GeLLMO

GeLLMO — styled GeLLM3O in the paper — is a family of instruction-tuned large language models for molecule optimization, developed by the Ning Lab at Ohio State University. The task is lead optimization: given a hit molecule with some initial activity, produce a structurally similar molecule that is better on every property a medicinal chemist cares about.

Most computational methods for this handle one or two properties at a time. Real programs do not work that way. A CNS candidate has to cross the blood-brain barrier while staying drug-like, non-mutagenic, and orally absorbable, and those objectives conflict — improving lipophilicity to aid permeability tends to degrade solubility and drug-likeness. Methods built for single-property optimization scale badly to these combinations and generally have to be retrained whenever the property set changes, which is often, because priorities differ between programs.

The authors' bet is that instruction tuning gives LLMs the compositional handle this task needs. They assembled MuMOInstruct, an instruction-tuning dataset built specifically around tasks requiring three or more simultaneous property improvements, and fine-tuned open base models on it. The resulting checkpoints follow natural-language optimization instructions and, critically, extend to property combinations they were never trained on.

#Key Features

• Multi-property by design: Tasks require simultaneous improvement across three to six properties, in contrast to prior instruction-tuning corpora restricted to single- and double-property objectives.
• Generalists outperform specialists: Checkpoints trained across many property combinations beat their single-task counterparts on the harder tasks, because exposure to varied combinations teaches shared trade-offs that per-task data is too sparse to convey.
• Zero-shot transfer to unseen combinations: Training data for the broadest checkpoints excludes every property combination used in out-of-domain evaluation, so those results measure genuine transfer rather than memorization.
• Instruction-phrasing robustness: Each task appears under multiple phrasings during training, preventing overfitting to one wording and allowing chemists to describe objectives in their own terms.
• Similarity-constrained pairs: Every training pair satisfies a Tanimoto similarity threshold above 0.6, so the model learns to modify a scaffold rather than replace it.

#Technical Details

MuMOInstruct spans six properties — penalized logP, QED, blood-brain barrier permeability, mutagenicity, human intestinal absorption, and DRD2 binding — combined into 63 optimization tasks built from 255,174 molecule pairs. Forty-two tasks require at least three simultaneous improvements, and ten of those are partitioned into five in-domain and five out-of-domain tasks for evaluation. Checkpoints are produced by LoRA fine-tuning of Mistral-7B-Instruct-v0.3 and Llama-3.1-8B-Instruct, with adapters on all projection layers and the language modeling head; variants are trained on the power sets of three, four, and six properties, and generation uses beam search with twenty candidates per input. The metric is success rate, the fraction of inputs optimized with improvement in every desired property, evaluated zero-shot without in-context examples. On in-domain tasks the generalist Mistral checkpoints average 76.8% and 76.1% success, beating the strongest baseline by 113.2% and 108.8% relative, with gains up to 186.6% over the best baseline overall and up to 176% on the hardest four-property task. On out-of-domain tasks the six-property generalists reach 88.7% (Mistral) and 90.8% (Llama) average success rate, outperforming Claude 3.5 with five-shot prompting and the chemistry-tuned LlaSMol by as much as 159.9% on individual tasks.

#Applications

GeLLMO is aimed at computational lead optimization, where a program needs analogues of a hit that fix several liabilities at once. Stating objectives in natural language means the model can be driven from a target product profile and slotted into agentic discovery loops, and the zero-shot behavior on unseen combinations means it does not need refitting when a program's priorities shift. The dataset is a contribution in its own right, giving the field a benchmark for realistic multi-property optimization where previous instruction-tuning corpora tested only one or two objectives.

#Impact

GeLLMO established that instruction-tuned open LLMs can perform multi-property molecule optimization well enough to outperform much larger closed models on the task, and that training across many property combinations produces transferable understanding of chemical trade-offs rather than memorized task recipes. The Ning Lab's later GeLLMO-C builds directly on this line, adding property-specific control over which properties to improve and which to hold steady. The limitation to note is the evaluation basis: success rates are computed against computational property predictors, not measured assays, so they reflect the oracles used to score them. Code and data are released under MIT terms with the checkpoints published on HuggingFace.