnostic benchmark built from real patent Markush structures, to evaluate how well AI models ground variable R-groups across molecular, textual, and symbolic modalities. Results show large gaps between easy recognition and hard grounding tasks, with even chemistry-domain VLMs performing poorly and open-ended Markush generation accuracy remaining below 20%.

OPDiv is an algorithm that optimizes the selection of diverse high-scoring compounds in virtual screening by using integer optimization to balance score and diversity, addressing the trade-off in cheminformatics pipelines.

The paper introduces MFP, a method using role-aware Morgan fingerprints for predicting reaction yields in chemistry, achieving high accuracy and faster training compared to existing methods like YieldBERT and GNAN.

Fraglingo introduces an attachment-aware autoregressive model for molecular design that generates molecules by jointly predicting fragment identity and attachment in a continuous latent space, enhancing property control and flexibility.

GraphNOSE is an open-source graph transformer framework that predicts multi-label odor descriptors from molecular structures, achieving superior performance over existing methods with fewer parameters and improved generalization to out-of-distribution compounds.

A blog post implementing the Extended Connectivity Fingerprint (ECFP) in APL, with references to the original Rogers and Hahn paper and several other resources.

Anthropic launches Claude Science, a desktop app for macOS and Linux that provides a unified research environment for life sciences, integrating AI, databases, HPC, and tools for genomics, proteomics, structural biology, and more.

Google DeepMind has open-sourced Science Skills, a collection of agent skills for scientific research tasks including genomics, structural biology, and cheminformatics, to accelerate agentic workflows with scientific grounding and higher token efficiency.