FRAGMENTA: Efficient End-to-end Fragmentation-based Generative Model with Agentic Tuning for Drug Lead Optimization in Small Data Regime

Yuto Suzuki; Paul Awolade; Daniel V. LaBarbera; Farnoush Banaei-Kashani

cs.AINov 25, 2025

FRAGMENTA: Efficient End-to-end Fragmentation-based Generative Model with Agentic Tuning for Drug Lead Optimization in Small Data Regime

Authors: Yuto Suzuki, Paul Awolade, Daniel V. LaBarbera, Farnoush Banaei-Kashani

Organizations: Department of Computer Science and Engineering University of Colorado Denver Denver, CO, USA · Skaggs School of Pharmacy University of Colorado Anschutz Medical Campus Aurora, CO, USA

Abstract

Molecule generation from extremely limited training data is a key challenge in drug discovery. Existing fragment-based methods are more suitable than atom-based approaches in this regime, but typically optimize fragment selection separately from downstream generation. Expert feedback is also especially valuable with limited data, yet translating such feedback into model objectives usually requires AI engineering expertise. We introduce FRAGMENTA, an end-to-end framework for small-data drug lead optimization with two components: (1) LVSEF, a fragment-based generator that jointly optimizes fragmentation and generation through a tabular reward-update mechanism, and (2) an agentic system that converts conversational expert feedback into updated generative objectives. Across three small-data datasets (11--104 molecules), LVSEF outperforms state-of-the-art methods in the smallest-data settings, matches them at larger scales, and trains ${\sim}16\times$ faster. On three public protein targets, iterative closed-loop optimization improves final-round discovery yield by up to ${\sim}16%$ over one-shot LVSEF-only on kinase, with gains depending on how well feedback matches target chemistry. In a real-world cancer drug-discovery deployment, Human-Agent FRAGMENTA identified nearly twice as many molecules with favorable docking scores ( $< -6$ ) as baseline methods.

Figures & tables

Figure 2. Overview of FRAGMENTA configurations. (a) Traditional human-in-the-loop approach with human medicinal chemists and AI engineers. (b) Semi-autonomous system where our agentic framework replaces the AI engineer. (c) Fully autonomous agent-to-agent system where both roles are automated.

Model	Dis w/ $\uparrow$	Dis w/o $\uparrow$	Valid (%) $\uparrow$	RS $\uparrow$	Unique (%) $\uparrow$	Novel (%) $\uparrow$	Cham. $\uparrow$	Div. $\uparrow$	Mem. (%) $\uparrow$
GraphNVP	-	-	0	-	-	-	-	-	-
HierVAE (w/ft)	1.7	3.7	100	0.93	5	81	0.56	0.77	44
HierVAE (w ft)	6.3	6.4	100	0.54	31	98	0.48	0.87	98
MiCaM	0	0.3	100	0.67	0.7	98	0.81	0.46	0
DEG	6.0	6.1	100	0.51	31	98	0.48	0.87	97
LVSEF (ran)	9.7	12.1	100	0.55	47	96	0.56	0.88	87

Table 1. Comparative Results with Chain Extenders Dataset

Model	Dis w/ $\uparrow$	Dis w/o $\uparrow$	Valid (%) $\uparrow$	RS $\uparrow$	Unique (%) $\uparrow$	Novel (%) $\uparrow$	Cham. $\uparrow$	Div. $\uparrow$	Mem. (%) $\uparrow$
GraphNVP	-	-	0	-	-	-	-	-	-
HierVAE (w/ft)	0	2.6	100	0.98	3	100	0.44	0.67	0
HierVAE (w.ft)	0	7.8	100	0.56	32	95	0.20	0.77	0
MiCaM	0	0.8	100	0.67	1.2	100	0.73	0.58	0
DEG	3.9	13.6	100	0.33	70	100	0.63	0.87	45
LVSEF (ran)	6.2	19.2	100	0.35	72	100	0.60	0.84	51

