ToolMol: Evolutionary Agentic Framework for Multi-objective Drug Discovery

Abstract: Advances in 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.

• The paper introduces ToolMol, which combines a multi-objective genetic algorithm with an LLM agent restricted to seven RDKit-backed molecular-editing tools, eliminating nearly all invalid SMILES generations.
• ToolMol achieved the best average rank of 1.56 across nine metrics, leading filtered affinity and hypervolume on c-MET, BRD4, and ACAA1 while optimizing binding, QED, and synthesizability under a 1,000-evaluation budget.
• The framework’s top candidates surpassed MF-LAL in absolute binding free energy by more than 19% on c-MET and 35% on BRD4, although results remain limited by reliance on predicted affinity and limited out-of-distribution validation.

Motivation and problem statement

De novo small-molecule design is inherently multi-objective: a viable candidate must bind its target strongly while remaining drug-like (QED) and synthesizable (SA). Existing LLM-based approaches to this problem, most notably MOLLEO, embed an LLM directly into a genetic algorithm as the crossover/mutation operator, but suffer from a persistent failure mode: LLMs frequently emit syntactically invalid SMILES strings. The authors report that direct molecular string generation fails in more than 30% of attempts on average, even on strong reasoning models such as GPT-OSS-120B. In MOLLEO, each failure triggers a fallback to weaker deterministic Graph-GA operators, diluting the benefit of LLM chemical knowledge. ToolMol addresses this by removing the LLM from direct string manipulation entirely.

Method

ToolMol couples a multi-objective genetic algorithm with an agentic LLM operator ("AgentGen") that modifies molecules exclusively through seven deterministic, RDKit-backed tools: add_atom, replace_atom, add_functional_group, add_substructure, replace_substructure, remove_substructure, and crossover_molecules. Because all edits occur in RDKit's molecular graph space, every returned molecule is guaranteed to be valid; syntactic failure is essentially eliminated except in the rare case that the LLM fails to call any tool correctly within ten iterations.

The optimization loop seeds 60 molecules from ZINC 250K, samples parents with probability proportional to $k^{\Phi(m)}$ ($k=10$), where $\Phi$ is a sum of objectives scaled to $[0,1]$, and maintains the population as the non-dominated Pareto frontier over three objectives: Boltz-2-predicted binding affinity ($\Delta G$, kcal/mol), QED, and SA. The search terminates at a fixed oracle budget of 1000 Boltz-2 evaluations. The agent receives dense per-atom structural context (indices, substitutable hydrogens, valences, ring membership, betweenness centrality) plus RDKit descriptors, enabling it to specify tool parameters precisely rather than emitting full molecular strings.

Main results

Across three targets — c-MET, BRD4, and ACAA1 — and seven baselines (Pocket2Mol, TAGMol, PAFlow, Graph-GA, two ShinkaEvolve variants, and MOLLEO), ToolMol achieves the best average rank (1.56) over nine metrics. Key findings:

Result	Detail
Filtered affinity (FA)	Best on all three targets (e.g., −10.35 kcal/mol on c-MET vs. −9.62 for MOLLEO)
Hypervolume	Best on all three targets (0.62, 0.60, 0.54)
Single-objective BA	Best or second-best everywhere; generative baselines like TAGMol match raw BA but collapse under QED/SA filtering

The FA metric is the most consequential: structure-based generative baselines (Pocket2Mol, TAGMol, PAFlow) achieve competitive raw affinity but their candidates largely fail the QED > 0.5 / SA < 3.0 filter, indicating that their high-scoring compounds are not practically synthesizable drug-like ligands. ToolMol is the only method that dominates both multi-objective metrics consistently.

Absolute binding free energy validation

Although ABFE feedback never enters the optimization loop, ToolMol's top-15 molecules (ranked by Boltz-2 alone) outperform MF-LAL — a state-of-the-art ABFE-targeted VAE method — by a large margin: −7.96 vs. −6.7 kcal/mol on c-MET and −8.4 vs. −6.2 kcal/mol on BRD4, gains exceeding 19% and 35% respectively. A filtered top-15 set beats MF-LAL on every metric simultaneously (ABFE, QED, SA). This is a strong claim: optimizing a surrogate oracle transfers to gold-standard free-energy calculations without explicit guidance toward them.

Ablations isolating the toolbox

Two ablations separate the contribution of tool-calling from the surrounding GA machinery:

• Toolbox inside MOLLEO's GA: swapping only the modification mechanism yields large affinity improvements over vanilla MOLLEO (e.g., −11.14 vs. −10.15 kcal/mol on c-MET), confirming the toolbox itself drives the gain.
• Retry-failure MOLLEO: forcing the LLM to retry until it emits valid SMILES reduces invalid generations from ~350/1000 to single digits, yet performance degrades across nearly every metric relative to vanilla MOLLEO. This is a notable negative result: validity alone does not explain ToolMol's advantage.

Why tool-calling helps: reasoning–action concordance

Case studies of GPT-OSS-120B reasoning traces reveal that MOLLEO frequently executes modifications inconsistent with its stated plan — e.g., hallucinating piperidine SMILES while actually inserting pyrrolidine. With ToolMol, executed edits match the stated plan exactly. Quantitatively, across 11 input pairs, ToolMol showed trace–modification discrepancies in 2 of 11 cases versus 7 of 11 for MOLLEO ($p = 0.02$, two-sided independent t-test). The authors' interpretation is that closing the gap between planned chemistry and executed chemistry allows the LLM's pretrained medicinal-chemistry priors to be exploited more faithfully.

Oracle justification and limitations

The choice of Boltz-2 as the primary oracle is supported by a correlation study on c-MET: Boltz-2 achieves $r^2 = 0.42$ against ABFE versus $r^2 = 0.09$ for AutoDock, with binder/non-binder ROC-AUC of 0.95 versus 0.84. However, the paper concedes a significant limitation: recent work shows Boltz-2 degrades substantially on novel, out-of-distribution scaffolds and protein targets — precisely the regime de novo generation operates in. The correlation analysis covers only c-MET due to ABFE computational cost (~12 hours per calculation on 8 GPUs). Additionally, the ShinkaEvolve adaptation is admittedly simplistic (no novelty rejection sampling, no LLM ensemble), so its baseline performance likely understates what that framework could achieve. The ACAA1 results are valuable precisely because that target has no BindingDB measurements, but they remain entirely dependent on predicted rather than experimental affinities.

Conclusion

ToolMol demonstrates that constraining an LLM to act through deterministic, chemically validated tools — rather than generating molecular strings directly — resolves the invalid-generation bottleneck of LLM-based evolutionary drug design and yields state-of-the-art multi-objective optimization across three protein targets, with filtered-affinity and hypervolume wins on every target and ABFE scores surpassing a dedicated state-of-the-art method without any ABFE supervision. The central open questions left by the paper are whether these gains survive evaluation against experimental binding data rather than predicted oracles, and how robust the framework remains when the affinity predictor itself is unreliable on out-of-distribution scaffolds.