One Sequence, Many Decodings: CAGenMol-2 Recasts Drug Design as Masked Molecular Inference

Yanting Li; Enyan Dai; Lei Wang; Wen-Cai Ye; Li Liu

cs.LGSep 28, 2026

One Sequence, Many Decodings: CAGenMol-2 Recasts Drug Design as Masked Molecular Inference

Authors: Yanting Li, Enyan Dai, Lei Wang, Wen-Cai Ye, Li Liu

Organizations: The Hong Kong University of Science and Technology (Guangzhou) · Jinan University, Guangzhou

Abstract

Drug design couples property evaluation, conditional generation, structure-based design, and local optimization, yet machine learning systems typically address these capabilities with separate task-specific models. We introduce CAGenMol-2, a masked diffusion molecular language model that represents molecules, continuous scalar properties, and 3D protein pockets within a single wrapped sequence. Within this pretrained interface, downstream operations are selected by which sequence regions are observed or masked at inference, allowing one checkpoint to perform property prediction, property- and pocket-conditioned generation, and partial-constraint design without task-specific architectures or backbone fine-tuning. We further propose Adaptive Fragment Optimization (AdaFO), a gradient-free mask-and-refill search that turns the masked decoder into an iterative local molecular optimizer. On CrossDocked2020, AdaFO increases Success Rate from 30.2% to 70.8%, the best reported under this protocol, while largely preserving drug-likeness and diversity. Finally, scaffold-preserving directional editing and CRBN/VHL case studies demonstrate its use in compound design workflows spanning local molecular editing, structure-based prioritization, and downstream simulation-based screening.

Figures & tables

Figure 1: Overview of CAGenMol-2. (A) Molecules, scalar properties, and 3D protein pockets are mapped to role-tagged blocks within a single wrapped SAFE sequence via a tokeniser, value encoder, and frozen 3D encoder, respectively. At inference, varying the visible and hidden (grey) blocks recovers multiple tasks from a single checkpoint, where target blocks (colored borders) are decoded into tokens or continuous scalars with confidence. (B) The wrapped sequence is processed by a shared transformer backbone. The lower timeline illustrates a masked-diffusion trajectory unmasking the molecular block from $t{=}1$ to $t{=}0$ under a fixed condition prefix.

Figure 2: AdaFO Refinement Loop : score the pool, rerank by MMR diversity, keep Top- $K$ parents, and mask-and-refill fragments to generate children. Masking is annealed for pocket-conditioned optimization and confidence-driven by unsurety $u$ for directional editing. Top children are filtered back into the pool.

Figure 3: Quality–diversity Pareto frontier for de novo generation. Despite joint training with property and pocket conditioning, CAGenMol-2 closely tracks the GenMol frontier.

Figure 4: Representative property-conditioned distribution shifts. Conditioning moves the sampled molecular distributions toward the requested scalar targets.

Property	Hit-ID	Hit-OOD	MAE-ID	MAE-OOD	Shift
logP	0.96	0.86	0.13	0.27	+1.07
MW	0.93	0.47	0.14	0.95	+0.59
QED	0.81	0.68	0.31	0.46	+1.27
SA	0.94	0.89	0.20	0.25	+1.58
MR	0.84	0.69	0.21	0.43	+0.92
Mean	0.89	0.72	0.20	0.48	+1.09

Table 1: Single-property conditional generation. ID and OOD denote targets within the empirical distribution and its tails, respectively. MAE and directional Shift are reported in corpus standard-deviation units.

Method	Macro $R^{2}$ $\uparrow$	Macro Pearson $\uparrow$
Morgan + GBR	0.525	0.720
GIN + GBR	0.219	0.550
GIN + MLP	0.477	0.678
ChemBERTa + MLP	0.762	0.880
CAGenMol-2 (native)	0.910	0.974

Table 2: Physicochemical property prediction on the 5k held-out set. CAGenMol-2 uses its pretrained native scalar head without downstream fitting; reference predictors use separately fitted regressors. Macro averages are reported for the dimensionless

R^{2}

and Pearson metrics.

Property	Full MAE $\downarrow$	Top-25% MAE $\downarrow$	Relative drop
logP	0.235	0.167	29.1%
MW	3.89	2.10	46.0%
QED	0.036	0.014	59.9%
SA	0.364	0.223	38.9%
MR	0.981	0.683	30.3%

Table 3: Selective prediction using confidence from the native scalar head. Molecules are ranked by confidence, and MAE is reported for the top-25% highest-confidence subset.