Table 2. Comparative Results with Acrylates Dataset

Model	Valid (%) $\uparrow$	Unique (%) $\uparrow$	Novel (%) $\uparrow$	Diversity $\uparrow$	QED $\uparrow$	SA $\downarrow$	Lipinski (%) $\uparrow$	Scaffold Div. $\uparrow$
GraphNVP	-	-	-	-	-	-	-	-
HierVAE(w/o ft)	100.0	51.0	92.2	0.864	0.565	2.411	99.0	0.27
MiCaM	100.0	3.0	66.7	0.097	0.605	2.278	100.0	0.03
DEG	100.0	96.0	100.0	0.838	0.754	2.790	100.0	0.88
LVSEF (ran)	100.0	99.0	100.0	0.849	0.730	2.706	100.0	0.90

Table 3. Comparative Results with Our Internal Dataset

Condition	Hit rate (%) $\uparrow$	Hits ( $\leq-6$ ) $\uparrow$	Top-10 Vina $\downarrow$	Best Vina $\downarrow$	QED $\uparrow$	SA $\downarrow$	Valid (%) $\uparrow$	Unique (%) $\uparrow$
Surrogate-Expert–Agent	32.2 ± 6.2%	32.2 ± 6.2	-8.44 ± 0.38	-10.01 ± 0.92	0.541 ± 0.018	0.529 ± 0.021	100.0	99.6 $\pm$ 0.7
Agent-Agent	29.2 ± 3.9%	29.2 ± 3.9	-8.51 ± 0.33	-9.88 ± 0.65	0.532 ± 0.017	0.531 ± 0.021	100.0	99.8 $\pm$ 0.6
Binary feedback	30.0 ± 7.7%	30.0 ± 7.7	-8.38 ± 0.35	-9.53 ± 0.85	0.535 ± 0.032	0.536 ± 0.016	100.0	100.0 $\pm$ 0.0
Scalar feedback	28.8 ± 7.7%	28.8 ± 7.7	-8.46 ± 0.44	-10.34 ± 0.96	0.521 ± 0.021	0.535 ± 0.021	100.0	100.0 $\pm$ 0.0
DEG	26.6 ± 6.5%	26.6 ± 6.5	-8.13 ± 0.36	-9.51 ± 0.99	0.502 ± 0.019	0.513 ± 0.028	100.0	100.0 $\pm$ 0.0
LVSEF-only	27.8 ± 7.6%	27.8 ± 7.6	-8.30 ± 0.21	-9.71 ± 1.35	0.537 ± 0.031	0.535 ± 0.031	100.0	100.0 $\pm$ 0.0

Table 4. Final-round performance on Kinase (FAK1) (round 6 for feedback conditions; round 1 for one-shot baselines). Five seeds (

n{=}5

). Mean

\pm

95% CI half-width.

Condition	Hit rate (%) $\uparrow$	Hits ( $\leq-6$ ) $\uparrow$	Top-10 Vina $\downarrow$	Best Vina $\downarrow$	QED $\uparrow$	SA $\downarrow$	Valid (%) $\uparrow$	Unique (%) $\uparrow$
Surrogate-Expert–Agent	28.4 ± 4.0%	28.4 ± 4.0	-8.05 ± 0.18	-9.11 ± 0.14	0.502 ± 0.015	0.508 ± 0.014	100.0	98.6 $\pm$ 1.4
Agent-Agent	30.4 ± 5.6%	30.4 ± 5.6	-8.43 ± 0.22	-10.02 ± 1.01	0.501 ± 0.032	0.525 ± 0.018	100.0	99.6 $\pm$ 0.7
Binary feedback	25.2 ± 4.3%	25.2 ± 4.3	-7.98 ± 0.09	-9.11 ± 0.19	0.498 ± 0.032	0.505 ± 0.021	100.0	99.6 $\pm$ 0.7
Scalar feedback	27.4 ± 4.4%	27.4 ± 4.4	-8.16 ± 0.47	-9.58 ± 1.37	0.495 ± 0.023	0.526 ± 0.024	100.0	99.2 $\pm$ 2.2
DEG	28.2 ± 6.3%	28.2 ± 6.3	-8.43 ± 0.41	-9.60 ± 0.35	0.507 ± 0.010	0.523 ± 0.024	100.0	99.0 $\pm$ 1.8
LVSEF-only	26.4 ± 5.3%	26.4 ± 5.3	-8.15 ± 0.36	-9.09 ± 0.48	0.488 ± 0.011	0.504 ± 0.019	100.0	99.6 $\pm$ 0.7

Table 5. Final-round performance on GPCR (CXCR4) (round 6 for feedback conditions; round 1 for one-shot baselines). Five seeds (

n{=}5

). Mean

\pm

95% CI half-width.

