TRIM: Extracting Reasoning Capabilities from Interpretable Models to Empower AI Teaching Systems for Molecular Classification

TRIM (Teaching Reasoning from Interpretable Models) is a framework that combines Explainable Boosting Machines (EBM) with large language models, aiming to resolve the conflict between AI black boxes and interpretability. It generates high-quality reasoning data through global single-molecule analysis and local neighbor comparison, which is used to train AI agents with chemical reasoning capabilities, supporting interpretability research in scientific fields such as drug discovery.

In the AI field, powerful models are often difficult to interpret, while interpretable models lack sufficient performance. The decision-making process of deep learning models is a "black box". In scientific fields like drug discovery, researchers not only need to know the results of molecular properties but also understand the reasons behind them. The TRIM project was born to address this: it combines interpretable machine learning with large language models to build a framework that extracts reasoning knowledge and trains the next generation of AI systems.

TRIM adopts a three-tier architecture:

1. Global Single-Molecule Analysis: Use EBM (Explainable Boosting Machine) to analyze individual molecules, integrating RDKit descriptors, pKa features, and functional group features (compressed from 95 to 36), and providing feature contribution scores.
2. Local Neighbor Comparison: Retrieve the 6 most similar known molecules to the target molecule (based on Morgan fingerprints and feature similarity), construct pairwise comparison features, and use EBM training to output similarity reasoning predictions.
3. Fused Reasoning: Integrate global and local results to make intelligent decisions using complementarity. Experiments show that the fused mode achieves an average macro F1 of 0.7019 on the validation set, and the local mode achieves the best result of 0.6917 on the test set.

TRIM converts EBM reasoning into teaching data:

• Reasoning Evidence Extraction: Global (feature contribution direction, structured analysis), Local (neighbor similarity, pairwise comparison, prediction confidence).
• Reasoning Rewriting: Use large language models to convert structured evidence into natural language: Global Rewriting (feature contribution description), Local Rewriting (neighbor analogy reasoning), Fused Rewriting (complete decision chain). Rewriting follows quality control: select at least one correctly predicted sample, explicitly reference neighbors, baseline awareness, and no meta-discourse.

TRIM provides a toolchain to train AI agents:

• Tool Definitions:
  • get_mol_properties_and_fg(SMILES): Returns molecular descriptors and functional group information.
  • compare_similar_mols(SMILES): Returns the 6 most similar neighbors and comparison analysis.
• Task List: Defines task names, label semantics, neighbor retrieval configurations, and dense feature lists, supporting the expansion of new tasks.

Innovations of TRIM:

1. Balancing Interpretability and Performance: EBM's accuracy in molecular classification tasks is comparable to that of black-box deep learning models, and its decisions are transparent.
2. From Explanation to Teaching: Converting model explanations into teaching materials for training other AI systems is a new paradigm of knowledge distillation.
3. Formalization of Scientific Reasoning: Simulates chemists' thinking: global feature analysis (physical and chemical judgment), neighbor comparison (analogical reasoning), and fusion layer (comprehensive decision-making).
4. Complete Engineering Pipeline: Provides a complete pipeline for data preparation, model training, evaluation, visualization, and reasoning rewriting (e.g., scripts like train_global_ebm.py).

Application Scenarios: Drug discovery (accelerating lead compound optimization), toxicity prediction (meeting regulatory transparency), AI chemical assistant (intelligent consultation), scientific education (helping understand molecular structure and properties). Future Directions: Expand to more molecular property predictions, introduce 3D conformation information, develop interactive visualization tools, and build larger reasoning datasets to train stronger models.