Explainable Molecular Property Prediction: Aligning Chemical Concepts With Predictions via Language Models

Contribution

Aligns molecular predictions with chemical concepts for more faithful explanations.

Abstract

We study explainable molecular property prediction with a language-model-based predictor over Group SELFIES representations. By combining attention and gradient signals with a margin-aware objective, the method produces explanations that better align with chemically meaningful concepts while preserving predictive performance.

Why This Paper Matters

It gives a concrete route to explainable molecular property prediction without separating prediction quality from concept-level alignment, which is useful for trustworthy molecular AI.

BibTeX

@article{wang2026tpami,
  title={Explainable Molecular Property Prediction: Aligning Chemical Concepts With Predictions via Language Models},
  author={Wang, Zhenzhong and Lin, Zehui and Lin, Wanyu and Yang, Ming and Zeng, Minggang and Tan, Kay Chen},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year={2026},
  doi={10.1109/tpami.2026.3664098},
}