Molecular optimization beyond experience summaries: the agent writes and revises causal hypotheses as it experiments, lifting mean AUC@10 across five molecular tasks from 0.250 to 0.279
Synopsis
Large language models can now construct and keep revising scientific principles during a discovery run, rather than leaving that translation to human experts. Doing it by hand needs domain expertise, time and effort, and bottlenecks discovery. Principia compares related inputs and their observed outcomes to write causal hypotheses, clusters similar inputs to induce recurring principles, then revises them with new observations. Across molecular optimization, protein engineering, gene perturbation and molecular property prediction, it beat baselines using internal knowledge, experience summaries or web retrieval; on five molecular optimization tasks it averaged AUC@10 0.279 versus 0.250 for the experience-summary baseline.
Interpretation
In multi-round candidate optimization, the guidance handed to the next round is now a set of rewritable causal explanations rather than a single running text summary of earlier results. Earlier guidance came from one of three places: candidates compressed into a rolling experience summary, general scientific text pulled from the web, or no guidance at all beyond population-style mutation and crossover. Across five molecular optimization tasks, mean AUC@10, which tracks how the ten best candidates improve within a limited evaluation budget, is 0.279 for Principia, 0.250 for the experience-summary baseline ExLLM, and 0.222 for the unguided MolLEO.
The explanations are written as comparisons: two similar inputs and their experimental outcomes are placed side by side, and the model states which shared or differing features account for outcomes that match or diverge. These texts, called comparative causal hypotheses, are grouped by similarity — molecular fingerprints, protein sequences or gene embeddings — and the model generalizes the relationships recurring within a group into a principle, then revises the hypotheses it used once new observations arrive. On eight TDC molecular property prediction tasks, with training and test molecules split by scaffold and test labels never entering knowledge construction, Principia averages 69.75 accuracy against 63.26 for the textual-summary baseline.
Perspective
All evaluations run against in-silico benchmark scoring functions: PMO scoring functions serve as the experimental oracle, which returns a molecule's score, under a budget of 1,000 distinct molecule evaluations; protein tasks use a Potts model, a pretrained predictor or a sequence lookup table for fitness; gene perturbation uses recorded CRISPR screen scores; property prediction uses eight TDC tasks split by scaffold, with up to 200 training and 50 test molecules per class. The main experiments use GPT-5-mini and gpt-5.4-mini as backbones. The authors frame test-time hypotheses-to-principle construction as a direction for advancing LLM-based scientific discovery.
Hypotheses formed from observations covering a small fraction of the input space can miss factors relevant to other inputs, and that limitation carries into the principles induced from them, which is why principles are rebuilt as new observations arrive. Scientific plausibility is shown qualitatively: an epoxide-bearing aryl ether motif is induced as a driver of mutagenicity, consistent with prior studies on such glycidyl compounds. The gene-perturbation hit ratio excludes the initial population and measures recall of benchmark hit genes.