Condition	Hit rate (%) $\uparrow$	Hits ( $\leq-6$ ) $\uparrow$	Top-10 Vina $\downarrow$	Best Vina $\downarrow$	QED $\uparrow$	SA $\downarrow$	Valid (%) $\uparrow$	Unique (%) $\uparrow$
Surrogate-Expert–Agent	18.4 ± 3.4%	18.4 ± 3.4	-7.66 ± 0.27	-8.89 ± 0.28	0.519 ± 0.017	0.527 ± 0.021	100.0	99.8 $\pm$ 0.6
Agent-Agent	18.2 ± 7.5%	18.2 ± 7.5	-7.48 ± 0.93	-8.47 ± 1.43	0.526 ± 0.028	0.524 ± 0.024	100.0	100.0 $\pm$ 0.0
Binary feedback	21.2 ± 4.2%	21.2 ± 4.2	-7.64 ± 0.36	-8.67 ± 0.84	0.520 ± 0.017	0.518 ± 0.045	100.0	99.8 $\pm$ 0.6
Scalar feedback	19.6 ± 4.4%	19.6 ± 4.4	-7.84 ± 0.31	-9.28 ± 0.69	0.534 ± 0.023	0.516 ± 0.020	100.0	99.8 $\pm$ 0.6
DEG	17.4 ± 2.4%	17.4 ± 2.4	-7.42 ± 0.19	-8.74 ± 0.68	0.520 ± 0.007	0.520 ± 0.026	100.0	100.0 $\pm$ 0.0
LVSEF-only	18.0 ± 5.6%	18.0 ± 5.6	-7.74 ± 0.40	-9.61 ± 1.60	0.513 ± 0.023	0.509 ± 0.023	100.0	99.8 $\pm$ 0.6

Table 6. Final-round performance on Nuclear receptor (MCR) (round 6 for feedback conditions; round 1 for one-shot baselines). Five seeds (

n{=}5

). Mean

\pm

95% CI half-width.

Model	Valid (%) ↑	Unique (%) ↑	Novel (%) ↑	Diversity ↑	QED ↑	SA ↓	Lipinski (%) ↑	Scaffold Div. ↑	Dock <-6 ↑
Baseline	100.0	99.0	100.0	0.849	0.730	2.706	100.0	0.90	-
Human-Human	100.0	99.0	100.0	0.867	0.738	2.897	100.0	0.87	7
Human-Agent	100.0	100.0	100.0	0.849	0.785	2.723	100.0	0.93	13
Agent-Agent	100.0	100.0	100.0	0.840	0.754	2.666	100.0	0.86	11

Table 7. Agent-based Molecular Generation Results

Obj. ver.	Intent distilled from review	Executable change
0 (round 1)	Balance drug-likeness and diversity	Shared composite surrogate: QED, SA, Lipinski, novelty, batch diversity.
2 (round 2)	Reject reactive warheads and PAINS	First LLM edit: PAINS filter, reactive-warhead hard SMARTS, hinge-binding rewards.
3 (round 3)	Favor permeable, hinge-binding cores	Adds TPSA band for permeability; expands hinge/scaffold reward SMARTS.
6 (round 6)	Refine ADME and kinase pharmacophores	Refines weights (QED, logP, TPSA, Fsp 3 , SA); expands toxicophore filters.

