🧬 biology

ToolMol: Evolutionary Agentic Framework for Multi-objective Drug Discovery

ToolMol is an evolutionary agentic framework that integrates a multi-objective genetic algorithm with an LLM-driven operator and a comprehensive RDKit-based toolbox to generate high-quality, synthesizable drug candidates with significantly improved binding affinity and free energy scores compared to existing methods.

Original authors: Andrew Y. Zhou, Sharvaree Vadgama, Sumanth Varambally, Peter Eckmann, Michael K. Gilson, Rose Yu

Published 2026-05-14

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Andrew Y. Zhou, Sharvaree Vadgama, Sumanth Varambally, Peter Eckmann, Michael K. Gilson, Rose Yu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are trying to invent a new key that fits perfectly into a very specific, complex lock (a protein in the human body). This is the heart of drug discovery. Usually, this process is like throwing thousands of random keys at the lock, hoping one works, then spending years testing the few that look promising. It's slow, expensive, and full of trial and error.

Recently, scientists started using "Large Language Models" (LLMs)—the same kind of AI that writes essays or code—to help design these keys. The idea was to ask the AI to "write" a new key. However, the paper explains that these AIs are like brilliant but clumsy writers: they often produce "keys" that look like words on a page but are actually nonsense when you try to build them. They make syntax errors, creating molecules that can't exist in the real world.

Enter ToolMol: The "Architect and Builder" Team

The authors introduce ToolMol, a new framework that fixes this problem by changing how the AI works. Instead of letting the AI try to write the entire chemical formula (the "string of text") from scratch, ToolMol turns the AI into a project manager who gives instructions to a robot builder.

Here is how it works, using a simple analogy:

The Evolutionary Game (The "Survival of the Fittest" Key Shop):
Imagine a shop that makes keys. Every day, the shop takes its best keys, breaks them in half, and swaps the pieces to make new "offspring" keys. This is called a Genetic Algorithm. The goal is to keep the best parts of the old keys and mix them to find something even better.
The Problem with the Old AI:
In previous attempts (like a system called MOLLEO), the AI was asked to write the new key's design instructions directly. It would say, "Combine part A and part B," but often it would mess up the grammar, resulting in a design that a builder couldn't follow. About 30% of the time, the AI produced "invalid" keys that had to be thrown away.
The ToolMol Solution (The Toolbox):
ToolMol changes the rules. The AI is no longer allowed to write the chemical formula. Instead, it is given a digital toolbox with seven specific, reliable tools (like "Add a Methyl Group," "Swap a Chlorine for Fluorine," or "Cut and Paste this section").
- The AI looks at the current keys and the lock it needs to fit.
- It thinks through its plan: "I need to cut this key here and attach that piece there."
- It then calls the tool to do the work.
- Because the tools are built by a strict chemical engine (RDKit), they always produce a valid, buildable key. The AI just has to say what to do, not how to write the code.

Why This Matters: The Results

The paper tested ToolMol against other methods on three different protein "locks." Here is what they found:

Better Keys: ToolMol found keys that fit the locks significantly better (over 10% stronger binding) than previous methods.
Realistic Keys: The keys it designed were not only strong but also "drug-like" (meaning they look like real medicines) and "synthesizable" (meaning a human chemist could actually build them in a lab).
The Gold Standard Test: The researchers took the best keys ToolMol made and ran them through the most expensive, high-precision computer simulation available (called Absolute Binding Free Energy). Even though ToolMol wasn't explicitly told to optimize for this specific test, its keys scored 35% better than the current state-of-the-art methods.

The Secret Sauce: Thinking Before Acting

The authors looked closely at how the AI "thought" (its reasoning trace). They found that when the AI used the toolbox, its actions matched its thoughts perfectly.

Old Way (MOLLEO): The AI would say, "I will add a fluorine atom here," but the resulting molecule would accidentally have a chlorine atom instead. The AI was "hallucinating" the result.
ToolMol Way: The AI says, "I will add a fluorine atom," and the tool actually adds a fluorine atom. The plan and the result are identical.

In Summary

ToolMol is a new way to use AI for drug discovery. Instead of asking the AI to be a clumsy writer of chemical formulas, it asks the AI to be a smart supervisor that uses a set of reliable, pre-programmed tools to build molecules. This simple change prevents errors, creates better drug candidates, and allows the AI's vast knowledge of chemistry to be used effectively without the risk of generating nonsense.

Technical Summary: ToolMol

Problem Statement

Small-molecule drug discovery is a resource-intensive process requiring compounds to simultaneously satisfy multiple crucial properties: high binding affinity, drug-likeness (QED), and synthetic accessibility (SA). While Large Language Models (LLMs) have shown promise in generating drug-like molecules by leveraging their pretraining on scientific text, existing LLM-based approaches suffer from a critical limitation: a high rate of syntactically invalid molecular strings (SMILES). Even strong reasoning models fail to generate valid structures in over 30% of attempts, often forcing frameworks to fall back on weaker, non-LLM operators. This syntactic failure hinders the efficient execution of property-based optimization, as the LLM's inherent chemical knowledge cannot be reliably translated into valid molecular modifications.

Methodology

The authors introduce ToolMol, an evolutionary agentic framework designed to overcome the syntactic limitations of direct LLM generation while leveraging their chemical reasoning capabilities. ToolMol integrates a multi-objective genetic algorithm (MOGA) with an agentic LLM operator that utilizes a deterministic, RDKit-backed toolbox.

