AI in drug discovery is undergoing a significant shift—from tools that retrieve information to systems that generate novel scientific ideas. These emerging inference engines and bio-prediction platforms, often referred to as “AI co-scientists,” represent a new frontier for how discovery can happen. In preclinical research and development (R&D), the challenge is no longer the amount of data available but the difficulty in interpreting it in ways that meaningfully advance research. Across the vast and expanding landscape of biomedical information, critical insights often remain hidden.
The ambition for the next-generation inference systems is clear: to generate novel, evidence-grounded, testable hypotheses that scientists are unlikely to propose from manual literature review. Industry initiatives such as Google's recent AI co-scientist work highlight the growing demand for systems that can reason, connect, and hypothesize at scale. We share this vision and have made significant investments to bring it to life. Our partnership with the Mila Institute and the creation of our dedicated Inference Lab reflect our commitment to advancing AI’s role in disease biology and drug discovery.
However, today’s AI inference systems still fall short of meeting the needs of real-world scientific environments. Many focus on idea generation while overlooking the fidelity, completeness, or accessibility of the underlying evidence. Others rely mainly on public datasets, leaving major blind spots where critical knowledge—often found only in closed-access content or proprietary sources—goes unseen. Additionally, most systems stop at generating hypotheses without tackling the harder questions: Are they experimentally feasible? What is the fastest path to validation?
Our approach closes these gaps. By integrating licensed closed-access scientific data, internal biopharma data, and evidence-backed insights into our Biology Evidence Knowledge Graph (BEKG), we provide a comprehensive, high-fidelity map of disease biology. While idea generation is becoming increasingly abundant, the real breakthrough lies in identifying which hypotheses are of the highest quality and determining how to validate them efficiently and confidently.
This work culminated in LEAP, our AI engine built for evidence-backed hypothesis generation and validation. LEAP navigates the full breadth of biological knowledge, uncovering relationships that would otherwise remain undiscovered. By combining the structured rigor of the BEKG with the advanced reasoning in a neuro-symbolic architecture, LEAP automates the cognitive “leap” scientists employ when tackling unknowns. And beyond generating hypotheses, LEAP predicts the most critical experiments to run and their likelihood of success—reducing uncertainty and accelerating research timelines by years.
Today, we’re excited to share the first real-world application of LEAP—a breakthrough that we believe marks a major step forward in advancing IPF research.
What is LEAP?
LEAP is an inference engine built on a predictive knowledge graph, designed to uncover novel biological connections and identify robust experimental paths. Its agentic architecture consists of independent, specialized agents that work together to perform the complex reasoning required to model biology with high precision, enabling the generation of novel predictions.
Biological relationships within LEAP are derived from BenchSci’s extensive corpus through a pipeline of custom machine learning models that distill scientific findings into contextualized, structured, and directional relationships between biological entities. Experimental details are extracted from figure panels, methods sections, and vendor recommendations, enriching the BEKG with evidence-backed experimental planning capabilities. The extracted insights are grounded in the BEKG and further enriched by comprehensive ontology libraries that cover diseases, genes, pathways, and other related entities.
All of this information, drawn from millions of publications and hundreds of datasets, is synthesized into a unified, scalable knowledge graph where every piece of evidence remains fully traceable and reproducible. This structure provides the high-fidelity foundation required to support credible and confident biological predictions.
How does LEAP work?
The ability to infer new knowledge from indirect relationships is a hallmark of human cognition. It allows scientists to form complex mental models of biology even when direct evidence is incomplete—linking pathways and interactions through reasoning. For example, a gene may not be directly linked to a disease, yet it might interact with a protein that regulates a pathway ultimately implicated in that condition. Identifying these multi-step connections is central to modern biological investigation, and it is exactly the type of reasoning LEAP performs at scale.
Figure 1. Architectural schematic of the LEAP engine and agent-driven question workflow
Step 1: Retrieval
The process begins with a user’s natural-language question, which LEAP’s Question Handling Agent interprets to identify intent and biological context (Figure 1a). The system autonomously decomposes the query into a structured, reasoning plan to explore multifaceted biological problems.
To maximize recall, the query undergoes alias normalization, aligning all terms with the underlying BEKG (Figure 1b). The normalized query is then used to retrieve graph evidence, surfacing evidence-backed biological relationships and inferring plausible multi-step connections by “threading” linked entities to bridge gaps in current knowledge (Figure 1c).
Step 2: Hypothesis Generation
Next, the Scaled Hypothesis Generation Agent evaluates the retrieved relationships for biological coherence and synthesizes them into descriptive, multi-step mechanistic hypotheses (Figure 1d). Supporting evidence is subsequently summarized into comprehensive, biologically grounded narratives ready for ranking and analysis (Figure 1e). To help users prioritize, LEAP ranks hypotheses using multiple criteria, including contextual relevance, biological significance, and novelty, guided by a Reference Network (Figure 1f).
Step 3: Experiment Design
Finally, LEAP supports experimental design by proposing studies that test both individual components and the full hypothesis (Figure 1g).
LEAP in Action
To demonstrate the potential of LEAP, we applied this technology to study idiopathic pulmonary fibrosis (IPF). IPF is a devastating and fatal disease, where the median survival is just 3-5 years. It's characterized by progressive and irreversible lung scarring, with multiple pathologies, including inflammation and epithelial injury, contributing to its progression. This interplay of heterogeneous disease pathologies has stalled therapeutic advances, and to date, only 3 drugs have been approved.
LEAP enabled us to holistically understand the mechanisms driving IPF and apply this knowledge to contextually investigate the recent clinical failure of a connective tissue growth factor (CTGF) inhibitor. LEAP revealed potential therapeutic shortcomings associated with CTGF inhibition and generated hypotheses for future IPF therapies, while suggesting that holistic, systems-based approaches may prove more effective.
The full IPF report demonstrates how LEAP can explore hidden mechanistic drivers, highlight overlooked therapeutic opportunities, and refine research strategy with evidence-backed precision.
Conclusion
By combining advanced agentic reasoning with an enriched, evidence-grounded knowledge graph, LEAP fundamentally redefines how scientists uncover and validate biological insights. It cuts through the noise of vast, fragmented datasets to surface clear, testable hypotheses and the experiments most likely to validate them. Connections that were once hidden within biological complexity become immediately visible and actionable. As LEAP incorporates additional powerful datasets, such as omics and proprietary experimental data, it will power even more distant but biologically plausible predictions.
LEAP empowers scientists to move beyond the constraints of documented knowledge, navigate disease biology with unprecedented precision, and accelerate the journey from curiosity to breakthrough discoveries. In doing so, it redefines what is possible in modern R&D—turning complexity into insight, and insight into impact.