Table 8. How LLM review is operationalized into training objectives (Surrogate-Expert–Agent, kinase FAK1, seed 0). Appendix A.7 lists full source code.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Condition	Hit rate (%) $\uparrow$	Hits ( $\leq-6$ ) $\uparrow$	Top-10 Vina $\downarrow$	Best Vina $\downarrow$	QED $\uparrow$	SA $\downarrow$	Valid (%) $\uparrow$	Unique (%) $\uparrow$
Surrogate-Expert–Agent	33.6 ± 3.2%	33.6 ± 3.2	-8.29 ± 0.39	-9.75 ± 0.76	0.541 ± 0.019	0.538 ± 0.043	100.0	99.8 $\pm$ 0.6
Agent-Agent	32.6 ± 4.4%	32.6 ± 4.4	-8.79 ± 0.49	-10.15 ± 0.23	0.533 ± 0.016	0.533 ± 0.029	100.0	99.8 $\pm$ 0.6
Binary feedback	37.4 ± 5.2%	37.4 ± 5.2	-8.61 ± 0.67	-9.70 ± 1.11	0.541 ± 0.018	0.553 ± 0.033	100.0	99.6 $\pm$ 1.1
Scalar feedback	35.8 ± 4.2%	35.8 ± 4.2	-8.53 ± 0.51	-10.13 ± 0.68	0.556 ± 0.028	0.557 ± 0.025	100.0	99.2 $\pm$ 1.0
DEG	26.6 ± 6.5%	26.6 ± 6.5	-8.13 ± 0.36	-9.51 ± 0.99	0.502 ± 0.019	0.513 ± 0.028	100.0	100.0 $\pm$ 0.0
LVSEF-only	27.8 ± 7.6%	27.8 ± 7.6	-8.30 ± 0.21	-9.71 ± 1.35	0.537 ± 0.031	0.535 ± 0.031	100.0	100.0 $\pm$ 0.0

Appendix

Table 9. Best-round performance on Kinase (FAK1) (argmax pre-registered selection score over rounds 1–6). Five seeds (

n{=}5

). Mean

\pm

95% CI half-width.

Condition	Hit rate (%) $\uparrow$	Hits ( $\leq-6$ ) $\uparrow$	Top-10 Vina $\downarrow$	Best Vina $\downarrow$	QED $\uparrow$	SA $\downarrow$	Valid (%) $\uparrow$	Unique (%) $\uparrow$
Surrogate-Expert–Agent	32.0 ± 2.3%	32.0 ± 2.3	-8.27 ± 0.14	-9.07 ± 0.50	0.505 ± 0.026	0.526 ± 0.015	100.0	99.4 $\pm$ 1.1
Agent-Agent	34.0 ± 2.3%	34.0 ± 2.3	-8.64 ± 0.21	-9.98 ± 0.23	0.520 ± 0.021	0.532 ± 0.025	100.0	98.4 $\pm$ 1.9
Binary feedback	32.2 ± 3.3%	32.2 ± 3.3	-8.29 ± 0.27	-9.42 ± 0.49	0.502 ± 0.026	0.512 ± 0.030	100.0	99.6 $\pm$ 0.7
Scalar feedback	33.4 ± 3.2%	33.4 ± 3.2	-8.35 ± 0.45	-9.52 ± 0.74	0.512 ± 0.024	0.518 ± 0.033	100.0	99.2 $\pm$ 1.6
DEG	28.2 ± 6.3%	28.2 ± 6.3	-8.43 ± 0.41	-9.60 ± 0.35	0.507 ± 0.010	0.523 ± 0.024	100.0	99.0 $\pm$ 1.8
LVSEF-only	26.4 ± 5.3%	26.4 ± 5.3	-8.15 ± 0.36	-9.09 ± 0.48	0.488 ± 0.011	0.504 ± 0.019	100.0	99.6 $\pm$ 0.7

Appendix

Table 10. Best-round performance on GPCR (CXCR4) (argmax pre-registered selection score over rounds 1–6). Five seeds (

n{=}5

). Mean

\pm

95% CI half-width.

