ProLigGPT: a language model for protein-targeted ligand design free from structural constraints

The deep-learning language model offers promising approaches to molecular design. However, existing models typically exhibit one or more of the following limitations: structural constraints, confinement to a single task, and lack of cross-domain capability (e.g., protein-to-ligand generation). Inspired by GPT's success in language modeling, we introduce ProLigGPT, a seq2seq autoregressive model based on GPT-2 and trained on ~1.9 million protein-ligand pairs without 3D structures. ProLigGPT supports tasks including de novo ligand generation and scaffold-based molecular design for target proteins. Benchmarking against three state-of-the-art structure-based models across eight representative target proteins, ProLigGPT demonstrated superior performance on comprehensive metrics. Through systematic evaluations of ProLigGPT's interpretability, we find that ProLigGPT uses the full protein sequence to generate target-aware ligands, unlike pocket-based strategies that depend only on residues near the binding site. Applying ProLigGPT to Pl3Kalpha; inhibitor design successfully yielded inhibitors with a novel scaffold and inhibitors derived from a known scaffold, demonstrating that ProLigGPT can reach beyond the boundaries of known chemical space. Finally, we validated the generalization ability of ProLigGPT on Target T1, which lacks a crystal structure and known inhibitors, confirming that ProLigGPT can design inhibitors solely from sequence information. These findings highlight the promise of language models in designing ligands for target proteins free from structural constraints, offering a flexible and scalable framework for exploring vast chemical spaces in drug discovery.