Can Decision Trees Teach Large Language Models? Distilling Verbalized Knowledge for Molecular Property Prediction

Khiem Le; Sreejata Dey; Marcos Martínez Galindo; Vanessa Lopez; Ting Hua; Nitesh V. Chawla; Hoang Thanh Lam

cs.LGMar 12, 2026

Can Decision Trees Teach Large Language Models? Distilling Verbalized Knowledge for Molecular Property Prediction

Authors: Khiem Le, Sreejata Dey, Marcos Martínez Galindo, Vanessa Lopez, Ting Hua, Nitesh V. Chawla, Hoang Thanh Lam

Organizations: University of Notre Dame, IN, USA · IBM Research

Abstract

Molecular Property Prediction (MPP) is a fundamental problem in drug discovery that has recently attracted growing attention. Large Language Models (LLMs), known for their impressive proficiency across domains, show promise as generalist models for MPP. However, their current performance remains below the threshold needed for practical adoption. To bridge this gap, we propose TreeKD for distilling the knowledge of tree-based specialist models into LLMs to complement the internal knowledge of LLMs and improve their predictive accuracy. For each property, we train a specialist decision tree using features derived from 40K functional groups in the input molecules. Then, the predictive rule learned by the decision tree, which encodes its knowledge, is verbalized and incorporated into the prompts for training LLMs. In addition, by replacing a single decision tree with a Random Forest, we introduce a test-time scaling technique called rule-consistency, which aggregates predictions generated from different prompts constructed with different rules. An extensive evaluation with two LLMs, Gemma-2-2B and Granite-3.3-2B, on the TDC benchmark with 22 prediction tasks shows that our method substantially enhances the performance of LLMs, advancing the development of generalist models for MPP.

Figures & tables

Figure 1: An example illustrates that the LLM sees a molecule as a whole, while a decision tree operates on each dimension of functional groups.

Figure 2: An overview of our proposed method, consisting of 4 stages: ① Extracting FGs, ② Building Specialist Models, ③ Distilling Knowledge to LLMs, and ④ Test-Time Scaling.

Figure 3: An example of rule verbalization. The decision tree is turned into a hierarchy of if-then conditions, where indentation indicates the depth of nodes.

Type	Property	Metric	Specialist Models		TxGemma-2B	Gemma-2-2B
Type	Property	Metric	MapLight	Decision Tree	TxGemma-2B	Naive	TreeKD	w/ self-consistency	w/ rule-consistency
Reg	Caco-2 Permeability	MAE (↓)	0.269	0.413	0.629	0.547	0.476 ( + 0.071 )	0.551 0.008 ( - 0.075 )	0.528 0.005 ( - 0.052 )
	Lipophilicity	MAE (↓)	0.547	0.848	0.598	0.672	0.579 ( + 0.093 )	0.634 0.009 ( - 0.055 )	0.600 0.006 ( - 0.021 )
	Solubility	MAE (↓)	0.793	1.771	0.950	2.921	0.921 ( + 2.000 )	0.982 0.019 ( - 0.061 )	0.868 0.025 ( + 0.053 )
	PPBR	MAE (↓)	7.654	11.492	9.986	7.556	7.503 ( + 0.053 )	8.380 0.168 ( - 0.877 )	7.459 0.106 ( + 0.044 )
	VDss	Spearman (↑)	0.722	0.414	0.387	0.522	0.647 ( + 0.125 )	0.602 0.013 ( - 0.045 )	0.675 0.007 ( + 0.028 )

Table 1: The results on the TDC benchmark. ( green ) and ( red ) values for TreeKD indicate performance gaps relative to naive training; boldface denotes better results in a direct comparison with TxGemma2-2B. In the 2 rightmost columns, ( green ) and ( red ) values indicate performance gaps relative to TreeKD for test-time scaling techniques; underlining denotes better results between them. The subscripts report standard deviations across 3 runs.

Figure 4: An example of a complete prompt on Caco-2 Permeability. All numerical values are rounded to 6 decimal places.

Type	Property	Gemma-2-2B
		Naive	Naive	Naive
			w/ FGs	w/ FGs
				w/ a predictive rule
Reg	Caco-2 Permeability	0.547	+ 0.055	+ 0.016
	Lipophilicity	0.672	- 0.005	+ 0.098
	Solubility	2.921	+ 2.012	- 0.012

Table 2: The performance gains on the TDC benchmark with evolved prompt variants. ( green ) and ( red ) values indicate performance gaps relative to the preceding one.