Condition	Hit rate (%) $\uparrow$	Hits ( $\leq-6$ ) $\uparrow$	Top-10 Vina $\downarrow$	Best Vina $\downarrow$	QED $\uparrow$	SA $\downarrow$	Valid (%) $\uparrow$	Unique (%) $\uparrow$
Surrogate-Expert–Agent	24.8 ± 3.2%	24.8 ± 3.2	-8.12 ± 0.30	-9.52 ± 1.22	0.541 ± 0.011	0.527 ± 0.033	100.0	99.8 $\pm$ 0.6
Agent-Agent	23.6 ± 3.2%	23.6 ± 3.2	-7.87 ± 0.52	-9.00 ± 1.05	0.540 ± 0.015	0.525 ± 0.018	100.0	100.0 $\pm$ 0.0
Binary feedback	25.6 ± 1.9%	25.6 ± 1.9	-8.02 ± 0.27	-9.23 ± 0.57	0.534 ± 0.026	0.527 ± 0.016	100.0	100.0 $\pm$ 0.0
Scalar feedback	23.8 ± 4.2%	23.8 ± 4.2	-7.76 ± 0.32	-8.62 ± 0.53	0.548 ± 0.009	0.528 ± 0.039	100.0	99.6 $\pm$ 0.7
DEG	17.4 ± 2.4%	17.4 ± 2.4	-7.42 ± 0.19	-8.74 ± 0.68	0.520 ± 0.007	0.520 ± 0.026	100.0	100.0 $\pm$ 0.0
LVSEF-only	18.0 ± 5.6%	18.0 ± 5.6	-7.74 ± 0.40	-9.61 ± 1.60	0.513 ± 0.023	0.509 ± 0.023	100.0	99.8 $\pm$ 0.6

Appendix

Table 11. Best-round performance on Nuclear receptor (MCR) (argmax pre-registered selection score over rounds 1–6). Five seeds (

n{=}5

). Mean

\pm

95% CI half-width.

Comparison	$\Delta$ hit rate (%)	$\Delta$ hits	$\Delta$ top-10 Vina	$\Delta$ QED
vs. LVSEF-only	+4.4 ± 5.2%	+4.4 ± 5.2	+0.13 ± 0.32	+0.004 ± 0.029
vs. scalar feedback	+3.4 ± 2.4%	+3.4 ± 2.4	-0.03 ± 0.33	+0.020 ± 0.015
vs. binary feedback	+2.2 ± 8.8%	+2.2 ± 8.8	+0.06 ± 0.32	+0.006 ± 0.011

Appendix

Table 12. Surrogate-Expert–Agent paired improvement vs. automated baselines on Kinase (FAK1) (final-round view,

n{=}5

matched seeds).

\Delta

= Surrogate-Expert–Agent minus baseline; for docking scores,

\Delta

= baseline minus Surrogate-Expert–Agent (positive = stronger binding).

Comparison	$\Delta$ hit rate (%)	$\Delta$ hits	$\Delta$ top-10 Vina	$\Delta$ QED
vs. LVSEF-only	+2.0 ± 5.6%	+2.0 ± 5.6	-0.10 ± 0.31	+0.014 ± 0.012
vs. scalar feedback	+1.0 ± 4.2%	+1.0 ± 4.2	-0.12 ± 0.23	+0.006 ± 0.023
vs. binary feedback	+3.2 ± 4.2%	+3.2 ± 4.2	+0.06 ± 0.10	+0.004 ± 0.027

Appendix

Table 13. Surrogate-Expert–Agent paired improvement vs. automated baselines on GPCR (CXCR4) (final-round view,

n{=}5

matched seeds).

\Delta

= Surrogate-Expert–Agent minus baseline; for docking scores,

\Delta

= baseline minus Surrogate-Expert–Agent (positive = stronger binding).

Comparison	$\Delta$ hit rate (%)	$\Delta$ hits	$\Delta$ top-10 Vina	$\Delta$ QED
vs. LVSEF-only	+0.4 ± 4.4%	+0.4 ± 4.4	-0.08 ± 0.34	+0.006 ± 0.022
vs. scalar feedback	-1.2 ± 4.6%	-1.2 ± 4.6	-0.18 ± 0.24	-0.016 ± 0.022
vs. binary feedback	-2.8 ± 4.4%	-2.8 ± 4.4	+0.02 ± 0.25	-0.002 ± 0.005

Appendix

Table 14. Surrogate-Expert–Agent paired improvement vs. automated baselines on Nuclear receptor (MCR) (final-round view,

n{=}5

matched seeds).

\Delta

= Surrogate-Expert–Agent minus baseline; for docking scores,

\Delta

= baseline minus Surrogate-Expert–Agent (positive = stronger binding).

