Programming Drugs with Generative Biology Stacks

Описание: NotebookLM: "The 2025 update of the AlphaFold Protein Structure Database (AFDB) represents a significant leap in synchronizing high-accuracy structural predictions with the latest biological data from UniProt. This release broadens the scientific scope of the repository by incorporating isoform-specific predictions and the underlying multiple sequence alignments, providing researchers with deeper insights into protein diversity and evolution. A major technical highlight is the integration of the Encyclopedia of Domains (TED), which adds over 361 million domain annotations and quality metrics to help users interpret functional architecture. To make this vast wealth of information more accessible, the platform features a redesigned user interface that pairs interactive 3D visualizations with organized data tabs. Ultimately, these enhancements aim to democratize structural biology by providing an open-access, sustainable resource that supports everything from drug discovery to fundamental life science research."

"ODesign is introduced as a pioneering biomolecular world model designed to move beyond traditional protein-centric frameworks by enabling the cross-modality generation of proteins, nucleic acids, and small molecules. By utilizing a unified all-atom architecture and a sophisticated hierarchical masking strategy, the model can perform diverse tasks ranging from de novo design to precise motif scaffolding across different molecular types. A core breakthrough of the system is its computational efficiency, achieving an order-of-magnitude improvement in design throughput compared to existing state-of-the-art methods like RFDiffusion."

"OMTRA is a versatile multi-task generative model designed to unify various stages of structure-based drug design into a single computational framework. By utilizing multi-modal flow matching, the model can simultaneously handle different molecular data types—such as 3D atom positions and discrete chemical identities—to perform tasks like de novo ligand design, molecular docking, and conformer generation. A central innovation of this work is the support for pharmacophore conditioning, which allows researchers to guide the generation of new drugs using specific, human-interpretable interaction patterns. To power this system, the authors curated the Pharmit Dataset, a massive collection of 500 million 3D molecular conformers that significantly expands the chemical diversity available for machine learning."

"ALLSites is a unified deep learning framework designed to identify protein binding sites across the entire human proteome using only protein sequence data. By integrating the ESM-2 protein language model with a gated convolutional network and transformer architecture, the system captures both local contextual patterns and global residue interactions without needing expensive or rare high-resolution structural information. This structure-free approach allows the tool to predict binding locations for a comprehensive range of drug modalities, including small molecules, proteins, peptides, carbohydrates, and nucleic acids. Ultimately, ALLSites achieves state-of-the-art accuracy that rivals structure-based methods, providing a high-speed resource to expand the druggable proteome and accelerate drug discovery."

https://advanced.onlinelibrary.wiley....
https://doi.org/10.1093/nar/gkaf1226
https://arxiv.org/abs/2510.22304
https://arxiv.org/abs/2512.05080