My Chemical Harness: Evolutionary Molecular Design over Synthetic Pathways with Large Language Model Agents

The paper introduces "My Chemical Harness," a route-native evolutionary framework that leverages large language models as high-level strategy controllers to guide the construction of executable synthetic pathways from building blocks, thereby achieving state-of-the-art molecular design performance without hallucinations or the need for model fine-tuning.

The Gist

Imagine you are trying to invent a new, super-effective medicine. In the past, scientists (or AI) would try to dream up a perfect molecule shape first, like sketching a car in a dream. Then, they would try to figure out how to actually build it in a factory. Often, the dream car was impossible to build because the parts didn't exist or the assembly instructions were nonsense.

"My Chemical Harness" is a new way of doing this. Instead of dreaming up the finished car first, this system starts with the assembly instructions and the parts catalog.

Here is how it works, using simple analogies:

1. The Search is for "Recipes," Not Just "Cakes"

Most AI tries to guess the final cake (the molecule) and hopes it tastes good. This system, however, treats every candidate as a recipe.

• The Ingredients: A list of real, purchasable chemicals (like flour, sugar, eggs).
• The Steps: A list of real, proven cooking methods (like "mix," "bake," "fold").
• The Rule: You can only write a recipe if you can actually buy the ingredients and if the steps are physically possible in a kitchen.

If a recipe calls for "magic dust" or a step that burns the kitchen down, the system rejects it immediately. The "search" isn't looking for a shape; it's looking for the best sequence of steps to make a useful product.

2. The AI is the "Chef Manager," Not the "Cook"

This is the most important part of the paper. The Large Language Model (the AI) is not allowed to just write down a random molecule. That would be like asking a chef to invent a new dish without knowing what ingredients are in the pantry.

Instead, the AI acts as a Strategy Manager:

• It looks at the current "recipes" in the database.
• It decides on a plan: "Let's try swapping the sugar for honey," or "Let's try a baking method we haven't used much yet," or "Let's keep the recipes short."
• It tells the computer: "Go try these specific changes."

The AI never actually "cooks" the molecule. It just gives high-level directions.

3. The "Robot Kitchen" Does the Real Work

Once the AI Manager gives a plan, a deterministic robot kitchen (local code) takes over. This robot:

• Checks if the ingredients actually exist.
• Follows the steps exactly to see if the recipe works.
• Builds the molecule.
• Tests if the final product is good (does it bind to the target disease?).
• Throws away any recipe that fails or produces a duplicate.

This separation is crucial. If the AI hallucinates (makes things up), the robot kitchen catches it immediately because the recipe won't work. The AI guides the direction, but the robot ensures the reality.

4. Learning from Mistakes (The "Reflection" Loop)

The system uses a smart loop called "Reflection."

1. Try: The AI suggests a strategy, and the robot tries 1,000 recipes.
2. Review: The robot tells the AI, "Hey, your idea to use 'honey' worked great, but 'baking at 500 degrees' failed every time."
3. Adjust: The AI reads this report, learns from it, and changes its strategy for the next 1,000 recipes.
4. Repeat: This happens over and over, getting smarter with every round.

What Did They Find?

The researchers tested this on a specific enzyme target (sEH) and a set of standard drug-design challenges.

• Better Results: Their system found better molecules than systems that just guessed the shape first, or systems that didn't use the AI's "reflection" ability.
• Easier to Build: The molecules found were not only effective but also much easier to actually synthesize (build in a lab).
• No Training Needed: The AI didn't need to be retrained or taught new chemistry. It just used its existing knowledge to act as a smart manager for the robot kitchen.

The Bottom Line

Think of this system as a team where the AI is the experienced project manager and the code is the precise construction crew. The manager decides where to look and what to try, but the crew ensures that every building block is real and every step is safe. This prevents the AI from dreaming up impossible things and ensures that the final discoveries are actually buildable in the real world.

Technical Summary

Technical Summary: My Chemical Harness

Problem Statement
The central challenge in drug and materials discovery is designing molecules with targeted chemical properties that are also synthetically accessible. While recent advances in generative models (transformers, diffusion models) have shown success in sampling new molecular structures, they often produce candidates that lack feasible synthetic routes. Traditional approaches typically generate molecular structures first and then attempt to map them to synthesizable pathways a posteriori, or they treat synthesizability as an external penalty. This paper argues that molecular design should be formulated directly as a search over executable synthetic pathways, where the candidate object is the route itself, not just the final molecular graph.

