Scaffold-Conditioned Preference Triplets for Controllable Molecular Optimization with Large Language Models

Yi Xiong; Liang Xiong; Xiaohong Ji; Sen Yang; Zhifeng Gao; Huaimin Wang; Kele Xu

cs.LGApr 14, 2026

Scaffold-Conditioned Preference Triplets for Controllable Molecular Optimization with Large Language Models

Authors: Yi Xiong, Liang Xiong, Xiaohong Ji, Sen Yang, Zhifeng Gao, Huaimin Wang, Kele Xu

Organizations: National Key Laboratory of Parallel and Distributed Processing College of Computer Science and Technology National University of DefenseTechnology · DP Technology

Abstract

Molecular property optimization is central to drug discovery, yet many deep learning methods rely on black-box scoring and offer limited control over scaffold preservation, often producing unstable or biologically implausible edits. While large language models (LLMs) are promising molecular generators, optimization remains constrained by the lack of chemistry-grounded preference supervision and principled data curation. We introduce Scaffold-Conditioned Preference Triplets (SCPT), a pipeline that constructs similarity-constrained triplets $\langle\text{scaffold}, \text{better}, \text{worse}\rangle$ via scaffold alignment and chemistry-driven filters for validity, synthesizability, and meaningful property gains. Using these preferences, we align a pretrained molecular LLM as a conditional editor, enabling property-improving edits that retain the scaffold. Across single- and multi-objective benchmarks, SCPT improves optimization success and property gains while maintaining higher scaffold similarity than competitive baselines. Compared with representative non-LLM molecular optimization methods, SCPT-trained LLMs are better suited to scaffold-constrained and multi-objective optimization. In addition, models trained on single-property and two-property supervision generalize effectively to three-property tasks, indicating promising extrapolative generalization under limited higher-order supervision. SCPT also provides controllable data-construction knobs that yield a predictable similarity-gain frontier, enabling systematic adaptation to diverse optimization regimes.

Figures & tables

Figure 1: Optimization failure cases. (a) High similarity, negligible gain. (b) High gain, low similarity.

Figure 2: The overall framework of our method. It consists of three main modules: (a) the SCPT pipeline, where training data are prepared through molecular pair construction, pair filtering, and scaffold extraction; (b) the alignment framework, which fine-tunes the LLM based on the constructed data formats.

Panel of Figure 3 — Figure 3: Density of Property Differences Between Source and Optimized Molecules Across Tasks under different similarity condition. The figure shows kernel density distributions of the property difference for each task.

Similarity	DRD2			HIA			Mutag $\downarrow$			pLogP			GSK3 $\beta$
Similarity	SR	SIM	RI	SR	SIM	RI	SR	SIM	RI	SR	SIM	RI	SR	SIM	RI
0.3–0.4	99.8	0.19	25.55	98.2	0.41	5.09	97.6	0.39	0.59	97.6	0.27	12.98	93.4	0.57	7.25
0.4–0.5	96.0	0.41	12.87	97.6	0.56	3.76	98.0	0.57	0.55	96.4	0.50	8.99	97.4	0.65	5.32
0.5–0.6	89.0	0.60	6.08	97.6	0.57	3.83	98.2	0.60	0.55	95.4	0.56	9.26	97.4	0.66	4.60
0.6–0.7	84.4	0.64	5.28	97.6	0.61	3.60	99.6	0.64	0.52	98.8	0.60	8.44	99.2	0.67	4.05
0.7–0.8	82.8	0.68	4.12	99.4	0.65	3.41	99.4	0.67	0.44	99.6	0.61	8.46	99.2	0.70	4.08

Table 1: Single-property optimization results by similarity bin. Rows indicate similarity intervals and columns correspond to tasks. Full results appear in Table 10 .

