MolSC: Leveraging Substituent Contributions to Enhance Fine-grained Molecular Understanding in LLMs

Hyuntae Park; Sooyeon Kim; Jiwon Park; SangKeun Lee

cs.LGSep 19, 2026

MolSC: Leveraging Substituent Contributions to Enhance Fine-grained Molecular Understanding in LLMs

Authors: Hyuntae Park, Sooyeon Kim, Jiwon Park, SangKeun Lee

Organizations: Department of Artificial Intelligence, Korea University, Seoul, Republic of Korea · Department of Computer Science and Engineering, Korea University, Seoul, Republic of Korea

Abstract

Recent advances in natural language processing have led to molecular Large Language Models (LLMs) with strong performance across diverse chemistry tasks. However, they still struggle to capture fine-grained structure-property relationships, particularly how small, localized modifications alter a molecule's behavior. To address this limitation, we introduce MolSC, a dataset of substituent contributions, defined as property changes induced by attaching specific substituents to molecular scaffolds. Curated from manually annotated bioactivity records, MolSC spans structural-alert liability, target-specific bioactivity, and physicochemical descriptors, and contains 181K substituent-level examples for training. We further propose MolSC-Bench, a held-out evaluation benchmark of 1,541 examples disjoint from MolSC at the scaffold, substituent, and molecule levels. Our experiments show that existing molecular LLMs and strong proprietary models such as GPT-5.2 and Gemini-3-Flash show limited reliability in substituent contribution prediction. In contrast, training on MolSC substantially improves this ability and achieves strong performance across diverse downstream molecular tasks. These results highlight substituent contribution learning as a key component of fine-grained molecular understanding.

Figures & tables

Figure 1: A MolSC-Bench example where our MolSC -trained model (Ours) correctly predicts property changes induced by substituent attachment, while GPT-5.2 predicts the opposite direction across all three property axes and misses the newly introduced BMS alert. Ground-truth (GT) changes are derived from ChEMBL.

Figure 2: Data generation process of MolSC . We divide each molecule into a scaffold and a substituent, and define the substituent contribution as the property difference between the molecule and the scaffold. We then build an instruction-tuning dataset with predefined templates. Templates and examples are provided in Appendix H .

	MolSC	MolSC-Bench
Contributions	181,098	1,541
Unique scaffolds	100,076	1,541
Unique substituents	20,541	1,541
Unique molecules	164,650	1,518
Instructions	149,232	1,541

Table 1: Statistics of MolSC and MolSC-Bench . Additional statistics are in Appendix B .

Method	Task 1: Property Profiling				Task 2: Substituent Contribution Prediction
Method	SR (%)	propMAE $\downarrow$	alertAcc $\uparrow$	bioMAE $\downarrow$	SR (%)	$\Delta$ propMAE $\downarrow$	$\Delta$ alertAcc $\uparrow$	$\Delta$ bioMAE $\downarrow$	dirAcc $\uparrow$
Mol-Instruction-8B ( 2023 )	0 40.6	24.019	0.667	–	0 65.1	9.263	0.075	0.735	0.454
ChemLLM-20B ( 2024 )	0 99.0	15.854	0.700	6.326	100.0	8.300	0.049	0.960	0.146
ChemDFM-v2-14B ( 2025 )	0 98.8	5.627	0.627	7.312	100.0	7.928	0.068	0.804	0.199
Gemini-3-Flash	100.0	9.013	0.667	1.480	100.0	2.535	0.100	0.799	0.722
GPT-5.2	100.0	10.531	0.517	0.657	100.0	3.519	0.219	2.471	0.706

Table 2: Results on MolSC-Bench . SR denotes Success Rate, the proportion of outputs that follow the required format. SC denotes Substituent Contribution; w/o SC refers to a model trained only on property profiling from MolSC . dual-view refers to our training strategy described in § 3.3 . Bold and underline indicate the best and second-best results, respectively. ‘–’ denotes no valid response, making the metric unmeasurable.

