Oct 14, 2026 · Latent Space: The AI Engineer Podcast · 5 min read

LeashBio: Scaling AI Drug Discovery via DELs

LeashBio leverages DNA-encoded libraries and rigorous data hygiene to generate high-quality training data for small molecule discovery. The company addresses the 'Clever Hans' problem in ML by prioritizing internal data over noisy public repositories, enabling robust generalization and novel chemical reasoning.

Technology Business Entrepreneurship

The takeaways

Internal Data Beats Public Repositories. Public chemical datasets contain significant sociological bias, causing models to learn chemist habits rather than physics. Generating proprietary, high-quality internal data is essential for robust generalization.
Scale via DNA-Encoded Libraries. DELs enable the screening of millions of molecules against hundreds of proteins simultaneously. This scalability provides the dense, high-signal data required to train effective binding models.
Rigorous Evaluation Prevents Model Cheating. Standard random splits allow models to memorize building blocks. Using cluster-based splits and activity cliffs ensures models learn true biophysical interactions rather than dataset artifacts.
Language Models Enhance Chemical Reasoning. Integrating LLMs with experimental data allows agents to reason about structure-activity relationships in natural language. This approach improves interpretability and reduces reliance on memorization.
Activity Cliffs Drive Model Learning. Intentionally introducing subtle mutations and chemical changes creates 'activity cliffs.' These high-contrast data points force models to identify critical binding determinants rather than surface-level patterns.
Data Quality as Competitive Moat. The primary barrier to entry is not capital but the ability to manage experimental noise and batch effects. A culture of epistemic rigor creates a defensible advantage in AI-driven discovery.

The Data Quality Imperative in AI Drug Discovery

LeashBio has emerged as a critical player in the AI-driven small molecule discovery space by addressing a fundamental flaw in current machine learning approaches: the reliance on noisy, sociologically biased public data. Traditional models trained on repositories like ChEMBL often learn the 'fingerprints' of individual chemists rather than the underlying physics of molecular binding. LeashBio’s strategy pivots on generating proprietary, high-fidelity data using DNA-Encoded Libraries (DELs), which allow for the simultaneous screening of millions of molecules against hundreds of protein targets. This approach provides the dense, high-signal dataset necessary to train models that generalize effectively to novel chemical spaces.

Strategic Differentiation Through Rigor

The company’s competitive advantage lies not just in scale, but in epistemic rigor. By meticulously tracking batch effects, protein misfolding, and experimental artifacts, LeashBio ensures that its training data reflects true biological interactions. This is exemplified by their use of 'activity cliffs'—intentional, subtle mutations in proteins or molecules that cause drastic changes in binding behavior. These high-contrast data points force models to learn causal relationships rather than memorizing surface-level patterns. Furthermore, their evaluation methodology, which includes cluster-based splits to prevent memorization of building blocks, sets a new standard for model validation in the field.

The Shift to Reasoning Agents

LeashBio is also pioneering the integration of Large Language Models (LLMs) into the discovery loop. By representing chemical structures and binding behaviors in natural language, they enable reasoning agents to explore structure-activity relationships (SAR) with greater interpretability. This 'lab-in-a-loop' approach allows AI to propose and test hypotheses rapidly, accelerating the optimization of lead compounds. The company has already demonstrated success in moving from library hits to mouse pharmacokinetics within six months, validating the practical utility of their data-driven engine.

Market Implications

For investors and industry leaders, LeashBio represents a shift from 'model-first' to 'data-first' AI drug discovery. The company’s ability to generate proprietary, high-quality data at scale creates a significant moat against competitors who rely on public datasets. As sequencing costs continue to drop, the scalability of DELs will only increase, further widening the gap between companies that prioritize data hygiene and those that do not. LeashBio’s model offers a blueprint for the next generation of biotech startups: those that treat data quality as their primary asset and leverage AI not just for prediction, but for rigorous, interpretable reasoning.

Key insights

Public chemical datasets are contaminated with sociological biases, leading models to learn chemist-specific habits rather than universal binding physics. This 'Clever Hans' effect severely limits the generalization capability of standard ML models.

Data Quality →

Impact: Companies relying on public data risk building models that fail on novel targets. Proprietary data generation is becoming a critical requirement for competitive AI drug discovery.
DNA-Encoded Libraries (DELs) offer a scalable solution for generating high-density binding data. By screening millions of molecules against hundreds of proteins, DELs provide the volume and diversity needed for robust model training.

Technology →

Impact: DELs enable the creation of proprietary datasets that are orders of magnitude larger than traditional screening methods, accelerating the pace of discovery and model improvement.
Standard random train-test splits allow models to cheat by memorizing building blocks. LeashBio uses cluster-based splits and activity cliffs to ensure models learn true biophysical interactions.

Machine Learning →

Impact: Rigorous evaluation methodologies are essential for validating model performance. Without them, apparent success may be due to memorization rather than genuine understanding.
Integrating Large Language Models with experimental data allows for interpretable reasoning about structure-activity relationships. This approach enhances the ability of AI agents to propose and test chemical hypotheses.

AI Strategy →

Impact: LLM-based reasoning agents can accelerate lead optimization by providing interpretable insights into why certain molecules bind, reducing the trial-and-error nature of traditional drug discovery.
The primary barrier to entry in AI drug discovery is not capital but the ability to manage experimental noise and batch effects. A culture of epistemic rigor creates a defensible competitive moat.

Business Strategy →

Impact: Investors should prioritize companies with strong data hygiene cultures and proprietary data generation capabilities over those with large funding but poor data quality.

Action items

Audit existing ML models for sociological bias by testing their ability to predict the identity of the chemist who created the molecules. If the model can identify the chemist, it is likely learning habits rather than physics.

Impact: Identifying and mitigating sociological bias in training data will improve model generalization and reduce the risk of failure on novel targets.
Implement cluster-based splits in model evaluation to prevent memorization of building blocks. Ensure that test sets contain no shared building blocks with training sets.

Impact: This will provide a more accurate assessment of model performance and ensure that the model is learning true biophysical interactions rather than dataset artifacts.
Integrate Large Language Models into the drug discovery workflow to enable interpretable reasoning about structure-activity relationships. Use natural language to represent chemical structures and binding behaviors.

Impact: This will enhance the interpretability of AI models and allow for more efficient hypothesis generation and testing, accelerating the lead optimization process.
Invest in DNA-Encoded Library (DEL) technology to generate proprietary, high-density binding data. Focus on screening millions of molecules against hundreds of protein targets.

Impact: This will provide the volume and diversity of data needed to train robust ML models and create a competitive moat through proprietary data assets.
Establish a culture of epistemic rigor by meticulously tracking batch effects, protein misfolding, and experimental artifacts. Implement automated filters to identify and mitigate data quality issues.

Impact: This will ensure that the training data reflects true biological interactions and improve the reliability and generalization capability of ML models.

Quotes

“When you train models on the exhaust, you might be in trouble.”

“We really think that it's learning a lot of sociology. It's not learning ground truth.”

“The winner was like the 1,200 just random person who submitted a Jupyter notebook”