Deciphering Scientific Reasoning Steps from Outcome Data for Molecule Optimization
TLDR
DESRO infers scientific reasoning steps from outcome data using LLMs, achieving high success in molecule optimization across 18 tasks.
Reasoning
The paper presents a novel framework (DESRO) that addresses the supervision gap in training reasoning models by recovering intermediate reasoning from grouped outcome data, with strong empirical results on molecule optimization. Its strengths include large-scale evaluation and generalization to out-of-distribution scenarios, but it is limited to a specific domain (molecule optimization) and does not cover broader scientific discovery tasks.
Read-first score
Read-first score 44.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 40.
Field roles
Rank sensitivity
Stability: volatile; rank range: 39.
Keyword Scores
Deep Analysis
Innovations
- Framework DESRO that deciphers scientific reasoning steps from outcome data by grouping data and using LLMs to recover underlying logic.
- Instantiation in molecule optimization: inferring optimization rationales from 2.3M molecular property records by grouping molecules with shared fragments and analyzing structural-property correlations with an LLM.
- Training a model that performs interpretable reasoning for molecule optimization, achieving state-of-the-art success rates and robust out-of-distribution generalization.
- Demonstration of framework generality by extending to reaction ligand selection.
Methodology
DESRO groups molecules with shared fragments from 2.3 million property records, then uses a large language model to analyze how structural variations correlate with property differences, inferring optimization rationales. From the derived reasoning data, a model is trained to conduct molecule optimization through an interpretable reasoning process, evaluated on 18 single- and multi-property optimization tasks.
Key Results
DESRO achieves the highest success rates on 15 out of 18 tasks, robustly generalizes to out-of-distribution scenarios (novel property combinations, unseen targets, natural language-defined properties), and autonomously reconstructs expert-level lead optimization trajectories under strict temporal splits.