Method	# Params	Modality	Exact $\uparrow$	BLEU $\uparrow$	Lev $\downarrow$	RDK $\uparrow$	MAC $\uparrow$	Mor $\uparrow$	Validity $\uparrow$
Forward Reaction Prediction
Mol-Instructions ( Fang et al., 2023 )	8B	1D	0.503	0.883	13.410	0.756	0.863	0.708	1.000
InstructMol ( Cao et al., 2025 )	7B	1D + 2D	0.536	0.967	10.851	0.776	0.878	0.741	1.000
HIGHT ( Chen et al., 2025 )	7B	1D + 2D	0.037	0.869	23.759	0.590	0.394	0.340	0.993
UniMoT ( Guo et al., 2025 )	7B	1D + 2D	0.611	0.980	8.297	0.836	0.911	0.807	1.000
Omni-Mol ( Hu et al., 2025 )	2B	1D + 2D	0.733	0.980	5.550	0.895	0.947	0.870	1.000

Table 3: Results for reaction prediction tasks. Bold and underlined mark the best and second-best scores. 1D and 2D refer to 1D molecular representations and 2D molecular graphs.

Method	HOMO $\downarrow$	LUMO $\downarrow$	$\Delta\epsilon\downarrow$	Avg $\downarrow$
Mol-Instruction ( 2023 )	0.0210	0.0210	0.0203	0.0210
InstructMol ( 2025 )	0.0048	0.0050	0.0061	0.0050
HIGHT ( 2025 )	0.0056	0.0065	0.0077	0.0066
UniMoT ( 2025 )	0.0042	0.0047	0.0055	0.0049
Omni-Mol ( 2025 )	0.0038	0.0047	0.0049	0.0044
KnowMol ( 2025 )	0.0031	0.0037	0.0035	0.0034

Table 4: Results for molecular property regression tasks on the QM9 benchmark ( Fang et al., 2023 ) evaluated using Mean Absolute Error (MAE). Results for our variants are averaged over three random seeds. Bold and underline indicate the best and second-best results.

Method	BACE $\uparrow$	BBBP $\uparrow$	ACNet $\uparrow$	Avg $\uparrow$
Llama-3.2-1B-Instruct	83.50	71.00	74.12	76.21
Ours w/o SC, w/o dual-view	85.98	73.12	77.32	78.81
Ours w/o dual-view	88.59	70.71	77.93	79.08
Ours w/o SC	82.42	70.81	77.30	76.84
Ours-1B	86.58	72.01	79.89	79.49

Table 5: Results for drug-like molecular property tasks evaluated using ROC-AUC. BACE and BBBP are from MoleculeNet ( Wu et al., 2018 ) . Bold and underline indicate the best and second-best results.

Model	Single $\uparrow$	Interaction $\uparrow$	Comparison $\uparrow$
Mol-Instructions-8B ( 2023 )	0.107	0.059	0.469
ChemLLM-7B ( 2024 )	0.233	0.235	0.250
LlaSMol-Mistral-7B ( 2024a )	0.387	0.298	0.239
Qwen2.5-7B ( 2025 )	0.590	0.396	0.664
nach0-base ( 2024 )	0.606	0.543	0.041
Llama-3.2-1B-Instruct ( 2024 )	0.677	0.655	0.662

Table 6: Results for functional group-level property reasoning on FGBench ( Liu et al., 2025 ) , evaluated using boolean accuracy. Bold and underline indicate the best and second-best results.

Panel of Table 7 — Table 7: Property changes and model predictions after attaching morpholine to four different scaffolds. Ground-truth (GT) values are from ChEMBL. dir denotes the predicted direction of change; blue and red indicate correct and incorrect dir predictions. Details of the used molecules are provided in Table 24 .

Model	Direction acc. $\uparrow$	Conflict-pair acc. $\uparrow$
GPT-5.2	0.610	0.145
Ours-1B	0.804	0.582

Table 8: Results for context-dependent substituent effect prediction. Conflict-pairs denote (substituent, axis) pairs whose effect direction flips across the two scaffolds. Bold indicates the best results.

Appendix figures & tables20 assets

Supplementary material from the paper’s appendix.

Appendix

	mean	median	std	min	p95	max
Substituents per scaffold	1.810	1	1.391	1	5	10
Contributions per scaffold	9.611	6	8.255	0	27	111

Appendix

Table 9: Distribution of substituents and contributions per scaffold in MolSC .

Axis	Sub-axis	Mean $\Delta$	Std	% increase	% decrease
Physicochemical	PSA	22.98	22.91	83.2	16.8
	ALogP	0.87	1.03	78.2	21.8
	HBD	0.43	1.19	63.3	36.7
	QED	$-$ 0.15	0.14	9.1	90.9
	NP-likeness	$-$ 0.08	0.36	40.2	59.8
Bioactivity	IC50	0.07	0.85	53.7	46.3

Appendix

Table 10: Distribution of property changes in MolSC by axis. Structural alerts are binary and are therefore reported as the fraction of changes that introduce or remove an alert.