Gap Range	DRD2			HIA			Mutag $\downarrow$			pLogP			GSK3 $\beta$
Gap Range	SR	SIM	RI	SR	SIM	RI	SR	SIM	RI	SR	SIM	RI	SR	SIM	RI
0–0.1	76.0	0.67	3.88	99.4	0.64	3.75	98.8	0.68	0.45	99.8	0.60	8.97	99.6	0.70	5.17
0.1–0.2	76.4	0.68	4.12	99.4	0.66	3.37	99.6	0.66	0.49	99.4	0.63	7.69	98.0	0.72	2.71
0.2–0.3	79.6	0.69	3.30	99.0	0.66	3.55	99.8	0.67	0.48	99.8	0.64	6.57	98.6	0.73	2.74
0.3–0.4	75.8	0.70	2.92	99.4	0.66	3.36	99.6	0.67	0.48	99.8	0.64	6.46	97.4	0.72	2.20
0.4–0.5	75.8	0.70	2.65	99.4	0.66	2.97	99.6	0.67	0.50	99.8	0.66	5.74	97.2	0.72	1.55

Table 2: Single-property optimization results by property-gap thresholds. Each row corresponds to a property-difference percentile range (the first row is the top 10% largest gaps), and each column to a task. Full results are reported in Table 11 .

Model	BBBP			DRD2			HIA			Mutag			pLogP			QED			GSK3 $\beta$			JNK3
Model	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$
LLaMA (0-shot)	30.2	0.81	0.45	12.8	0.70	0.02	30.4	0.66	3.02	37.4	0.76	0.24	17.0	0.74	2.41	18.0	0.75	0.19	24.6	0.70	0.72	19.6	0.68	0.26
LLaMA (5-shot)	78.8	0.23	1.57	71.4	0.15	15.08	72.2	0.27	5.46	74.4	0.18	0.63	62.6	0.18	8.93	71.6	0.15	1.12	74.0	0.16	11.02	69.6	0.15	11.50
Mistral (0-shot)	30.4	0.72	0.51	17.6	0.72	0.72	57.2	0.72	1.28	41.0	0.69	0.33	32.2	0.74	2.71	12.8	0.67	0.21	29.0	0.67	0.71	23.8	0.64	0.27
Mistral (5-shot)	70.6	0.67	0.93	28.6	0.64	1.92	72.2	0.68	1.88	59.8	0.63	0.39	64.0	0.70	3.27	50.0	0.68	0.27	50.6	0.69	1.40	49.6	0.67	1.58
GPT-4o (0-shot)	56.0	0.70	0.66	19.0	0.73	0.25	58.4	0.71	1.40	44.0	0.70	0.27	43.2	0.72	3.20	38.4	0.75	0.24	41.8	0.68	0.91	30.4	0.67	0.27

Table 3: Single-property results. Best values are bold, DPO results are highlighted in gray.

Model	BDP			BDQ			BPQ			DPQ			BDPQ
Model	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$
Mistral (0-shot)	6.6	0.81	0.68	3.0	0.76	0.53	15.8	0.73	0.51	2.2	0.65	0.41	3.2	0.77	0.87
LLaMA (0-shot)	22.0	0.73	0.74	2.2	0.64	0.53	28.4	0.64	0.72	2.6	0.62	0.32	5.2	0.80	0.62
Claude-3.5 (0-shot)	19.6	0.66	1.05	13.0	0.62	1.14	56.0	0.62	0.86	11.0	0.54	0.51	8.0	0.60	1.34
GPT-4o (0-shot)	7.8	0.69	0.90	2.0	0.69	0.62	36.4	0.73	0.42	2.8	0.57	0.50	1.8	0.71	0.39
Mistral (5-shot)	35.2	0.64	2.10	17.0	0.60	2.32	68.6	0.63	0.79	10.4	0.54	1.10	11.0	0.69	0.96

Table 4: Multi-property results. Best values are bold, DPO results are highlighted in gray.

Param	BDP			BDPQ			BDQ			BPQ			DPQ
Param	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$
lr1e-05+beta0.1	7.92	92.2	0.33	8.32	82.0	0.27	8.60	89.6	0.41	2.17	96.8	0.34	7.29	83.2	0.30
lr1e-05+beta0.3	6.46	85.4	0.40	7.02	72.6	0.32	6.99	88.2	0.47	1.98	97.4	0.40	5.49	77.2	0.36
lr1e-05+beta0.5	6.40	84.8	0.42	6.63	72.2	0.34	6.82	86.8	0.48	1.93	97.8	0.41	4.93	76.8	0.38
lr5e-06+beta0.1	6.11	85.6	0.42	6.42	74.4	0.35	7.17	88.0	0.46	1.93	97.2	0.41	5.49	77.6	0.37
lr5e-06+beta0.3	5.48	83.2	0.46	5.85	66.4	0.39	6.16	84.8	0.51	1.78	96.8	0.44	4.25	72.2	0.41

Table 5: DPO results under different hyperparameter settings. Rows correspond to hyperparameter combinations; columns correspond to tasks.