Method	BBBP $\uparrow$	Tox21 $\uparrow$	HIV $\uparrow$	ESOL $\downarrow$	Lipo $\downarrow$
ChemBERTa3	$0.700$	$0.718$	$0.740$	$0.920$	0.758
GenMol	$0.897$	$0.753$	$0.768$	0.725	$0.948$
CAGenMol-2	0.931	0.789	0.792	$0.789$	$0.790$

Table 4: Representative MoleculeNet transfer results. Classification tasks report ROC-AUC and regression tasks report RMSE. CAGenMol-2 and the GenMol use frozen representations with trained MLP readouts.

Method	Vina $\downarrow$	HA $\uparrow$ (%)	QED $\uparrow$	SA $\uparrow$	Div $\uparrow$	SR $\uparrow$ (%)
Reference	-7.45	–	0.48	0.73	–	25.0
Pocket2Mol	-7.15	48.4	0.56	0.74	0.69	24.4
DecompDiff	-8.39	64.4	0.45	0.61	0.68	24.5
MolCRAFT	-9.25	59.1	0.46	0.62	0.61	36.1
RGA + Vina	-8.01	64.4	0.57	0.71	0.41	46.2
DecompOpt	-8.98	73.5	0.48	0.65	0.60	52.5

Table 5: Structure-based drug design on CrossDocked2020, averaged over 100 held-out pockets. Success Rate jointly requires affinity, drug-likeness, and synthesizability thresholds. The RL and AdaFO rows represent parameter-updating and inference-time optimization, respectively.

Figure 5: Representative three-iteration directional-editing trajectory. AdaFO-S decreases logP through local molecular edits while retaining similarity to the seed scaffold.

Figure 6: Representative binding-pose inspection for CRBN. CAGenMol-2 candidates is shown together in the binding site, illustrating the structures carried forward by the design-to-screening workflow.

Appendix figures & tables37 assets

Supplementary material from the paper’s appendix.

Appendix

Symbol	Meaning
$s_{1:L}$	SAFE token sequence of a molecule, length $L$
$y_{k},z_{k}$	raw / $z$ -normalised value of property $k$
$K{=}6$	number of scalar property slots
$t\in[0,1]$	diffusion time ( $0$ clean, $1$ fully masked)
$\alpha_{t}$	survival probability at time $t$
$m$	mask token

Appendix

Table 6: Recurring notation.

Slot	Mol-only position	Pocket position
pocket ( $\times 4$ )	—	$2,3,4,5$
logP	$2$	$8$
MW	$5$	$11$
QED	$8$	$14$
SA	$11$	$17$
MR	$14$	$20$

Appendix

Table 7: Absolute value-slot positions in the wrapped sequence. Maximum length

320

(

\leq 26

prefix

+\leq 256

SAFE tokens

+2

suffix, with slack). Properties in order: logP, MW, QED, SA, MR, reserved. Special tokens and the prefix are never noised.

Property	Mean $\mu_{k}$	Std $\sigma_{k}$
logP	$2.333$	$2.483$
MW	$371.558$	$131.271$
QED	$0.684$	$0.166$
SA	$3.491$	$0.860$
MR	$98.646$	$37.334$

Appendix

Table 8: Empirical corpus statistics used for evaluation target shifts, hit tolerances, and normalized errors. Model encoding uses the separate checkpoint statistics specified in the text.

Group / parameter	Value
Backbone
layers / hidden / heads	$12$ / $768$ / $12$
intermediate size	$3072$
activation	GELU
dropout (hidden / attn)	$0.1$ / $0.1$
max position / vocab	$320$ / $1899$

Appendix

Table 9: Complete hyperparameter listing for training and both inference-time optimisers.

GenMol ( $N{=}1$ )
Method / $(\tau,r)$	Validity	Uniqueness	Quality	Diversity
SAFE-GPT	$94.0{\pm}0.4$	$100.0{\pm}0.0$	$54.7{\pm}0.3$	$0.879{\pm}0.001$
GenMol w/o conf.	$96.7{\pm}0.3$	$99.3{\pm}0.2$	$53.8{\pm}1.7$	$0.896{\pm}0.001$
$(0.5,0.5)$	$100.0{\pm}0.0$	$99.7{\pm}0.1$	$84.6{\pm}0.8$	$0.818{\pm}0.001$
$(0.5,1.0)$	$99.7{\pm}0.1$	$100.0{\pm}0.1$	$83.8{\pm}0.5$	$0.832{\pm}0.002$
$(0.5,10)$	$99.8{\pm}0.1$	$99.6{\pm}0.1$	$79.1{\pm}0.9$	$0.845{\pm}0.002$

Appendix

Table 10: De novo generation, full five-setting comparison.

