Polypharmacological target-aware molecule generation

Abstract

Abstract Motivation Designing compounds capable of modulating multiple target proteins simultaneously (i.e. polypharmacology) is critical for treating complex diseases such as cancer and neurodegenerative disorders. However, AI-driven multitarget drug design remains severely hindered by the scarcity of multitarget 3D structural data and the inefficiency of traditional “generate-then-filter” pipelines. There is a pressing need for scalable methods that can satisfy diverse binding constraints across multiple proteins without relying on scarce structural priors. Results To overcome these bottlenecks, we reformulate multitarget drug design as a sequence-based generation task and introduce PolyTamGen, an end-to-end autoregressive framework. We first curate a large-scale ligand-centric bioactivity dataset from ChEMBL and BindingDB, comprising 1.6 million unique ligands and 4.9 million bioactivity entries. PolyTamGen encodes target protein sets using ESM3, compresses these high-dimensional features into compact tokens via an attention module, and feeds them into a sequence-based compound generator to produce the final multitarget ligands. We further apply Direct Preference Optimization to improve docking-based affinity proxies and drug-likeness. Extensive experiments on dual- and triple-target benchmarks demonstrate that PolyTamGen improves joint success rates over representative baselines while maintaining high chemical validity and diversity. Availability and implementation The code is available at https://github.com/bjzgcai/PolyTamGen. The processed data, model checkpoints, and benchmark results are available at Zenodo: https://doi.org/10.5281/zenodo.20594128.