Axis	Sub-axis	$n$	Ratio (%)
Physicochemical		728,054	75.7
	ALogP	179,768	18.7
	QED	177,893	18.5
	NP-likeness	177,598	18.5
	PSA	131,952	13.7
	HBD	60,843	6.3

Appendix

Table 11: Number of axis-level annotations in MolSC by property axis.

Axis	Sub-axis	$\Delta_{P}$ std	Opp.-sign (%)	# eligible
Physicochemical	PSA	1.382	4.2	7,424
	ALogP	0.087	4.6	7,657
	HBD	0.025	0.0	4,474
	QED	0.081	25.8	7,642
	NP-likeness	0.210	46.6	7,617
Bioactivity	IC50	0.463	58.0	13,385

Appendix

Table 12: Variance of

\Delta_{P}

and frequency of opposite-sign effects for the same substituent across different scaffolds. Eligible substituents are those appearing on at least two scaffolds with a defined effect on the given axis.

Task	Model	Exact $\uparrow$	BLEU $\uparrow$	Lev $\downarrow$	RDK $\uparrow$	MAC $\uparrow$	Mor $\uparrow$	Validity $\uparrow$
Forward Reaction Prediction	Llama-3.2-1B-Instruct	0.778	0.986	4.993	0.906	0.950	0.889	1.000
Retrosynthesis	Llama-3.2-1B-Instruct	0.574	0.962	8.707	0.858	0.909	0.825	1.000
Reagent Prediction	Llama-3.2-1B-Instruct	0.206	0.685	16.196	0.480	0.587	0.466	1.000

Appendix

Table 13: Results for reaction prediction tasks of our backbone model.

Model	HOMO $\downarrow$	LUMO $\downarrow$	$\Delta\epsilon\downarrow$	Avg $\downarrow$
Llama-3.2-1B-Instruct	0.0036	0.0038	0.0045	0.0040

Appendix

Table 14: Results for property regression tasks of our backbone model.

Method	Forward		Retro
Method	Exact $\uparrow$	Lev $\downarrow$	Exact $\uparrow$	Lev $\downarrow$
Ours-1B w/o $\Delta_{P}$	0.917	1.333	0.658	6.192
Ours-1B	0.927	1.281	0.674	5.588

Appendix

Table 15: Effect of removing the property-change targets from the MolSC conversations. Bold indicates the best results.

Method	# Params	Modality	Exact $\uparrow$	BLEU $\uparrow$	Lev $\downarrow$	RDK $\uparrow$	MAC $\uparrow$	Mor $\uparrow$	Validity $\uparrow$
Forward Reaction Prediction
Text+Chem T5 ( Christofidellis et al., 2023 )	223M	1D	0.239	0.782	20.413	0.705	0.789	0.652	0.762
BioT5+ ( Pei et al., 2024 )	252M	1D	0.864	0.993	3.403	0.949	0.975	0.935	1.000
Ours-1B	1B	1D	0.927	0.990	1.281	0.977	0.986	0.969	1.000
Ours-3B	3B	1D	0.948	0.995	0.760	0.984	0.992	0.979	1.000
Retrosynthesis

Appendix

Table 16: Comparison with small language models pretrained on large-scale molecular data on the reaction prediction tasks. Bold and underlined mark the best and second-best scores within each task.

Method	# Params	Modality	HOMO $\downarrow$	LUMO $\downarrow$	$\Delta\epsilon\downarrow$	Avg $\downarrow$
MoLFormer-XL ( Ross et al., 2022 )	48M	1D	0.0061	0.0063	0.0094	0.0073
MolBridge ( Park et al., 2025 )	48M	1D	0.0064	0.0058	0.0087	0.0070
Ours-1B	1B	1D	0.0033	0.0030	0.0045	0.0034
Ours-3B	3B	1D	0.0029	0.0031	0.0039	0.0033

Appendix

Table 17: Comparison with pretrained molecular encoders on the QM9 property regression tasks, fine-tuned on the same training data as our experimental setting. Bold and underline indicate the best and second-best results.