Panel of Figure 4 — Figure 4: Visualization of DPO hyperparameter trends. By plotting the data points from Table 5 and fitting a curve, we observe a clear negative correlation. (a) SIM vs. SR; (b) SIM vs. RI.

Method	DRD2			GSK3 $\beta$			pLogP			QED
Method	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$
GraphGA (20)	30.80	0.210	12.86	100.00	0.077	15.48	100.00	0.154	11.00	100.00	0.159	1.20
JTVAE (20)	99.60	0.111	477.19	84.60	0.122	5.86	100.00	0.124	15.79	99.80	0.151	1.16
REINVENT (20)	52.00	0.188	38.41	57.59	0.165	0.81	100.00	0.105	19.06	100.00	0.138	1.32
SFT-Mistral	98.10	0.570	8.65	98.80	0.590	10.26	99.95	0.570	9.81	100.00	0.570	0.81
DPO-Mistral	95.60	0.547	14.82	98.25	0.520	17.12	99.60	0.510	17.43	99.53	0.533	0.92

Table 6: Comparison with representative non-LLM baselines under a similarity-aware PMO objective. The main comparison uses a generation budget of 20.

Method	DRD2+pLogP+QED			GSK3 $\beta$ +DRD2+pLogP			GSK3 $\beta$ +QED+DRD2			GSK3 $\beta$ +QED+pLogP
Method	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$
GraphGA (20)	9.0	0.154	3.13	19.9	0.144	11.04	5.9	0.101	6.30	4.1	0.153	0.60
JTVAE (20)	51.0	0.101	2.44	69.3	0.115	8.02	30.7	0.101	13.82	91.8	0.094	1.79
REINVENT (20)	35.8	0.112	2.05	4.8	0.137	7.86	0.8	0.135	1.47	85.4	0.109	1.69
SFT-Llama	57.2	0.530	2.17	70.4	0.610	17.55	67.6	0.620	9.51	85.0	0.580	1.71
DPO-Llama	68.0	0.420	5.74	82.8	0.520	51.67	72.4	0.530	21.45	92.6	0.500	3.35

Table 7: Compositional generalization to unseen three-property optimization tasks. LLM models are trained only on single-property and pairwise-property optimization data. Non-LLM baselines are evaluated under the same PMO setting with a similarity-aware fitness, and the main comparison uses a matched generation budget of 20.

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

	Mutagenicity	pLogP	QED	BBBP	HIA	GSK3 $\beta$	JNK3	DRD2
N	110586	105949	157863	42296	13929	47168	22711	116755
mean	-0.21	1.88	0.18	0.32	0.34	0.17	0.10	0.40
std	0.10	1.52	0.09	0.11	0.21	0.07	0.08	0.17
var	0.01	2.32	0.01	0.01	0.05	0.01	0.01	0.03
min	-0.85	1.00	0.10	0.20	0.10	0.11	0.06	0.20
q1	-0.26	1.28	0.12	0.24	0.16	0.12	0.06	0.26

Appendix

Table 8: Summary statistics of single-property training data.

Properties	Count	Properties	Count
GSK3 $\beta$ +DRD2 +pLogP	9,469	JNK3+QED	8,152
GSK3 $\beta$ +DRD2 +QED	13,898	JNK3+pLogP	10,609
GSK3 $\beta$ +pLogP +QED	7,268	GSK3 $\beta$ +QED	23,251
DRD2+pLogP +QED	2,349	GSK3 $\beta$ +pLogP	19,571
JNK3+GSK3 $\beta$	9,885	pLogP+QED	15,874

Appendix

Table 9: Summary statistics of multi-property training data.

Panel of Figure 5 — Figure 5: Molecular pair examples optimized for the DRD2 property. Green denotes the original molecule, while red denotes the optimized molecule. (a) shows the SFT-optimized results, (b) shows the DPO-optimized results.

Panel of Figure 6 — Figure 6: Molecular pair example optimized with different hyperparameter configurations. In the subfigure, the three rows (top to bottom) correspond to the original and optimized molecules at $\beta=$ 0.1, 0.3 and 0.5 respectively. Green denotes the original molecule, while red denotes the optimized molecule.