Condition	Train (h)	Train + eval (h)
LVSEF-only	0.55 $\pm$ 0.12	0.72 $\pm$ 0.17
Surrogate-Expert–Agent	3.27 $\pm$ 0.76	4.25 $\pm$ 1.02
Scalar feedback	3.27 $\pm$ 0.77	4.23 $\pm$ 1.01
Binary feedback	3.28 $\pm$ 0.77	4.24 $\pm$ 1.02
Agent-Agent	3.25 $\pm$ 0.75	4.22 $\pm$ 1.00
DEG	8.81 $\pm$ 4.28	8.97 $\pm$ 4.29

Appendix

Table 15. Generative-model training cost per benchmark run (LVSEF/DEG train + evaluate subprocess wall time). Excludes Vina docking, LLM calls, and feedback logic. Mean

\pm

95% CI half-width over 15 runs per condition.

Model	Average number of atoms
DEG	3.0
LVSEF	4.2

Appendix

Table 16. Average Number of Atoms in Fragments

Condition	Kinase (FAK1)	GPCR (CXCR4)	Nuclear receptor (MCR)
Surrogate-Expert–Agent	v13: PAINS filter; 45 SMARTS patterns; 14 hard filters; TPSA shaping; logP shaping; weighted property mix	v4: 3 SMARTS patterns	v5: PAINS filter; 3 SMARTS patterns
Agent-Agent	v6: PAINS filter; 3 SMARTS patterns	v5: 4 SMARTS patterns	v6: PAINS filter; 3 SMARTS patterns; weighted property mix
Scalar feedback	v1: alpha=0.3 centroid Tanimoto	v1: alpha=0.3 centroid Tanimoto	v1: alpha=0.3 centroid Tanimoto
Binary feedback	v1: beta=0.4 accept-set Tanimoto	v1: beta=0.4 accept-set Tanimoto	v1: beta=0.4 accept-set Tanimoto
LVSEF-only	v0: weighted QED/SA/Lipinski/novelty/diversity composite	v0: LLM-authored composite objective	v0: LLM-authored composite objective
DEG	v0: LLM-authored composite objective	v0: LLM-authored composite objective	v0: LLM-authored composite objective

Appendix

Table 17. Final training objective per DUD-E condition and target (seed 0, highest objective_v* after six feedback rounds).

Round	Ver.	LLM-driven objective change	Hit (%) $\uparrow$	$\Delta$ hit	Best Vina $\downarrow$	$\Delta$ Vina
1	0	Shared composite surrogate: QED, SA, Lipinski, novelty, batch diversity.	25.0	—	-8.90	—
2	2	First LLM edit: PAINS filter, reactive-warhead hard SMARTS, hinge-binding rewards.	24.0	-1.0	-9.25	-0.3
3	3	Adds TPSA band for permeability; expands hinge/scaffold reward SMARTS.	25.0	+1.0	-9.43	-0.2
4	4	Further kinase pharmacophore rewards (pyridine/pyrimidine hinges, bioisosteres).	23.0	-2.0	-9.62	-0.2
5	5	Adds logP shaping and additional reward motifs from review feedback.	33.0	+10.0	-10.05	-0.4
6	6	Refines weights (QED, logP, TPSA, Fsp 3 , SA); expands toxicophore filters.	32.0	-1.0	-10.08	-0.0

Appendix

Table 18. Per-round objective updates and discovery gains for Surrogate-Expert–Agent on kinase (FAK1), seed 0. Each row is the objective active at the start of that round (v0 at round 1; v2–v6 after subsequent LLM/CodeAgent edits).

\Delta

columns compare to the prior round.