Method	Exact $\uparrow$	BLEU $\uparrow$	Lev $\downarrow$	RDK $\uparrow$	MAC $\uparrow$	Mor $\uparrow$
Ours-1B w/o SC, w/o dual-view	0.201 $\pm$ 0.009	0.690 $\pm$ 0.018	16.36 $\pm$ 0.46	0.484 $\pm$ 0.008	0.584 $\pm$ 0.006	0.458 $\pm$ 0.005
Ours-1B w/o dual-view	0.206 $\pm$ 0.005	0.702 $\pm$ 0.006	15.75 $\pm$ 0.24	0.497 $\pm$ 0.006	0.592 $\pm$ 0.002	0.469 $\pm$ 0.004
Ours-1B w/o SC	0.230 $\pm$ 0.002	0.726 $\pm$ 0.008	14.61 $\pm$ 0.28	0.506 $\pm$ 0.011	0.605 $\pm$ 0.006	0.489 $\pm$ 0.007
Ours-1B	0.235 $\pm$ 0.007	0.732 $\pm$ 0.004	14.27 $\pm$ 0.12	0.516 $\pm$ 0.002	0.612 $\pm$ 0.001	0.495 $\pm$ 0.001

Appendix

Table 18: Results for reagent prediction across three random seeds, reported as mean

\pm

standard deviation. Bold indicates the best results.

Method	HOMO $\downarrow$	LUMO $\downarrow$	$\Delta\epsilon\downarrow$	Avg $\downarrow$
Ours w/o SC, w/o dual-view	0.00333 $\pm$ 0.00009	0.00380 $\pm$ 0.00033	0.00400 $\pm$ 0.00012	0.00370 $\pm$ 0.00015
Ours w/o dual-view	0.00340 $\pm$ 0.00002	0.00350 $\pm$ 0.00008	0.00402 $\pm$ 0.00030	0.00361 $\pm$ 0.00013
Ours w/o SC	0.00271 $\pm$ 0.00002	0.00273 $\pm$ 0.00003	0.00322 $\pm$ 0.00007	0.00287 $\pm$ 0.00002
Ours-1B	0.00276 $\pm$ 0.00003	0.00299 $\pm$ 0.00035	0.00335 $\pm$ 0.00016	0.00303 $\pm$ 0.00016

Appendix

Table 19: Results for QM9 property regression across three random seeds, reported as mean

\pm

standard deviation. Bold indicates the best results.

Model	# Params	Pre-downstream stage	Data scale
InstructMol ( Cao et al., 2025 )	7B	Pretraining	264K
HIGHT ( Chen et al., 2025 )	7B	Pretraining	295K
UniMoT ( Guo et al., 2025 )	7B	Pretraining	324K
Omni-Mol ( Hu et al., 2025 )	2B	Pretraining	310K
KnowMol ( Yang et al., 2025 )	7B	Instruction tuning	100K
Ours	1B	Instruction tuning	149K

Appendix

Table 20: Training stage and data scale used by each baseline prior to downstream fine-tuning.

Panel of Figure 3 — Figure 3: A SFT example for the forward direction: starting from a scaffold profile, the model is then asked to attach a substituent and report the resulting original (combined) molecule together with the induced property changes. stype is short for standard type.

Panel of Figure 4 — Figure 4: A SFT example for the backward direction: starting from a profile of the original molecule, the model is then asked to decompose it into a scaffold and a substituent and report the property changes attributable to the substituent.

Panel of Figure 5 — Figure 5: An example of a MolSC-Bench Task 1 prompt used for existing model inference, with the ground-truth profile shown below it.

Panel of Figure 6 — Figure 6: An example of a MolSC-Bench Task 2 prompt used for existing model inference (e.g., GPT-5.2, Gemini-3-Flash), with the ground-truth profile shown below it.