N{=}1000

per setting, mean

\pm

std over three seeds.

Prop	Target (label)	mean $\pm$ std	MAE	bias	hit	shift
logP	$-2$ (ood-lo)	$-1.78{\pm}0.63$	$0.47$	$+0.22$	$0.93$	$+1.85$
logP	$0$ (p10)	$0.08{\pm}0.55$	$0.38$	$+0.08$	$0.95$	$+1.10$
logP	$2.3$ (p50)	$2.14{\pm}0.49$	$0.34$	$-0.16$	$0.96$	$+0.27$
logP	$4.4$ (p90)	$4.25{\pm}0.54$	$0.26$	$-0.15$	$0.95$	$+0.58$
logP	$7$ (ood-hi)	$6.67{\pm}1.52$	$0.88$	$-0.33$	$0.81$	$+1.56$
MW	$180$ (ood-lo)	$220{\pm}49$	$54.3$	$+40.2$	$0.69$	$+0.81$

Appendix

Table 11: Single-property conditional generation, all 25 configurations. hit

=

hit@

0.5\sigma

; shift is in

\sigma

units.

Figure 7: Property distributions for single-property conditioned generation (QED). Results for in-distribution and out-of-distribution targets are shown.

Figure 8: Property distributions for single-property conditioned generation (SA). Results for in-distribution and out-of-distribution targets are shown.

Figure 9: Property distributions for single-property conditioned generation (MW). Results for in-distribution and out-of-distribution targets are shown.

Figure 10: Property distributions for single-property conditioned generation (MR). Results for in-distribution and out-of-distribution targets are shown.

Figure 11: Property distributions for single-property conditioned generation (logP). Results for in-distribution and out-of-distribution targets are shown.

Figure 12: Realised property distributions under conditioning (coloured) versus unconditional generation (grey). The prefix shifts the entire probability mass toward each requested target.

Combo	Targets	valid/500	joint hit	regime
Lipinski	logP 2.5, MW 350, QED 0.75	497	$0.76$	ID
Lead-like	QED 0.7, SA 2.5	500	$0.84$	ID
Big-lipo	logP 5, MW 550, QED 0.5, SA 4	493	$0.52$	OOD
Tiny-drug	MW 200, QED 0.85	495	$0.44$	OOD

Appendix

Table 12: Multi-property conditional generation. Per-property hits for the Lipinski combo: logP

0.92

, MW

0.85

, QED

0.85

; for lead-like: QED

0.87

, SA

0.95

$\gamma$	hit@ $0.5\sigma$	MAE $/\sigma$	validity
$1.0$	$0.805$	$0.315$	$0.996$
$1.5$	$0.808$	$0.320$	$0.998$
$2.0$	$0.802$	$0.318$	$0.996$
$3.0$	$0.798$	$0.334$	$0.993$
$4.0$	$0.807$	$0.332$	$0.989$

Appendix

Table 13: Guidance sweep, aggregated over five single-property targets.

Figure 13: Hit rate, MAE, and validity versus CFG scale

\gamma

. The near-flat hit curve shows guidance is redundant given the strong prefix.

Method	Metric	logP	MW	QED	SA	MR	Macro
CAGenMol-2	MAE	$0.235$	$3.89$	$0.036$	$0.364$	$0.981$	—
	RMSE	$0.434$	$24.6$	$0.054$	$0.434$	$6.11$	—
	$R^{2}$	$0.966$	$0.963$	$0.896$	$0.755$	$0.971$	$\mathbf{0.910}$
	Pears.	$0.987$	$0.986$	$0.951$	$0.959$	$0.989$	$\mathbf{0.974}$
ChemBERTa+MLP	MAE	$0.750$	$25.8$	$0.081$	$0.268$	$7.40$	—
ChemBERTa+MLP	$R^{2}$	$0.771$	$0.794$	$0.594$	$0.839$	$0.813$	$0.762$

Appendix

Table 14: Native property prediction and fitted reference predictors on the same 5,000-molecule evaluation set. MAE and RMSE retain each property’s units; macro averages apply to

R^{2}

and Pearson correlation.