Round	Ver.	LLM-driven objective change	Hit (%) $\uparrow$	$\Delta$ hit	Best Vina $\downarrow$	$\Delta$ Vina
1	0	Shared composite surrogate: QED, SA, Lipinski, novelty, batch diversity.	19.0	—	-8.99	—
2	1	First LLM edit: hard filters on anionic O/N (permeability); thiophene/ester penalties.	20.0	+1.0	-9.86	-0.9
3	2	Adds Michael-acceptor, acetal, and additional reactive-warhead hard SMARTS.	27.0	+7.0	-9.71	+0.1
4	3	Rejects any formal charge; enforces MW 250–500 Da and Fsp ${}^{3}\geq 0.2$ ; adds epoxide filter.	17.0	-10.0	-10.15	-0.4
5	4	Adds pharmacophore reward SMARTS (phenol, sulfone, benzimidazole, fluorophenyl).	16.0	-1.0	-8.75	+1.4
6	5	Expands pharmacophore bonuses and rotatable-bond / complexity limits.	14.0	-2.0	-8.56	+0.2

Appendix

Table 19. Per-round objective updates and discovery gains for Surrogate-Expert–Agent on nuclear receptor (MCR), seed 0. Each row is the objective active at the start of that round (v0 at round 1; v1–v5 after subsequent LLM/CodeAgent edits).

\Delta

columns compare to the prior round.

Molecular Generation

Molecular generation develops computational methods to create chemical structures and conformations for discovery, optimization, and structure elucidation.

See all 231 papers

PairIT: Autoregressive Transformers for Low-Data Molecule Optimization
Oct 13, 2026Hannah Lawrence, Bodhi P. Vani, Ji Won Park +5Autoregressive Transformer Molecular Optimization
Genentech, South San Francisco, CA, USA · Massachusetts Institute of Technology, Cambridge, MA, USA · Prescient Design, Genentech, Basel, Switzerland +1
Sesame: Structure-Aware Molecular Generation via Spatial Density-Map Conditioning
Jun 22, 2026Konstantin Yatsenko, Arvind ThiagarajanMolecular Generation Structure-Based Drug Design
Tessel Biosciences, Inc., 750 Main Street, Cambridge, MA 02139
Fraglingo: Molecular Design via Attachment-Aware Autoregressive Fragment Generation
Sep 11, 2026Thao Nguyen, Jeonghwan Kim, Zhenhailong Wang +1Molecular Generation Molecule Design
Siebel School of Computing and Data Science University of Illinois Urbana-Champaign Urbana, IL 61801, USA

Oct 13, 2026J/K move · Enter open · S save

PairIT: Autoregressive Transformers for Low-Data Molecule Optimization

Hannah Lawrence, Bodhi P. Vani, Ji Won Park, Natasa Tagasovska+4

Genentech, South San Francisco, CA, USA · Massachusetts Institute of Technology, Cambridge, MA, USA · Prescient Design, Genentech, Basel, Switzerland+1

Autoregressive Transformer Molecular Optimization Drug Discovery Transformer

Open paper AskPDF arXiv

Generating novel molecules with desired properties is a central challenge in drug discovery. Although foundational work focused on de novo generation, practical use cases require a more complex setting: conditioning on an initial seed molecule, optimizing multiple properties, and operating in data-sparse regimes. We introduce PairIT, a flexible training framework for autoregressive transformers over SMILES sequences that enables seeded generation by learning from pairs of structurally similar molecules. By leveraging matched pairs, the model learns to generate local, chemically meaningful edits that improve a property of the seed molecule. Prompt tokens extend this framework by specifying the target property, desired direction, and approximate magnitude of change. Multiple prompt tokens can be concatenated to enable multi-property optimization at inference time. For properties where pair supervision alone is insufficient, reinforcement learning post-training can further improve optimization, providing a tunable trade-off between proximity to the seed and the optimization objective. PairIT naturally supports pretraining on a large dataset and fine-tuning on a smaller dataset, which is critical for performance in data-sparse regimes. Surprisingly, we observe diminishing returns to the scale of the pretraining dataset. We demonstrate empirically on MoleculeNet and Therapeutics Data Commons datasets that, with appropriate paired-data curation, a standard transformer model of modest size (68M parameters) can achieve robust seeded molecular optimization in low-data regimes.

Abstract

Molecular Generation

Explore similar work

PairIT: Autoregressive Transformers for Low-Data Molecule Optimization

Sesame: Structure-Aware Molecular Generation via Spatial Density-Map Conditioning

Fraglingo: Molecular Design via Attachment-Aware Autoregressive Fragment Generation