Panel of Figure 7 — Figure 7: Density of Property Differences in Training Data Across Tasks Under Different Similarity Conditions. The figure presents the kernel density distributions of property differences in the training dataset for each task.

Panel of Figure 8 — Figure 8: Density of Property Differences Between Source and Optimized Molecules Across Tasks under different similarity condition. The figure shows kernel density distributions of the property difference for each task. Subfig (f) is a reduction-oriented task, so its distribution trends in the opposite direction (toward negative values).

Panel of Figure 9 — Figure 9: Distribution of Tanimoto similarity between optimized molecules and original molecules under different similarity conditions.

Similarity	JNK3			BBBP			DRD2			HIA			Mutag $\downarrow$			pLogP			QED			GSK3 $\beta$
Similarity	SR	SIM	RI	SR	SIM	RI	SR	SIM	RI	SR	SIM	RI	SR	SIM	RI	SR	SIM	RI	SR	SIM	RI	SR	SIM	RI
0.3–0.4	93.0	0.48	2.94	98.8	0.39	1.83	99.8	0.19	25.55	98.2	0.41	5.09	97.6	0.39	0.59	97.6	0.27	12.98	99.0	0.35	1.10	93.4	0.57	7.25
0.4–0.5	95.8	0.57	2.39	98.4	0.57	1.50	96.0	0.41	12.87	97.6	0.56	3.76	98.0	0.57	0.55	96.4	0.50	8.99	99.4	0.49	0.88	97.4	0.65	5.32
0.5–0.6	96.6	0.59	2.31	99.8	0.59	1.56	89.0	0.60	6.08	97.6	0.57	3.83	98.2	0.60	0.55	95.4	0.56	9.26	99.8	0.55	0.87	97.4	0.66	4.60
0.6–0.7	97.2	0.62	2.23	99.6	0.61	1.57	84.4	0.64	5.28	97.6	0.61	3.60	99.6	0.64	0.52	98.8	0.60	8.44	99.8	0.60	0.83	99.2	0.67	4.05
0.7–0.8	99.0	0.66	2.27	100.0	0.65	1.49	82.8	0.68	4.12	99.4	0.65	3.41	99.4	0.67	0.44	99.6	0.61	8.46	99.6	0.61	0.80	99.2	0.70	4.08

Appendix

Table 10: Single-property optimization results by similarity bin. Rows indicate similarity intervals and columns correspond to tasks.

Property Range	BBBP			DRD2			HIA			Mutag			pLogP			QED			GSK3 $\beta$			JNK3
Property Range	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$
0–0.1	100.0	0.66	1.48	76.0	0.67	3.88	99.4	0.64	3.75	98.8	0.68	0.45	99.8	0.60	8.97	99.6	0.61	0.79	99.6	0.70	5.17	99.6	0.65	2.76
0.1–0.2	100.0	0.66	1.49	76.4	0.68	4.12	99.4	0.66	3.37	99.6	0.66	0.49	99.4	0.63	7.69	99.6	0.62	0.77	98.0	0.72	2.71	98.2	0.68	1.49
0.2–0.3	100.0	0.65	1.45	79.6	0.69	3.30	99.0	0.66	3.55	99.8	0.67	0.48	99.8	0.64	6.57	99.8	0.63	0.70	98.6	0.73	2.74	96.4	0.67	0.98
0.3–0.4	100.0	0.67	1.37	75.8	0.70	2.92	99.4	0.66	3.36	99.6	0.67	0.48	99.8	0.64	6.46	99.4	0.65	0.66	97.4	0.72	2.20	95.2	0.70	0.96
0.4–0.5	100.0	0.66	1.36	75.8	0.70	2.65	99.4	0.66	2.97	99.6	0.67	0.50	99.8	0.66	5.74	100.0	0.65	0.62	97.2	0.72	1.55	96.8	0.71	0.86

Appendix

Table 11: Single-property optimization results by property-gap thresholds. Each row corresponds to a property-difference percentile range (the first row is the top 10% largest gaps), and each column to a task.