Prop	RMSE/ std	Top-50% (rel. drop)	Top-25% (rel. drop)
logP	$2.27$	$0.185$ ( $21.3\%$ )	$0.167$ ( $29.1\%$ )
MW	$1.43$	$2.34$ ( $39.8\%$ )	$2.10$ ( $46.0\%$ )
QED	$1.82$	$0.020$ ( $43.5\%$ )	$0.014$ ( $59.9\%$ )
SA	$3.22$	$0.263$ ( $27.8\%$ )	$0.223$ ( $38.9\%$ )
MR	$1.17$	$0.684$ ( $30.2\%$ )	$0.683$ ( $30.3\%$ )

Appendix

Table 15: Confidence reliability (RMSE/std ratio;

>1

over-confident) and selective-prediction MAE with relative reduction versus the full-set MAE.

Figure 14: Reliability: predicted std versus empirical RMSE per bin. Points above the diagonal indicate over-confidence.

Figure 15: Selective prediction: rolling MAE as a function of retained coverage (sorted by confidence). The curves show the accuracy–coverage trade-off as low-confidence molecules are dropped.

Classification (ROC-AUC $\uparrow$ )
Method	BACE $\uparrow$	BBBP $\uparrow$	ClinTox $\uparrow$	SIDER $\uparrow$	Tox21 $\uparrow$	HIV $\uparrow$	MUV $\uparrow$
D-MPNN	$0.809$	$0.710$	$0.906$	$0.570$	$0.759$	$0.771$	$0.786$
AttentiveFP	$0.784$	$0.663$	$0.847$	$0.606$	$0.781$	$0.757$	$0.786$
GROVER ${}_{\text{base}}$	$0.821$	$0.700$	$0.812$	$0.648$	$0.743$	$0.625$	$0.673$
MolCLR	$0.824$	$0.722$	$0.912$	$0.589$	$0.750$	$0.781$	$0.796$
GEM	$0.856$	$0.724$	$0.901$	$0.672$	$0.781$	$0.806$	$0.817$

Appendix

Table 16: MoleculeNet transfer and published reference scores. Ours and the GenMol re-run use frozen representations with trained MLP readouts; published models follow their own training protocols. Classification: ROC-AUC (

\uparrow

); regression: RMSE (

\downarrow

). Baseline numbers transcribed from the KA-GNN and ChemBERTa3 papers and a Uni-Mol survey; GenMol is our own re-run under the identical protocol. Bold

=

best per column.

Metric	Baseline	AdaFO	$\Delta$
Vina (kcal/mol) $\downarrow$	$-7.36{\pm}1.14$	$-8.76{\pm}1.26$	$-1.40$
High Affinity (%) $\uparrow$	$53.5{\pm}37.4$	$85.1{\pm}31.7$	$+31.6$
Success Rate (%) $\uparrow$	$30.2{\pm}31.0$	$70.8{\pm}40.7$	$+40.6$
QED $\uparrow$	$0.734{\pm}0.008$	$0.705{\pm}0.026$	$-0.029$
SA (norm) $\uparrow$	$0.844{\pm}0.004$	$0.833{\pm}0.013$	$-0.011$
Diversity $\uparrow$	$0.838{\pm}0.008$	$0.818{\pm}0.017$	$-0.020$

Appendix

Table 17: SBDD 100-pocket aggregate, mean

\pm

std across pockets, pretrained baseline versus five AdaFO iterations.

Figure 16: Per-pocket Success Rate under AdaFO (y) versus the pretrained baseline (x). Each point represents one of the 100 evaluation pockets; the diagonal denotes equal success rates.

Figure 17: Per-pocket metric distributions (grey: baseline; green: AdaFO). The Vina and Success-Rate distributions shift one-sidedly.

Setting	Vina	SR (%)	Diversity
Full AdaFO	$-8.52$	$80.2$	$0.84$
$-$ annealing	$-7.34$	$71.8$	$0.83$
$-$ MMR	$-8.46$	$79.6$	$0.78$
$-$ elite	$-8.48$	$76.4$	$0.82$

Appendix

Table 18: AdaFO innovation ablation on the pilot pocket, five iterations.