Type	Property	Specialist Models		Gemma-2-2B
		Decision Tree		TreeKD
		$d$ = 3	$d$ = 6	$d$ = 3	$d$ = 6
Reg	Caco-2 Permeability	0.566	0 0.413 ( + 0.153 )	0.442	0.476 ( - 0.034 )
	Lipophilicity	0.913	0 0.848 ( + 0.065 )	0.636	0.579 ( + 0.057 )
	Solubility	1.863	0 1.771 ( + 0.092 )	0.949	0.921 ( + 0.028 )
	PPBR	11.734	11.492 ( + 0.242 )	7.827	7.503 ( + 0.324 )

Table 3: The results of TreeKD on the TDC benchmark using maximum depths (

d

) of 3 and 6. ( green ) and ( red ) values indicate performance gaps when using

d=3

relative to using

d=6

Figure 5: The distribution of 10498 molecules in the test set of 13 classification tasks from the TDC benchmark, according to whether they are correctly predicted by the decision tree and the LLMs.

Figure 6: The distribution of 55 pairs of molecules exhibiting activity cliffs that the decision tree fails to distinguish, according to whether both molecules in each pair are correctly predicted by the LLMs.

Figure 7: A few examples of pairs of molecules exhibiting activity cliffs on Ames Mutagenicity (differences between them are highlighted in red ), showing whether both molecules in each pair are correctly predicted.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Type	Property	Dataset Size	Dataset Split
Reg	Caco-2 Permeability	906	Scaffold
	Lipophilicity	4200	Scaffold
	Solubility	9982	Scaffold
	PPBR	1797	Scaffold
	VDss	1130	Scaffold
	Half Life	667	Scaffold

Appendix

Table 4: Detailed information and statistics of datasets.

Type	Property	Gemma-2-2B
		$N$ = 50		$N$ = 100
		self-*	rule-*	self-*	rule-*
Reg	Caco-2 Permeability	0.551	0.528	0.537 ( + 0.014 )	0.520 ( + 0.008 )
	Lipophilicity	0.634	0.600	0.627 ( + 0.007 )	0.595 ( + 0.005 )
	Solubility	0.982	0.868	0.978 ( + 0.004 )	0.869 ( - 0.001 )
	PPBR	8.380	7.459	7.913 ( + 0.467 )	7.329 ( + 0.130 )

Appendix

Table 5: The results of TreeKD on the TDC benchmark with test-time scaling techniques using

N=50

and

N=100

. Note that * stands for consistency. underlining denotes better results between techniques when using the same

N

. ( green ) and ( red ) values indicate performance gaps when using

N=100

relative to using

N=50

Type	Property	Random Forest
Type	Property	$N$ = 50	$N$ = 100
Reg	Caco-2 Permeability	0.355	0 0.354 ( + 0.001 )
	Lipophilicity	0.842	0 0.841 ( + 0.001 )
	Solubility	1.702	0 1.703 ( - 0.001 )
	PPBR	11.043	11.123 ( - 0.080 )
	VDss	0.489	0 0.491 ( + 0.002 )

Appendix

Table 6: The results of Random Forest on the TDC benchmark using

N=50

and

N=100

. ( green ) and ( red ) values indicate performance gaps when using

N=100

relative to using

N=50

Type	Property	Metric	Gemma-2-2B
Type	Property	Metric	Naive	TreeKD
Reg	Caco-2 Permeability	MAE (↓)	000 5.847	000 0.607 ( 000 + 5.240 )
	Lipophilicity	MAE (↓)	000 1.687	000 0.981 ( 000 + 0.706 )
	Solubility	MAE (↓)	00 42.173	000 1.937 ( 00 + 40.236 )
	PPBR	MAE (↓)	00 58.217	00 10.364 ( 00 + 47.853 )
	VDss	MAE (↓)	0 796.001	00 11.765 ( 0 + 784.236 )

Appendix

Table 7: The results on the TDC benchmark (on regression tasks) in the zero-shot setting. ( green ) and ( red ) values for TreeKD indicate performance gaps relative to naive inference.

