Instruction-tuned LLM series for multi-property molecule optimization that improves named drug properties while preserving those already in range.
No providers recorded yet. Browse all providers
GeLLMO-C is a series of instruction-tuned large language models for controllable molecule optimization, developed by the Ning Lab at Ohio State University. Given a starting molecule and a natural-language instruction, the model returns a modified molecule that improves the properties the instruction names — solubility, permeability, drug-likeness — without degrading the ones that already meet their thresholds.
That distinction is the point of the work. Real lead optimization is rarely "maximize everything": a chemist typically has a compound that is already acceptable on several axes and needs to fix one or two specific liabilities without breaking what works. Prior computational methods and instruction-tuned chemistry LLMs treat multi-property optimization as uniform improvement across all properties, which does not match how medicinal chemistry campaigns are actually run and produces molecules that trade away hard-won properties for gains elsewhere.
To train for the realistic version of the task, the authors built C-MuMOInstruct, an instruction-tuning dataset in which every example carries explicit, property-specific objectives — which properties to improve, and which to hold. The resulting GeLLMO-C checkpoints handle these targeted instructions and, notably, generalize zero-shot to property combinations and instruction phrasings never seen during training.
GeLLMO-C checkpoints are produced by LoRA instruction tuning of two open base models, Mistral-7B-Instruct-v0.3 and Llama-3.1-8B-Instruct, on C-MuMOInstruct. The 10 properties covered are penalized logP, QED, PAMPA permeability, blood-brain barrier penetration, human intestinal absorption, hERG inhibition, carcinogenicity, mutagenicity, drug-induced liver injury, and DRD2 inhibition. Evaluation spans 119 tasks — 51 in-distribution across five property combinations and 68 out-of-distribution across five novel ones. Against general-purpose and chemistry-specific LLM baselines, GeLLMO-C achieves up to 126% higher success rate on in-distribution tasks, with the largest margins on the difficult DRD2-focused combinations, and 35% higher success rate with 77% better relative improvement on out-of-distribution tasks. The released generalist checkpoints, GeLLMO-C-P10-Llama and GeLLMO-C-P10-Mistral, are published on HuggingFace with model cards; the codebase is MIT-licensed, but the weights themselves carry a CC BY-NC 4.0 license and are therefore not available for commercial use.
The natural setting is computational lead optimization, where a medicinal chemist has a hit with a known liability and needs analogues that fix it. Because the objectives are stated in natural language, the model can be driven directly by a project's target product profile and slotted into agentic drug discovery loops without retraining for each new property combination. The zero-shot generalization matters here: property priorities shift between programs, and a model that must be refit for every new combination is impractical.
GeLLMO-C reframes molecule optimization around the constraint that makes it hard in practice — preserving what already works — and shows that instruction tuning with explicit per-property objectives is enough to teach that behavior, including for combinations never seen in training. The release of both the dataset and the checkpoints makes the line directly extensible. The evaluation, however, rests entirely on computational property predictors rather than measured assays, so reported success rates reflect the oracles used to score them; the non-commercial weight license also limits industrial adoption, even though C-MuMOInstruct itself is released under the more permissive CC BY 4.0 with a dataset card documenting its composition and evaluation splits.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.