$w_{q}{=}w_{a}$	Vina	SR (%)	$\Delta$ QED	$\Delta$ SA
$0.0$	$-8.62$	$82.4$	$-0.04$	$-0.02$
$0.3$	$-8.55$	$83.1$	$-0.03$	$-0.01$
$0.7$	$-8.52$	$80.2$	$+0.02$	$+0.00$

Appendix

Table 19: Reward-weight sweep on the pilot pocket;

\Delta

relative to the pretrained baseline.

Figure 18: AdaFO score trajectory over iterations on the pilot pocket: best-in-pool climbs and mean-of-Top-

K

improves as the mask anneals.

Figure 19: A three-iteration AdaFO-S trajectory decreasing logP from $3.45$ to $1.77$ while preserving the scaffold. Red marks the atoms changed at each step.

Config	Random-span SR/Sim	AdaFO-S SR/Sim
QED $\uparrow 0.5\sigma$	34 / 0.81	40 / 0.78
QED $\uparrow 1.0\sigma$	28 / 0.76	24 / 0.79
QED $\downarrow 0.5\sigma$	44 / 0.78	42 / 0.75
QED $\downarrow 1.0\sigma$	66 / 0.64	72 / 0.60
SA $\uparrow 0.5\sigma$	8 / 0.92	10 / 0.91
SA $\uparrow 1.0\sigma$	56 / 0.73	62 / 0.70

Appendix

Table 20: Directional editing, all twelve configurations, small-mask regime (

\alpha\in[0.06,0.20]

, 50 seeds). SR is the scaffold-preserving directional success rate (%); Sim is mean Tanimoto to the seed. Random-span is the former AdaFO-D (confidence radius, random location).

	DIR-only %		$\|\Delta\|/\sigma$		Sim
Axis	Rand	AdaFO-S	Rand	AdaFO-S	Rand	AdaFO-S
QED $\uparrow$	12.5	33.3	0.21	0.50	0.82	0.65
QED $\downarrow$	32.7	81.2	0.27	0.74	0.81	0.64
SA $\uparrow$	51.1	91.4	0.33	0.44	0.79	0.67
SA $\downarrow$	29.8	59.1	0.31	0.67	0.76	0.58
logP $\uparrow$	47.9	100.0	0.16	0.52	0.81	0.63

Appendix

Table 21: Single forced edit, direction-only success (60 seeds). One saliency-guided edit moves the property the correct way about twice as often as a random-span edit, and twice as far, at a modest similarity cost.

Figure 20: Auxiliary random-span experiment (100 seeds per configuration): confidence-guided versus uniform mask fractions. Saliency-guided small-mask results are reported separately in Table 20 .

Figure 21: Auxiliary random-span experiment (100 seeds): success versus seed similarity for confidence-guided and uniform mask fractions. The saliency-guided results are reported separately in Table 20 .

Figure 22: Worked directional edits (QED up, SA down, logP up/down). Red marks the atoms changed relative to the seed; each is a local, chemically interpretable single-group modification that achieves the requested shift at high similarity.

Figure 23: The model localises its own edit. Top: contrastive token saliency $s_{i}$ for a “raise logP” target; the model flags both protonated amines ( $[\mathrm{NH}^{+}]$ ) highest. Bottom: masking the flagged window and refilling under the target removes one amine, raising logP $-0.83\!\to\!2.07$ at Tanimoto $0.69$ .

ID	$\Delta G$	$\Delta G-\Delta G_{\rm ref}$	BSA	Polar contact	RMSD	Ref. sim.	QED	Raw SA
11	-26.73	2.99	0.992	0.95	8.1	0.143	0.648	3.89
2	-24.13	5.59	1.04	0.96	11.18	0.143	0.754	2.67
10	-24.12	5.6	0.987	0.89	13.48	0.145	0.763	3.24
22	-22.52	7.2	0.97	0.69	9.02	0.15	0.748	3.19
25	-22.52	7.2	0.996	0.91	—	0.129	0.922	2.92
29	-21.78	7.94	0.873	0.82	—	0.143	0.826	3.46

Appendix

Table 22: CRBN shortlist. Energies are in kcal/mol; RMSD is in Å. BSA is retained fraction, polar contact is frame occupancy, and reference similarity is ECFP Tanimoto. Missing RMSD entries are denoted by dashes.