Model	BBBP			QED			pLogP			HIA
Model	SR $\uparrow$	RI $\uparrow$	SIM $\uparrow$	SR $\uparrow$	RI $\uparrow$	SIM $\uparrow$	SR $\uparrow$	RI $\uparrow$	SIM $\uparrow$	SR $\uparrow$	RI $\uparrow$	SIM $\uparrow$
$\text{SFT-LLM}_{Mistral}$	99.55 $\pm$ 0.10	1.755 $\pm$ 0.013	0.560 $\pm$ 0.000	100.00 $\pm$ 0.00	0.810 $\pm$ 0.000	0.570 $\pm$ 0.000	99.95 $\pm$ 0.10	9.808 $\pm$ 0.076	0.570 $\pm$ 0.000	98.95 $\pm$ 0.10	5.048 $\pm$ 0.072	0.560 $\pm$ 0.000
$\text{DPO-LLM}_{Mistral}$	99.533 $\pm$ 0.115	1.973 $\pm$ 0.015	0.520 $\pm$ 0.010	99.533 $\pm$ 0.115	0.920 $\pm$ 0.000	0.533 $\pm$ 0.006	99.60 $\pm$ 0.20	17.427 $\pm$ 0.051	0.510 $\pm$ 0.065	97.333 $\pm$ 0.416	6.740 $\pm$ 0.207	0.493 $\pm$ 0.006
$\text{SFT-LLM}_{LLaMA}$	99.733 $\pm$ 0.115	1.757 $\pm$ 0.006	0.577 $\pm$ 0.006	99.80 $\pm$ 0.00	0.860 $\pm$ 0.000	0.570 $\pm$ 0.000	99.933 $\pm$ 0.115	9.017 $\pm$ 0.047	0.570 $\pm$ 0.000	98.467 $\pm$ 0.306	5.687 $\pm$ 0.093	0.560 $\pm$ 0.000
$\text{DPO-LLM}_{LLaMA}$	99.533 $\pm$ 0.115	1.987 $\pm$ 0.047	0.510 $\pm$ 0.006	99.80 $\pm$ 0.503	0.977 $\pm$ 0.015	0.543 $\pm$ 0.006	99.467 $\pm$ 0.260	18.097 $\pm$ 0.027	0.480 $\pm$ 0.006	99.467 $\pm$ 0.115	7.460 $\pm$ 0.166	0.520 $\pm$ 0.006

Appendix

Table 12: Appendix single-property results (Part I): BBBP, QED, pLogP, and HIA. Results are reported as mean

\pm

std over runs. Best mean values in each column are bold.

Model	DRD2			Mutag			GSK3 $\beta$			JNK3
Model	SR $\uparrow$	RI $\uparrow$	SIM $\uparrow$	SR $\uparrow$	RI $\uparrow$	SIM $\uparrow$	SR $\uparrow$	RI $\uparrow$	SIM $\uparrow$	SR $\uparrow$	RI $\uparrow$	SIM $\uparrow$
$\text{SFT-LLM}_{Mistral}$	98.10 $\pm$ 0.258	8.655 $\pm$ 0.142	0.570 $\pm$ 0.000	99.80 $\pm$ 0.00	0.675 $\pm$ 0.006	0.570 $\pm$ 0.000	98.80 $\pm$ 0.033	10.256 $\pm$ 0.153	0.590 $\pm$ 0.000	98.20 $\pm$ 0.10	3.940 $\pm$ 0.024	0.580 $\pm$ 0.000
$\text{DPO-LLM}_{Mistral}$	95.60 $\pm$ 0.529	14.820 $\pm$ 0.108	0.547 $\pm$ 0.006	99.40 $\pm$ 0.20	0.853 $\pm$ 0.006	0.520 $\pm$ 0.016	98.25 $\pm$ 0.486	17.116 $\pm$ 0.982	0.520 $\pm$ 0.000	98.90 $\pm$ 0.166	8.590 $\pm$ 0.049	0.490 $\pm$ 0.006
$\text{SFT-LLM}_{LLaMA}$	96.60 $\pm$ 0.529	7.760 $\pm$ 0.313	0.583 $\pm$ 0.006	99.533 $\pm$ 0.115	0.670 $\pm$ 0.008	0.557 $\pm$ 0.006	98.60 $\pm$ 0.447	12.790 $\pm$ 0.046	0.510 $\pm$ 0.005	98.60 $\pm$ 0.156	5.770 $\pm$ 0.066	0.550 $\pm$ 0.000
$\text{DPO-LLM}_{LLaMA}$	97.80 $\pm$ 1.858	10.630 $\pm$ 0.154	0.550 $\pm$ 0.012	99.00 $\pm$ 0.00	0.780 $\pm$ 0.006	0.520 $\pm$ 0.010	98.00 $\pm$ 0.146	20.010 $\pm$ 0.650	0.490 $\pm$ 0.002	98.20 $\pm$ 0.450	7.810 $\pm$ 0.136	0.470 $\pm$ 0.005

Appendix

Table 13: Appendix single-property results (Part II): DRD2, Mutag, GSK3

\beta

, and JNK3. Results are reported as mean

\pm

std over runs. Best mean values in each column are bold.

