Multi-modal large language model for drug development
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Molecular discovery has long been important to the advancement of drug development, where the goal is often to identify molecules that can act as ligands and bind effectively to a biological target such as a protein receptor or enzyme. Traditional discovery methods often depend on fixed templates or manually designed rules, making them limited in creativity, slow to scale, and costly to implement. Recent advancements in LLMs such as MolT5 (Edwards et al., 2022) offer promising alternatives by generating candidate molecules directly from natural language descriptions. However, most LLM-based approaches still behave like black boxes, producing output molecules without showing the intermediate reasoning steps behind their design. This lack of transparency and interpretability is a barrier in drug development, where understanding the design process is needed to justify decisions and ensure that generated molecules are chemically feasible. However, chemical feasibility alone is insufficient in target-directed drug discovery, since the usefulness of a generated molecule ultimately depends on whether it can achieve meaningful interaction with the target. Moreover, existing approaches often treat generation and evaluation as loosely coupled steps, providing limited controllability over target relevance. As a result, molecules generated by most LLM-based approaches may appear chemically plausible yet fail to exhibit useful ligand-receptor binding behaviour, leading to wasted effort in generation processes and downstream evaluation. Therefore, target-directed generation becomes crucial for prioritizing molecules that are both chemically feasible and more likely to bind to the intended receptor. Consequently, an effective LLM-based approach for target-directed molecule design system should (i) be interpretable, by reporting intermediate design traces and explicit constraints, and (ii) be target-aware, by incorporating target information and systematic evaluations that prioritise candidates with higher likelihood of binding to the intended receptor. To address the limitations of black-box LLM molecule generation and the lack of target-aware prioritisation, this project builds upon Intern-S1-mini (Bai et al., 2025), a recently developed LLM with strong general reasoning capabilities across scientific modalities, including molecule generation. The aim is to develop an interpretable end-to-end workflow for target-directed molecule design, where natural language inputs are translated into candidate SMILES molecules while preserving traceability of the generation process.
Fields of Study & Subjects
Computational Drug Discovery Methods
Multimodal Machine Learning Applications
Machine Learning in Materials Science
Publication details
- Journal / source
- DR-NTU (Nanyang Technological University)
- Publication date
- 1 January 2026
- Language
- EN