Goal-oriented de novo molecule design with an instruction-tuned LLM that honors property targets, substructure constraints, and numeric values.
No providers recorded yet. Browse all providers
Goal-oriented molecule design asks for compounds that satisfy explicit constraints: a target binding affinity, a drug-likeness score above some threshold, a scaffold that must be retained. The dominant approaches — Bayesian optimization over a learned latent space, or reinforcement learning against a reward model — each require training property predictors, and both handle substructure requirements awkwardly, since a latent vector has no natural slot for "keep this ring system."
ChatMol, from researchers at Soochow University, reframes the problem as instruction following. A pretrained large language model is fine-tuned to read a natural-language prompt describing the desired properties and substructures and to emit a molecule that satisfies them. Two additions make this work in practice. First, the authors design a molecule string format derived from SELFIES, stripping the square brackets and separating elements with spaces, which keeps the validity guarantees of SELFIES while producing shorter, more language-like sequences that fit an existing LLM vocabulary without extending it. Second, they address a well-known weakness of language models — treating numbers as arbitrary tokens — by adding a dedicated positional encoding for numeric values in the prompt, so that a requested property value is represented as a quantity rather than as digits.
Note that an unrelated project also called ChatMol exists: a PyMOL plugin that acts as a conversational assistant for molecular modeling. It shares only the name.
ChatMol fine-tunes Llama3 backbones ranging from 1B to 8B parameters using LoRA, training on molecules drawn from ZINC250k and MOSES with prompts constructed for each constraint setting. Across single-property, substructure-property, and multi-property benchmarks it outperforms variational autoencoder and reinforcement learning baselines. In a multi-objective binding-affinity task against the estrogen receptor ESR1, ChatMol reaches a dissociation constant of 0.25 while retaining the best overall performance, a 4.76% improvement over prior methods. The numerical encoding contributes a Pearson correlation gain of up to 0.49 between instructed property values and the properties of generated molecules, the paper's clearest evidence that the mechanism does what it claims. Performance improves with backbone size, consistent with LLM scaling behavior. The authors state that code, data, and model weights will accompany the final version; as of the preprint no checkpoints or repository have been released.
The natural fit is early-stage hit generation and lead optimization, where a medicinal chemist can state requirements in the terms they already use — retain this scaffold, push logP down, keep QED high — and receive candidate structures without configuring a separate optimizer for each objective. The substructure conditioning is what distinguishes it from latent-space samplers in practice: scaffold hopping and R-group enumeration around a fixed core map directly onto the prompt. The ESR1 case study shows the same interface driving a multi-objective affinity campaign against a named protein target.
ChatMol belongs to a broader shift in which general-purpose language models, adapted cheaply, replace bespoke generative architectures for chemistry. Its specific contribution is diagnostic as much as generative: by isolating numerical representation as the failure mode and fixing it with an encoding change, the work quantifies how much of an LLM's difficulty with quantitative design instructions is representational rather than a limit of scale. The practical caveat is availability — until the promised code and weights appear, the results cannot be reproduced or built on, and evaluation to date is entirely in silico against computed property scores rather than measured assays.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.