Methodology: My Chemical Harness
The authors introduce "My Chemical Harness," a route-native evolutionary framework that integrates Large Language Models (LLMs) as strategy controllers within a deterministic chemistry pipeline. The system operates as a genetic algorithm where the population consists of executable synthetic routes rather than isolated molecules.

• Search Space Representation: The search space is defined by a library of purchasable building blocks ($B$), a set of reaction templates ($R$) represented as SMARTS patterns, and a compatibility matrix ($M$). A candidate is a route $r = (s_1, \dots, s_L)$, a sequence of instantiated reaction steps. The "genotype" is the route, and the "phenotype" is the final product molecule obtained by executing the route using deterministic tools (e.g., RDKit).
• LLM Role as Strategy Controller: The LLM is explicitly restricted from generating molecules or reaction steps directly to prevent hallucinations. Instead, it acts as a high-level planner that selects preferences for the search, such as route length, move types (e.g., substitution, extension), reaction families, and exploration pressure.
• Deterministic Execution: Local code handles all heavy lifting: sampling compatible reactants, validating reaction slots, executing RDKit transformations, removing duplicates, and scoring products using task-specific oracles.
• Reflective Optimization Loop: The framework employs a "reflective" iteration process:
  1. Context Assembly: The system gathers island memory (history of successful/failed routes).
  2. Planning: The LLM proposes a search strategy based on the task and context.
  3. First Pass: Local code executes the strategy, generating and scoring candidates.
  4. Reflection: The LLM reviews the outcomes (successes, failures, motifs) and revises its strategy for the remaining budget.
  5. Second Pass & Selection: The revised strategy is executed, and the best candidates are committed to the population.
  6. Learning: The LLM generates a report to guide future iterations.

Key Contributions

1. Route-Native Evolutionary Search: The paper proposes a paradigm shift where the genetic algorithm operates on executable synthesis programs (routes), ensuring that synthesizability is structural to the search space rather than a post-hoc constraint.
2. Constrained LLM Integration: It demonstrates a "harness" architecture where LLMs guide exploration through high-level strategy without direct access to chemical generation, thereby eliminating the risk of hallucinated products or unsupported reaction steps.
3. Reflective Agent Mechanism: The introduction of a mid-iteration reflection step allows the LLM to adapt its search policy based on immediate feedback from the deterministic chemistry engine, improving the allocation of search effort.

Results
The framework was evaluated on two benchmarks:

• Soluble Epoxide Hydrolase (sEH) Proxy: The reflective LLM agent achieved state-of-the-art performance compared to deterministic controllers and non-reflective LLM baselines. It attained a top-1000 mean sEH score of $0.984 \pm 0.005$, the lowest Synthetic Accessibility (SA) score of $1.996 \pm 0.027$, and the highest AiZynthFinder success rate of $0.994 \pm 0.004$. The reflective agent successfully enriched chemically interpretable motifs (e.g., urea and amide groups) consistent with known sEH inhibitor chemistry.
• Therapeutics Data Commons (TDC) Oracles: The method was applied to 13 diverse molecular design oracles. The reflective LLM approach outperformed baseline synthesis-constrained methods (ReaSyn, SynthesisNet, SynNet) on several targets, including amlodipine_mpo (0.6436 vs. 0.620) and sitagliptin_mpo (0.4522 vs. 0.314), demonstrating the generalizability of the strategy control across different optimization objectives.

Significance and Claims
The paper claims that constrained LLM agents can play a significant role in molecular discovery without requiring training, fine-tuning, or dedicated generative models. The primary value of the LLM is identified as adaptive prioritization over a constrained synthetic action space. By separating the LLM's strategic role from the deterministic execution of chemistry, the system achieves a balance where the LLM guides the search toward promising regions of chemical space while local tools enforce chemical reality and validity. The authors conclude that this approach facilitates the rational design of synthesizable drugs and molecularly imprinted polymers, offering a reliable, auditable, and hallucination-free pathway for AI-assisted scientific discovery.