Panel of Figure 10 — Figure 10: Density of Property Differences Between Source and Optimized Molecules Across Tasks. The figure shows kernel density distributions of the property difference for each task. Subfig (f) is a reduction-oriented task, so its distribution trends in the opposite direction (toward negative values).

Model	BDP			BDQ			BPQ
Model	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$
$\text{SFT-LLM}_{Mistral}$	81.60 $\pm$ 0.67	0.560 $\pm$ 0.000	3.87 $\pm$ 0.05	83.95 $\pm$ 0.77	0.550 $\pm$ 0.000	4.98 $\pm$ 0.07	97.05 $\pm$ 0.34	0.538 $\pm$ 0.005	1.45 $\pm$ 0.01
$\text{DPO-LLM}_{Mistral}$	81.93 $\pm$ 0.50	0.497 $\pm$ 0.006	5.15 $\pm$ 0.19	85.87 $\pm$ 0.50	0.517 $\pm$ 0.006	6.22 $\pm$ 0.06	97.07 $\pm$ 0.12	0.467 $\pm$ 0.006	1.74 $\pm$ 0.03
$\text{SFT-LLM}_{LLaMA}$	78.60 $\pm$ 0.59	0.558 $\pm$ 0.005	3.64 $\pm$ 0.10	78.95 $\pm$ 1.20	0.563 $\pm$ 0.005	4.49 $\pm$ 0.09	94.55 $\pm$ 0.47	0.495 $\pm$ 0.006	1.55 $\pm$ 0.02
$\text{DPO-LLM}_{LLaMA}$	84.07 $\pm$ 0.44	0.470 $\pm$ 0.010	6.67 $\pm$ 0.33	86.67 $\pm$ 1.70	0.510 $\pm$ 0.000	6.52 $\pm$ 0.04	97.13 $\pm$ 0.99	0.450 $\pm$ 0.010	1.98 $\pm$ 0.06

Appendix

Table 14: Multi-property results in the appendix (Part I). Results are reported as mean

\pm

std over runs. Best mean values in each column are bold.

Model	DPQ			BDPQ
Model	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$
$\text{SFT-LLM}_{Mistral}$	62.85 $\pm$ 1.91	0.513 $\pm$ 0.005	2.42 $\pm$ 0.11	57.00 $\pm$ 1.49	0.523 $\pm$ 0.010	3.05 $\pm$ 0.07
$\text{DPO-LLM}_{Mistral}$	70.13 $\pm$ 0.42	0.443 $\pm$ 0.006	4.30 $\pm$ 0.19	65.13 $\pm$ 1.36	0.430 $\pm$ 0.010	5.65 $\pm$ 0.18
$\text{SFT-LLM}_{LLaMA}$	57.60 $\pm$ 0.94	0.510 $\pm$ 0.006	2.12 $\pm$ 0.09	46.15 $\pm$ 0.87	0.498 $\pm$ 0.010	3.36 $\pm$ 0.11
$\text{DPO-LLM}_{LLaMA}$	68.33 $\pm$ 1.94	0.460 $\pm$ 0.010	3.39 $\pm$ 0.19	59.20 $\pm$ 1.80	0.420 $\pm$ 0.010	5.97 $\pm$ 0.41

Appendix

Table 15: Multi-property results in the appendix (Part II). Results are reported as mean

\pm

std over runs. Best mean values in each column are bold.

Method	Budget	DRD2			GSK3 $\beta$			pLogP			QED
Method	Budget	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$
GraphGA	20	30.80	0.210	12.86	100.00	0.077	15.48	100.00	0.154	11.00	100.00	0.159	1.20
	50	69.00	0.187	84.70	100.00	0.077	15.48	100.00	0.113	16.18	100.00	0.164	1.34
	100	93.80	0.138	324.20	100.00	0.078	15.39	100.00	0.113	16.18	100.00	0.191	1.32
	200	93.40	0.142	453.98	100.00	0.091	15.63	100.00	0.112	19.90	100.00	0.199	1.31
JTVAE	20	99.60	0.111	477.19	84.60	0.122	5.86	100.00	0.124	15.79	99.80	0.151	1.16

Appendix

Table 16: Budget sensitivity of non-LLM baselines on single-property optimization tasks under the PMO framework.