Core Components

Multi-Objective Genetic Algorithm (MOGA):
- Initialization: The process begins with a population of 60 molecules sampled from the ZINC 250K dataset.
- Selection: Parent molecules are sampled with probability proportional to their fitness, defined by a weighted sum of scaled objectives (Binding Affinity, QED, SA).
- Optimization Goal: The algorithm seeks the non-dominated Pareto frontier ( $M^*$ ) within the chemical subspace, avoiding arbitrary weight selection by strictly defining dominance ( $m' \succ m$ if $m'$ is better in all objectives).
- Termination: The process runs until a fixed "oracle budget" (1,000 binding affinity evaluations) is exhausted.
Agentic LLM Operator (AGENTGEN):
- Tool-Calling Mechanism: Instead of asking the LLM to output raw SMILES strings, ToolMol provides the LLM with a set of 7 deterministic functions backed by RDKit. These include add atom, replace atom, add functional group, add substructure, replace substructure, remove substructure, and crossover molecules.
- Contextual Input: Before each tool call, the LLM receives detailed structural information (atom indices, valences, centrality, neighboring atoms) and property metrics (QED, SA, LogP) for the input molecules.
- Iterative Process: The agent performs a sequence of tool calls (up to a maximum of 10 steps, typically fewer) to modify molecules. It is encouraged to perform a crossover first, followed by mutations.
- Validity Guarantee: Because modifications occur in the deterministic graph space of RDKit Mol objects rather than text generation, the resulting molecules are guaranteed to have valid SMILES encodings, provided the tool calls themselves are syntactically correct.
Evaluation Pipeline:
- Affinity Prediction: Binding affinity ( $\Delta G$ ) is estimated using Boltz-2, a biomolecular foundation model chosen for its high correlation with gold-standard Absolute Binding Free Energy (ABFE) calculations.
- Metrics: Performance is measured by Binding Affinity (BA), Filtered Affinity (FA, considering only molecules with QED > 0.5 and SA < 3.0), and Hypervolume (HV) of the Pareto frontier.
- Validation: A subset of top molecules undergoes rigorous ABFE calculations using molecular dynamics (OpenMM) to validate the predicted affinity.

Key Contributions

Framework Design: The presentation of ToolMol, the first work to utilize LLM tool-calling within an evolutionary framework for multi-objective drug discovery. It successfully abstracts the molecular modification process, allowing the LLM to act as a precise structural editor rather than a text generator.
Syntactic Robustness: By shifting from direct string generation to tool invocation, ToolMol eliminates the high rate of invalid SMILES strings that plague previous LLM-based methods like MOLLEO.
Reasoning Alignment: The framework ensures that the LLM's "chain-of-thought" reasoning traces faithfully correspond to the actual structural modifications made to the molecule, a significant improvement over methods where the LLM's plan often diverges from the result.

Results

ToolMol was evaluated on three protein targets: c-MET, BRD4, and ACAA1 (a target with limited pretraining data for LLMs).

Multi-Objective Performance: ToolMol achieved state-of-the-art (SOTA) results across all three targets. It outperformed baselines including Pocket2Mol, TAGMol, PAFlow, Graph-GA, and MOLLEO.
- Binding Affinity: ToolMol ligands demonstrated >10% stronger predicted binding affinity compared to existing methods.
- Filtered Affinity & Hypervolume: ToolMol consistently achieved the best scores in Filtered Affinity (FA) and Hypervolume (HV), indicating a superior ability to generate molecules that are both high-affinity and drug-like/synthesizable.
Gold-Standard Validation: In ABFE evaluations (the gold standard for binding affinity), ToolMol ligands outperformed the SOTA method MF-LAL by over 35% in ABFE scores, despite ToolMol not being explicitly optimized for ABFE during the search process.
Ablation Studies:
- Replacing MOLLEO's direct generation with ToolMol's toolbox significantly improved results, confirming the value of the tool-calling approach.
- Simply forcing MOLLEO to retry invalid generations did not improve performance, suggesting that the quality of the modification logic (enabled by tools) is more critical than just the validity of the output.
- ToolMol's reasoning traces showed significantly higher concordance with actual molecular changes compared to MOLLEO (p = 0.02).

Significance and Claims

The paper claims that ToolMol demonstrates the practical utility of LLMs for de novo drug design by effectively harnessing their inherent chemical knowledge through structured tool use. The authors argue that the primary advantage of LLMs lies not in generating strings, but in their ability to reason about chemical transformations and heuristics (e.g., hinge-binding motifs, metabolic stability). By constraining the LLM to a valid chemical space via tools, ToolMol bridges the gap between the LLM's reasoning capabilities and the rigorous requirements of molecular generation.

The work suggests that this agentic approach offers a distinct advantage over recent generative models (diffusion, flow-based) which often struggle to balance binding affinity with drug-likeness and synthesizability. The authors posit that ToolMol's ability to generate molecules that align with the reasoning of a medicinal chemist—evidenced by the high ABFE scores—marks a significant step toward efficient, interpretable, and high-quality computational drug discovery.

Limitations Noted: The authors acknowledge that the accuracy of the Boltz-2 oracle may degrade on novel, out-of-distribution scaffolds, though they provide evidence of its superior correlation with ABFE compared to AutoDock Vina. They also note that the current implementation uses a simplified genetic algorithm setup, leaving room for future exploration of more complex evolutionary strategies.