ID	$\Delta G$	$\Delta G-\Delta G_{\rm ref}$	BSA	Polar contact	RMSD	Ref. sim.	QED	Raw SA
35	-27.01	3.57	0.934	0.99	3.41	0.105	0.558	3.09
13	-26.69	3.89	0.953	0.94	4.26	0.133	0.729	3.17
6	-25.15	5.43	0.889	0.98	3.38	0.102	0.771	2.92
4	-24.3	6.28	0.876	1	4.37	0.124	0.503	3.04
40	-23.89	6.69	0.901	0.99	3.89	0.108	0.745	2.23
23	-22.56	8.02	1.043	0.79	4.44	0.119	0.439	4.12

Appendix

Table 23: VHL shortlist. Energies are in kcal/mol; RMSD is in Å. BSA is retained fraction, polar contact is frame occupancy, and reference similarity is ECFP Tanimoto. Missing RMSD entries are denoted by dashes.

Figure 24: Six CRBN candidates selected by ascending MM-GBSA energy while retaining distinct scaffold strings. Structures are rendered directly from the supplied SMILES. Each label gives the original candidate ID, MM-GBSA energy (kcal/mol), and ECFP similarity to EF2. All 25 shortlisted candidates are listed in Table 22 and the accompanying CSV.

Figure 25: Six VHL candidates selected by ascending MM-GBSA energy with distinct scaffold strings, rendered from the supplied SMILES. Labels give the original ID, MM-GBSA energy (kcal/mol), and ECFP similarity to EXE. All 21 shortlisted candidates are listed in Table 23 and the accompanying CSV.

Explore similar work

ToolMol: Evolutionary Agentic Framework for Multi-objective Drug Discovery
May 12, 2026Andrew Y. Zhou, Sharvaree Vadgama, Sumanth Varambally +3Drug Design Drug Discovery
Department of Computer Science and Engineering, UC San Diego, La Jolla, California, United States · Department of Computer Science, Stanford University, Stanford, California · Skaggs School of Pharmacy and Pharmaceutical Sciences, UC San Diego, La Jolla, California
Mol-Debate: Multi-Agent Debate Improves Structural Reasoning in Molecular Design
Apr 22, 2026Wengyu Zhang, Xiao-Yong Wei, Qing LiMolecular Design Drug Discovery
♠The Hong Kong Polytechnic University · ♡Sichuan University
FORGE: Fragment-Oriented Ranking and Generation for Context-Aware Molecular Optimization
May 11, 2026Qingchuan Zhang, He Cao, Hao Li +6Molecular Optimization Precise Editing
University of Science and Technology of China · International Digital Economy Academy · Peking University +5

May 12, 2026, cs.LGJ/K move · Enter open · S save

ToolMol: Evolutionary Agentic Framework for Multi-objective Drug Discovery

Andrew Y. Zhou, Sharvaree Vadgama, Sumanth Varambally, Peter Eckmann+2

Department of Computer Science and Engineering, UC San Diego, La Jolla, California, United States · Department of Computer Science, Stanford University, Stanford, California · Skaggs School of Pharmacy and Pharmaceutical Sciences, UC San Diego, La Jolla, California

Drug Design Drug Discovery Molecular Generation Self-Evolving Agents Discovery

Open paper AskPDF arXiv

Advances in large language models (LLMs) have recently opened new and promising avenues for small-molecule drug discovery. Yet existing LLM-based approaches for molecular generation often suffer from high rates of invalid and low-quality ligand candidates, a result of the syntactic limitations of current models with regard to molecular strings. In this paper, we introduce

\texttt{ToolMol}

, an evolutionary agentic framework for de novo drug design.

\texttt{ToolMol}

combines a multi-objective genetic algorithm with an agentic LLM operator that iteratively updates the ligand population. We build a comprehensive toolbox of RDKit-backed functions that allows our agentic operator to consisently make precise ligand modifications.

\texttt{ToolMol}

achieves state-of-the-art performance on multi-objective property optimization tasks, discovering drug-like and synthesizable ligands that have

>10\%

stronger predicted binding affinity compared to existing methods, evaluated on three protein targets.

\texttt{ToolMol}

ligands additionally achieve state-of-the-art results in gold-standard Absolute Binding Free Energy scores, gaining over existing methods by over

35\%

. By studying chain-of-thought reasoning traces, we observe that tool-calling enables the model to more faithfully execute its planned modifications, efficiently exploiting the strong chemical prior knowledge in LLMs.

Abstract

Explore similar work

ToolMol: Evolutionary Agentic Framework for Multi-objective Drug Discovery

Mol-Debate: Multi-Agent Debate Improves Structural Reasoning in Molecular Design

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