Loading…
Multi-modal LLM answering free-form questions about a compound's indications, pharmacodynamics and mechanism of action from its SMILES string.
No providers recorded yet. Browse all providers
Ask a pharmacologist three routine questions about a candidate compound — what does it treat, what does it do to the body, and by what mechanism — and the honest answers are paragraphs, not labels. Computational drug-property prediction has answered a different question instead: it assigns a compound to one of a few discrete classes, one model per endpoint. That framing captures binary bioactivity and toxicity reasonably well, and has nothing to say about an indication or a mechanism of action, which exist in the literature only as prose.
DrugChat reframes those predictions as text generation. A user supplies a SMILES string and a question in plain English, and the model writes a free-form answer; because the answer is generated rather than selected, one fixed checkpoint handles indication, pharmacodynamics, mechanism of action, cytotoxicity, administration route and prodrug status by prompt alone. It comes from Pengtao Xie's group at UC San Diego, with collaborators in UCSD's bioengineering, pharmacy and medicine schools and at Scripps Research. A prototype appeared in 2023; the full system, adding a visual encoder, human-expert evaluation and GPT-4 comparisons, followed as a bioRxiv preprint in October 2024.
The design is the encoder–adaptor–LLM recipe standard in multimodal language models, with molecule-aware encoders in place of the usual vision tower. The same group later applied the pattern in GeneChat for nucleotide sequences and MetaboliteChat for metabolites; DrugChat is the earliest of the three.
The molecular graph is encoded by a five-layer graph isomorphism network (roughly 0.5M parameters, 300-dimensional output) pretrained by context prediction on 2 million unlabeled ZINC15 molecules; the rendered image goes through ImageMol, an 18-layer ResNet (11M parameters, 512-dimensional output) pretrained on images of 10 million PubChem compounds. Two MLP adapters — 28M and 29M parameters, both projecting to 5120 dimensions — turn those vectors into a "molecule token" appended to the tokenized prompt and read by Vicuna-13B. The released training configuration freezes both molecular encoders, so training fits the adapters bridging them to the language model. Training data was curated from ChEMBL, PubChem and DrugBank into 91,365 (molecule, prompt, answer) triplets over roughly 14,000 unique compounds, and ran for 10 epochs on one 80GB A100.
On 572 held-out DrugBank molecules, expert graders scored DrugChat 1.05, 0.94, 0.80 and 0.92 out of 2 on indication, pharmacodynamics, mechanism of action and overview, against GPT-4's 0.38, 0.82, 0.45 and 0.46; automated metrics agree in direction (semantic similarity 0.47 versus 0.32). On discrete endpoints it reaches F1 of 0.57 to 0.71 for cytotoxicity across three human cell lines and 0.59 to 0.67 for administration routes, ahead of LLM baselines and of ImageMol as a task-specific classifier.
The intended user is a medicinal chemist or pharmacology researcher triaging compounds whose annotations are thin — asking, of a structure with no DrugBank entry, what activity to expect and what to screen for next. Answering in prose and accepting follow-ups makes it an exploratory interface rather than a scoring function, and its discrete-endpoint performance also makes it a single stand-in for several routine filters — cytotoxicity, route, prodrug status — early in a screening cascade.
DrugChat is a preprint, not peer reviewed, and the authors describe it as a prototype not yet validated by pharmaceutical experts. A candid reading of its own numbers matters: an expert score near 1.0 on a 0–2 scale means the typical answer is partially correct, so predictions are leads to check rather than conclusions, and the model gives no signal about which statements to distrust. Training and inference code are released under BSD-3-Clause and the instruction-tuning corpus is on HuggingFace under an MIT license, but no trained checkpoint is distributed — the repository instructs users to train their own — raising the cost of reproducing the results. What the work establishes is that grounding a chat model in molecule-specific encoders beats prompting it with raw SMILES by a wide margin, and that free-form generation is a workable output format for attributes categorical predictors cannot address at all.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.