Unified Discovery: How ASCEND's Conversational AI Bridges Internal and Public Data for Accelerated Oncology Insights

Over the past decade, BenchSci has been comprehensively collecting, decoding, harmonizing, and structuring insights from tens of millions of biomedical publications and datasets into a machine‑readable map of disease biology—an effort that ultimately led to LENS, BenchSci’s breakthrough system for extracting complete, evidence-based scientific knowledge at scale. That map encodes hundreds of millions of traceable biological associations and underpins ASCEND’s Conversational AI, a question-answering layer specifically designed for scientific research. Unlike general‑purpose tools such as ChatGPT, ASCEND uses an agentic AI architecture designed specifically for scientific reasoning. Grounded in deeply curated biomedical data and proprietary AI tuned to scientific language, its specialized agents work together to interpret complex research questions and return contextually relevant, citable answers directly within the scientist’s workflow.

Scientific R&D generates massive volumes of data, including lab notes, reports, slide decks, inventories, and procurement records. In most organizations, these assets are scattered across systems, each with its own schema and vocabulary. Against this backdrop of fragmentation, each new question forces scientists to spend valuable time reconciling information, increasing the risk of duplicating past work and slowing progress. 

This can be solved by aligning internal records and external publications to a shared set of semantics and provenance, then making them accessible through a unified evidence graph that supports intent-based querying. 

BenchSci applies this in ASCEND, combining and harmonizing internal experimental records with licensed closed-access sources and open data sources, including clinical trials, in a unified Biological Evidence Knowledge Graph (BEKG). By incorporating closed-access publications and proprietary data, results reflect the broader evidence landscape rather than just an open-source subset. Many emerging AI systems for science, such as Google’s Co‑Scientist or agents from FutureHouse, focus on generating new hypotheses from publicly available literature and tools. In this case, ASCEND instead returns traceable summaries of what is already known, rather than proposing new candidate mechanisms. This integration creates a proprietary, therapeutically agnostic map of disease biology that spans all major data types. Because the BEKG is updated regularly across all sources using scalable extraction methods, it remains comprehensive, current, and aligned with emerging biological insights. This unified perspective helps organizations see what has already been explored, where evidence converges or diverges, and where promising opportunities lie. 

Built on this unified knowledge graph, ASCEND’s Conversational AI allows scientists to ask natural-language questions and retrieve findings grounded in both internal and external evidence. Because experimental results are mapped to standardized biological ontologies and linked with literature and clinical data, each answer generates a single, explainable view of the relevant biology and prior work that is grounded in the knowledge graph, enabling scientists to see what has already been done and where it makes sense to explore next.

Unifying Internal Records with External Scientific Data

Integrating internal experimental records with external scientific datasets creates a unified, proprietary knowledge graph that organizes and connects all evidence into a comprehensive view of research activity. ASCEND decodes unstructured internal sources (ELNs, documents, slides, inventories, procurement records, licensing data) and extracts entities such as targets, models, reagents, readouts, and outcomes.

With internal data displayed alongside external content in the same graph, ASCEND can surface in-house results alongside publicly available data sources for a given question. This reduces redundant experiments and maintains project continuity when teams change, as well as supports direct comparison of assay performance and variability between internal studies and published benchmarks.

Rather than relying on predictive inference alone, the system uses a GraphRAG approach, translating natural-language questions into graph traversals along predefined biological relationship patterns, known as meta-paths. This neuro-symbolic method ensures that answers are grounded in explicit, traceable relationships within the evidence-backed knowledge graph, rather than statistical associations alone. 

By grounding biological entities in a structured knowledge base and following defined relationship paths (e.g., compound → target → pathway → phenotype), ASCEND links related evidence in a way that reflects biological logic. This helps break down silos and provides a consistent, interpretable framework for exploring scientific questions.

Because answers are assembled from graph-based evidence paths, scientists can fluidly move from high-level responses down to the original documents, experiments, and relationships, which allows for rigorous review and reduces cognitive overhead in navigating fragmented data sources. In regulated environments, the system’s transparency and repeatability provide a more predictable alternative to generative prompts.

Security and Governance

Internal experimental data can often contain sensitive information.ASCEND is designed to keep source files contained while scientists collaborate on structured outputs and shared reports, rather than passing around raw documents. The system operates on Google Cloud in a SOC 2 Type 2–certified environment, featuring access controls and isolated deployments that adhere to standard security practices. This allows teams to collaborate, align on results, and scale knowledge sharing without compromising security, compliance, or auditability.

With this infrastructure, what does it actually look like to use ASCEND's Conversational AI in context?

Using ASCEND to Resolve a Research Question

Consider a translational team investigating immune checkpoint regulation in cancer. They want to understand how PD-L1 expression is influenced by intracellular signaling pathways and are looking for mechanistic evidence across internal studies and external datasets.

With ASCEND, the scientist can simply ask: “How is PD-L1 expression related to AKT and MAPK1 phosphorylation?”

ASCEND searches across internal and external sources, including full-text publications, preprint articles, gene/protein information datasets, clinical trials, internal records, and more to retrieve only the most relevant evidence. In seconds, the scientist gains a contextualized view of relevant signalling pathways, established mechanisms of regulation, and modifiers of PD-L1 expression. When internal records are available, ASCEND displays them alongside the literature. 

Because retrieval is optimized around the graph, answers typically arrive in around 30 seconds in BenchSci’s internal measurements, rather than the multi‑minute to hour‑scale cycles reported for some hypothesis‑generation agents. Scientists can phrase questions in their own terms without extensive prompt iteration, since the system interprets them through predefined biological relationships in the BEKG, rather than relying on user-crafted guiding prompts to steer a general-purpose model toward a specific research task.

ASCEND’s Conversational AI reflects a different pattern for scientific intelligence: not a general language model repurposed for scientific questions, but a system tuned from the outset for how scientists frame questions and review evidence. With proprietary and public data fused in a single, secure environment under shared ontologies, a single query can surface internal experiments alongside external findings, helping teams trace, verify, and extend their understanding of a pathway or target with full context while reducing redundant work.