Type	Property	Metric	Gemma-2-2B
Type	Property	Metric	Naive	TreeKD
Reg	Caco-2 Permeability	MAE (↓)	000 2.803	000 0.593 ( 000 + 2.210 )
	Lipophilicity	MAE (↓)	000 1.214	000 0.957 ( 000 + 0.257 )
	Solubility	MAE (↓)	000 2.567	000 1.930 ( 000 + 0.637 )
	PPBR	MAE (↓)	00 13.330	00 10.031 ( 000 + 3.299 )
	VDss	MAE (↓)	000 4.653	000 2.278 ( 000 + 2.375 )

Appendix

Table 8: The results on the TDC benchmark (on regression tasks) in the 3-shot setting. ( green ) and ( red ) values for TreeKD indicate performance gaps relative to naive inference.

Type	Property	Gemma-2-2B
		Naive	Naive	Naive
			w/ a predictive rule	w/ FGs
				w/ a predictive rule
Reg	Caco-2 Permeability	0.547	+ 0.014	+ 0.057
	Lipophilicity	0.672	- 0.026	+ 0.119
	Solubility	2.921	- 0.084	+ 2.084

Appendix

Table 9: The performance gains on the TDC benchmark with evolved prompt variants. ( green ) and ( red ) values indicate performance gaps relative to the preceding one.

Panel of Figure 8 — Figure 8: The inference times (in seconds) across different batch sizes.

Molecular Property Prediction

Molecular property prediction estimates chemical, physical, and biological properties from molecular structure and context.

See all 197 papers

Multi-Granular Rationale-Guided Molecular LLM for Property Prediction
Aug 11, 2026Junwoo Park, Minyoung Shin, Cheol Soon Lee +1Molecular Property Prediction Molecular Representation Learning
Sungkyunkwan University · Korea University
MolE-RAG: Molecular Structure-Enhanced Retrieval-Augmented Generation for Chemistry
Jun 4, 2026Joey Chan, Wonbin Kweon, Ashley Shin +4Retrieval-Augmented Generation LLM Inference
University of Illinois Urbana-Champaign · University of California, San Diego
Molecular Déjà Vu: Digit-Level Retrieval of Molecular Properties in Frontier Language Models
Sep 4, 2026Matthias Busch, Marius Tacke, Sviatlana V. Lamaka +4LLM Evaluation LLM Reliability
Institute for Artificial Intelligence and Simulation in Mechanics, Hamburg University of Technology, Eißendorfer Straße, 21073 Hamburg, Germany · Institute of Material Systems Modeling, Helmholtz-Zentrum Hereon, Max-Planck-Straße, 21502 Geesthacht, Germany · Institute of Surface Science, Helmholtz-Zentrum Hereon, Max-Planck-Straße, 21502 Geesthacht, Germany +3

Aug 11, 2026J/K move · Enter open · S save

Multi-Granular Rationale-Guided Molecular LLM for Property Prediction

Junwoo Park, Minyoung Shin, Cheol Soon Lee, Sujee Lee

Sungkyunkwan University · Korea University

Molecular Property Prediction Molecular Representation Learning LLMs for Graph Learning

Open paper AskPDF arXiv

Large language models (LLMs) are widely applied across chemical tasks, such as molecular property prediction, which underpins drug discovery. Molecular LLMs represent a molecule through several modalities, notably a 1D SMILES sequence or a 2D molecular graph. Both encode molecular information implicitly, so the contribution of individual substructures remains opaque. Retrieval and augmentation methods add context, but from external sources. However, the cues chemists reason over are the internal substructures that drive a property up or down. We propose MR-MoL, a multi-granular rationale-guided molecular LLM that supplies this evidence directly. A fine-tuned GNN scores each substructure through masking, and the most influential ones are serialized as a ranked, direction-tagged rationale that the LLM reads alongside the SMILES sequence and molecular graph. The rationale spans three levels of granularity: Murcko scaffolds with their side chains, BRICS fragments, and functional groups. This is, to our knowledge, the first method to expose GNN-derived attributions to an LLM as evidence for property prediction. On eight MoleculeNet tasks, MR-MoL achieves the best overall results among generalist models and narrows the gap to specialist models tuned for each task. Five diagnostics further confirm that the model reads the rationale rather than merely benefiting from its presence. Its direction, rank, and substructure each shape the prediction, and its attributions reproduce known structure-property relationships.

Abstract

Molecular Property Prediction

Explore similar work

Multi-Granular Rationale-Guided Molecular LLM for Property Prediction

MolE-RAG: Molecular Structure-Enhanced Retrieval-Augmented Generation for Chemistry

Molecular Déjà Vu: Digit-Level Retrieval of Molecular Properties in Frontier Language Models