Method	Budget	DRD2+pLogP+QED			GSK3 $\beta$ $\beta$ +DRD2+pLogP			GSK3 $\beta$ $\beta$ +QED+DRD2			GSK3 $\beta$ +QED+pLogP
Method	Budget	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$	SR $\uparrow$	SIM $\uparrow$	RI $\uparrow$
GraphGA	20	9.0	0.154	3.13	19.9	0.144	11.04	5.9	0.101	6.30	4.1	0.153	0.60
	50	2.6	0.128	1.57	0.0	–	–	9.3	0.133	8.76	0.0	–	–
	100	2.6	0.128	1.57	0.0	–	–	78.8	0.098	45.50	0.0	–	–
	200	26.2	0.099	3.15	41.6	0.110	32.18	79.3	0.095	63.65	40.1	0.097	1.93
JTVAE	20	51.0	0.101	2.44	69.3	0.115	8.02	30.7	0.101	13.82	91.8	0.094	1.79

Appendix

Table 17: Budget sensitivity of non-LLM baselines on three-property optimization tasks under the PMO framework. Entries with no successful molecules have undefined SIM/RI and are marked as –.

Molecular Optimization

Molecular optimization computationally designs or edits chemical structures to improve desired properties while respecting chemical and practical constraints.

See all 103 papers

C-MORAL: Controllable Multi-Objective Molecular Optimization with Reinforcement Alignment for LLMs
Apr 24, 2026Rui Gao, Youngseung Jeon, Swastik Roy +2RL for Language Models Molecular Optimization
University of California, Los Angeles · Amazon
FORGE: Fragment-Oriented Ranking and Generation for Context-Aware Molecular Optimization
May 11, 2026Qingchuan Zhang, He Cao, Hao Li +6Molecular Optimization Large Language Model-Guided Optimization
University of Science and Technology of China · International Digital Economy Academy · Peking University +5
Small Molecule Optimization with Large Language Models
Jul 26, 2024Philipp Guevorguian, Menua Bedrosian, Tigran Fahradyan +3Molecular Optimization Evolutionary Optimization
YerevaNN research lab, Yerevan, Armenia.

Apr 24, 2026J/K move · Enter open · S save

C-MORAL: Controllable Multi-Objective Molecular Optimization with Reinforcement Alignment for LLMs

Rui Gao, Youngseung Jeon, Swastik Roy, Morteza Ziyadi+1

University of California, Los Angeles · Amazon

RL for Language Models Molecular Optimization Large Language Model-Guided Optimization Multi-Objective Language Model Alignment Drug Discovery

Open paper AskPDF arXiv

Large language models (LLMs) show promise for molecular optimization, but aligning them with selective and competing drug-design constraints remains challenging. We propose C-Moral, a reinforcement learning post-training framework for controllable multi-objective molecular optimization. C-Moral combines group-based relative optimization, property score alignment for heterogeneous objectives, and bottleneck-sensitive non-linear reward aggregation to improve stability across competing molecular properties. Experiments on C-MuMOInstruct and S

^2

-Bench MolOpt show that C-Moral achieves the best performance among compared methods on both benchmarks. On C-MuMOInstruct, C-Moral achieves the best Success Optimized Rate (SOR) of 48.9% on in-domain tasks and 39.5% on out-of-domain tasks while preserving scaffold similarity. On S

^2

-Bench MolOpt, it also achieves the strongest results across LogP, MR, and QED optimization tasks. These results suggest that C-Moral is an effective way to align molecular LLMs with continuous and constrained molecular design objectives. Our code and models are publicly available at https://github.com/Rwigie/C-MORAL.

Abstract

Molecular Optimization

Explore similar work

C-MORAL: Controllable Multi-Objective Molecular Optimization with Reinforcement Alignment for LLMs

FORGE: Fragment-Oriented Ranking and Generation for Context-Aware Molecular Optimization

Small Molecule Optimization with Large Language Models