Turn	Paraphrasings
T0 (human) Profile request	(1) Predict the full molecular properties, structural alerts, and bioactivity of this molecule: {mol} . (2) What are the molecular properties, structural alerts, and target bioactivity of {mol} ? (3) For the molecule {mol} , give a complete profile of its physicochemical properties, alert hits, and bioactivity. (4) I have a molecule: {mol} . Provide its properties, alerts, and bioactivity profile. (5) Describe the molecular properties, structural alerts, and bioactivity of {mol} .
T1 (model) Profile response	(1) Properties: {props} . Structural alerts: {alerts} . Bioactivity: {bio} . (2) Property profile - {props} . Alert flags - {alerts} . Bioactivity - {bio} . (3) Physicochemical: {props} . Alerts: {alerts} . Activity: {bio} . (4) Here is the full profile. Properties: {props} . Alerts: {alerts} . Bioactivity: {bio} . (5) Computed properties: {props} . Structural alert hits: {alerts} . Measured bioactivity: {bio} .
T2 (human) Attachment request	(1) Apply substituent {sub} to the scaffold and tell me the resulting combined molecule. For the affected property axes ( {affected_props} ) and bioactivity stypes ( {affected_bio} ), report the signed change as combined $-$ scaffold. For every alert set (PAINS, Glaxo, BMS, Dundee, MLSMR), state whether it was introduced, removed, or preserved. (2) Attach fragment {sub} to the scaffold and give me the combined molecule. For each affected property axis ( {affected_props} ) and each affected bioactivity stype ( {affected_bio} ), list the signed change combined $-$ scaffold. Report the 3-way direction (introduced / removed / preserved) for all five alert sets. (3) Bond fragment {sub} onto the scaffold and write out the combined molecule. Give the signed change (combined $-$ scaffold) for: {affected_props} . Do the same for these bioactivity stypes: {affected_bio} . For each of the five alert sets, say introduced, removed, or preserved. (4) Combine the scaffold with fragment {sub} and output the resulting molecule. For each affected property axis ( {affected_props} ) and bioactivity stype ( {affected_bio} ), report combined $-$ scaffold. Finish with the 3-way direction for all five alert sets. (5) Take the scaffold, attach {sub} , and describe the resulting combined molecule. For the affected property axes ( {affected_props} ) and bioactivity stypes ( {affected_bio} ), give the signed difference combined $-$ scaffold. For every alert set, state whether the fragment introduced it, removed it, or preserved it.
T3 (model) Attachment response	(1) Combined molecule: {cm} . Effect: {chg} (2) After attachment, the molecule is {cm} . Effect: {chg} (3) The combined molecule is {cm} . Per-axis effect: {chg} (4) Result: {cm} . Per-axis effect: {chg} (5) Combined: {cm} . Effect: {chg}

Appendix

Table 21: Forward direction templates: each turn lists five paraphrasings sampled uniformly at conversation construction time. Placeholders: {mol} is the anchor molecule, {sub} the substituent, {cm} the combined molecule, {props}/{alerts}/{bio} the absolute profile fields, {affected_props}/{affected_bio} comma-separated axis names only, and {chg} the signed-delta list followed by the five alert labels.

Turn	Paraphrasings
T0 (human) Profile request	(1) Predict the full molecular properties, structural alerts, and bioactivity of this molecule: {mol} . (2) What are the molecular properties, structural alerts, and target bioactivity of {mol} ? (3) For the molecule {mol} , give a complete profile of its physicochemical properties, alert hits, and bioactivity. (4) I have a molecule: {mol} . Provide its properties, alerts, and bioactivity profile. (5) Describe the molecular properties, structural alerts, and bioactivity of {mol} .
T1 (model) Profile response	(1) Properties: {props} . Structural alerts: {alerts} . Bioactivity: {bio} . (2) Property profile — {props} . Alert flags — {alerts} . Bioactivity — {bio} . (3) Physicochemical: {props} . Alerts: {alerts} . Activity: {bio} . (4) Here is the full profile. Properties: {props} . Alerts: {alerts} . Bioactivity: {bio} . (5) Computed properties: {props} . Structural alert hits: {alerts} . Measured bioactivity: {bio} .
T2 (human) Decomposition request	(1) Find a scaffold + fragment decomposition for this molecule and tell me what each piece is. For the affected property axes ( {affected_props} ) and bioactivity stypes ( {affected_bio} ), report the signed change as combined $-$ scaffold. For every alert set (PAINS, Glaxo, BMS, Dundee, MLSMR), state whether it was introduced, removed, or preserved. (2) Decompose this molecule into scaffold and fragment. List the signed change (combined $-$ scaffold) for each affected property axis ( {affected_props} ) and each affected bioactivity stype ( {affected_bio} ). Report the 3-way direction for all five alert sets. (3) Break this molecule into a scaffold and a fragment. Give the signed change combined $-$ scaffold for: {affected_props} . Do the same for these bioactivity stypes: {affected_bio} . For each of the five alert sets, say introduced, removed, or preserved. (4) Identify a scaffold + fragment decomposition for this molecule and write both out. For each affected property axis ( {affected_props} ) and bioactivity stype ( {affected_bio} ), give combined $-$ scaffold. Finish by reporting all five alert sets as introduced, removed, or preserved. (5) Split this molecule into a scaffold and a fragment. Give the signed difference combined $-$ scaffold for the affected property axes ( {affected_props} ) and the affected bioactivity stypes ( {affected_bio} ). For every alert set, state whether the fragment introduced it, removed it, or preserved it.
T3 (model) Decomposition response	(1) Scaffold: {sc} . Fragment: {sub} . Effect: {chg} (2) The scaffold is {sc} and the fragment is {sub} . Effect: {chg} (3) After decomposition: scaffold {sc} , fragment {sub} . Per-axis effect: {chg} (4) Decomposition — scaffold: {sc} ; fragment: {sub} . Effect: {chg} (5) Decomposed: scaffold {sc} , fragment {sub} . Per-axis effect: {chg}

