Key Takeaways

  • Public chemistry repositories like ChEMBL train machine learning models on human publication habits rather than physical binding mechanics.
  • A human medicinal chemist oversees 20 to 50 structure-activity relationship (SAR) campaigns across an entire career; Leash Bio completed over 10,000 of them in a single weekend.
  • Open-source language models such as Qwen and GLM can reason over molecular differences and binding behavior in plain English instead of relying solely on black-box neural networks.
  • Pre-screening millions of molecules across thousands of protein constructs with DNA-encoded libraries creates a software-speed environment where agents run 100 to 200 sequential tool calls without wet-lab delays.
  • Teams execute this automated cycle through Leash Bio's LLM Closed-Loop Reasoning Sandbox.

The Leash Bio's LLM Closed-Loop Reasoning Sandbox

  • Step 1: Natural Language Representation: Teach an open-source LLM (such as Qwen or GLM) to describe small molecule structures, molecular diffs, and binding behaviors in natural English.
  • Step 2: Seed Hypothesis Generation: Provide the agent with 10 binding molecules and 10 structurally close non-binders (hard negatives), prompting it to articulate an explicit English hypothesis explaining which chemical motifs drive binding.
  • Step 3: Experimental Selection Budget: Grant the agent an experimental testing budget (e.g., 5 to 10 molecules from a massive catalog of millions of pre-measured compounds) to test its hypothesis.
  • Step 4: Instantaneous Wet-Lab Rollout Query: The agent queries the pre-measured empirical DEL database like an instant coding sandbox, receiving ground-truth binding results without wet-lab latency.
  • Step 5: Belief Updating and SAR Convergence: The agent analyzes the returned data, updates its hypothesis across 100 to 200 sequential tool calls, and isolates the most potent binder within its molecule budget.

When This Works (and When It Doesn't)

This system excels at rapid structure-activity relationship optimization campaigns using pre-screened chemical databases, bypassing physical synthesis delays and preventing model memorization. Quigley points out that the agent operates within “a reinforcement rollout of a large scale that is growing all the time that is full of these activity cliffs and full of different protein variations.” The model receives ground-truth feedback in milliseconds, testing whether tiny atomic substitutions retain or destroy target binding.

The method breaks down the moment you step outside the boundaries of your pre-measured chemical library. If the model wants to explore a chemical scaffold that was never synthesized or sequenced in your DEL runs, the feedback loop halts. You are forced back into slow, physical bench chemistry. It also fails when natural English fails to convey subtle three-dimensional stereochemistry or solvent interactions that simple textual diffs cannot capture.

What to Do With This

If you are a founder running early-stage drug discovery, stop treating wet-lab synthesis as your primary iteration bottleneck. Set up a pilot search for your lead program this week:

Take 10 known binders and 10 structurally similar non-binders from your target assay. Feed them into an open-source model like Qwen. As Quigley describes: “We taught them medicinal chemistry in ways where you say here is a structure describe it in natural English give me the diff between two molecules in natural English and then we got it to the point of where we say okay here are some molecules that are binders here are some that are not binders can you tell us why you think the binders are binding.”

Next, give the model an experimental budget of 8 compound selections from your existing proprietary screen. Let the agent query the database through automated tool calls. “The model will go into the data that we have already collected. It will pull out what it thinks are the most informative choices. It will get back the answer instantly like it's a coding sandbox and then it will update its hypothesis about the binding.” Check the updated hypotheses after 150 tool calls. If the agent isolates the active motif, you have validated your target mechanism before ordering a single physical tube.