CoMPASS: Collaborative Molecular Property Prediction via Adaptive Small-Large Model Synergy
2026-08-31 • Machine Learning
Machine LearningArtificial Intelligence
AI summaryⓘ
The authors developed CoMPASS, a method that combines two types of AI models to better predict molecular properties. They keep a graph neural network (GAT) as the main predictor but let a large language model (LLM) help only when the GAT is unsure. CoMPASS uses examples from the training data to guide the LLM, then carefully adjusts the GAT's output based on how much the two models agree. Their tests show this approach improves predictions by adding helpful LLM insights without letting it override confident GAT decisions. The authors emphasize that controlled collaboration, rather than replacing predictions, leads to better results.
Molecular property predictionGraph neural networks (GNN)Graph attention network (GAT)Large language models (LLM)Model calibrationRetrieval-based learningModel fusionChemical reasoningClassification and regression benchmarks
Authors
Wentao Li, Jiangjie Qiu, Yijun Li, Leyi Zhao, Xiaonan Wang
Abstract
Accurate molecular property prediction requires both statistical reliability and chemical reasoning. Graph neural networks can be calibrated directly on labeled assays but remain limited by the coverage of their training data. Large language models (LLMs) can compare molecular evidence and articulate chemical rationales, yet are unreliable as standalone quantitative predictors. The central challenge is therefore to determine when an LLM should influence a calibrated model and by how much. Here we present CoMPASS, a retrieval-calibrated framework for small-large model collaboration. CoMPASS retains a graph attention network (GAT) as the predictive anchor, retrieves locally relevant training molecules, provides attention-grounded evidence to an LLM, and converts its proposal into a bounded correction through an agreement-aware gate. Across six classification and two regression benchmarks, CoMPASS improves the GAT anchor in regions of correctable uncertainty while limiting LLM intervention in high-confidence regimes. Ablations show that the gains arise from validation-calibrated retrieval and bounded fusion rather than prompting alone. These results suggest that generative reasoning should augment calibrated prediction through evidence-grounded, controlled corrections rather than direct output replacement. Code is available at https://github.com/littlepeachs/CoMPASS.