Appendix

Table 22: Backward direction templates: each turn lists five paraphrasings. Placeholders follow Table 21 , with {sc} denoting the scaffold (used in T3).

Axis	Property	Unit / Range	Description
Physicochemical descriptors	PSA	Å 2	Topological polar surface area.
	ALogP	log unit	Atom-based octanol-water partition coefficient.
	HBD	count	Hydrogen bond donor count.
	QED	[0, 1]	Quantitative estimation of drug-likeness.
	NP-likeness	$\sim$ [ $-5$ , $+5$ ]	Natural-product-likeness score.
	HBA	count	Hydrogen bond acceptor count.

Appendix

Table 23: All properties annotated in MolSC , organized along the three property axes. HBA, MW, Ro5 violations, aromatic rings, heavy atoms, rotatable bonds, and Ro3 pass are not considered when computing substituent contributions, since their values change monotonically under substituent attachment in our data. Bioactivity values are reported on the pchembl scale,

-\log_{10}

of the molar concentration.

Label	Structure	Textual Representation
Substituent		SMILES: C1COCCN1 SELFIES: [C][C][O][C][C][N][Ring1][=Branch1]
Scaffold A		SMILES: CN1Cc2ccccc2C(c2cc3ccccc3s2)C1 SELFIES: [C][N][C][C][=C][C][=C][C][=C][Ring1][=Branch1][C][Branch1][=C][C][=C][C][=C][C][=C][C][=C][Ring1][=Branch1][S][Ring1][=Branch2][C][Ring2][Ring1][Ring1]
Scaffold B		SMILES: O=C(Nc1ccccc1)c1cn2ccccc2n1 SELFIES: [O][=C][Branch1][#Branch2][N][C][=C][C][=C][C][=C][Ring1][=Branch1][C][=C][N][C][=C][C][=C][C][Ring1][=Branch1][=N][Ring1][=Branch2]
Scaffold C		SMILES: O=c1c(-c2cccs2)nc2cncnc2n1-c1ccccc1 SELFIES: [O][=C][C][Branch1][Branch2][C][=C][C][=C][S][Ring1][Branch1][=N][C][=C][N][=C][N][=C][Ring1][=Branch1][N][Ring1][#C][C][=C][C][=C][C][=C][Ring1][=Branch1]
Scaffold D		SMILES: COc1ccccc1Nc1nc(Nc2ccccc2S(=O)(=O)C(C)C)c2cc[nH]c2n1 SELFIES: [C][O][C][=C][C][=C][C][=C][Ring1][=Branch1][N][C][=N][C][Branch2][Ring1][=Branch1][N][C][=C][C][=C][C][=C][Ring1][=Branch1][S][=Branch1][C][=O][=Branch1][C][=O][C][Branch1][C][C][C][=C][C][=C][NH1][C][Ring1][Branch1][=N][Ring2][Ring1][=Branch1]

Appendix

Table 24: Molecules used in the context-dependent contribution analysis (Table 7 ).

Molecular Property Prediction

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

See all 202 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
SciCore-Mol: Augmenting Large Language Models with Pluggable Molecular Cognition Modules
May 21, 2026Yuxuan Chen, Changwei Lv, Yunduo Xiao +5AI-Assisted Scientific Research Chemical Reaction Prediction
School of Electronic and Computer Engineering, Peking University, Shenzhen, China · Tsinghua University, Beijing, China · School of Intelligence Science and Technology, Nanjing University, Suzhou, China
Probing Chemical Language Models: Effects of Pre-training and Fine-tuning
Jul 2, 2026Anna Karnysheva, Dietrich Klakow, Ji-Ung LeeLanguage Model Pretraining Chemical Language Models
1RTG Neuroexplicit Models · 2Spoken Language Systems · Saarland University +1

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

SciCore-Mol: Augmenting Large Language Models with Pluggable Molecular Cognition Modules

Probing Chemical Language Models: Effects of Pre-training and Fine-tuning