Cell Type Knowledge in the Age of Foundation Models and Agentic AI

About the workshop

We are at a pivotal moment in cell biology. Massive single-cell and spatial omics datasets, analyzed with increasingly powerful AI methods, provide data-driven, unbiased definitions of cell phenotypes at unprecedented scale. To date, relating these cell phenotypes to the vast store of prior knowledge about cell structure, function and dysfunction in health and disease has required painstaking manual effort.

Advances in agentic AI are beginning to change this — dissolving the boundaries between literature, curated knowledge and data. Deep research tools can query and interpret the literature at scale, extracting information about cell types and their phenotypes to generate and extend ontologies and knowledge bases. They can gather and assess evidence to map omics-based definitions of cell types to classical types, find relevant datasets, and run bioinformatic analyses to support or refute proposed mappings. Importantly, they can do this at scales no human team can match.

This work can feed back into the construction and use of machine learning models. Structured knowledge in ontologies and knowledge bases provides a scaffold for interpreting the outputs of ML models and a training corpus for building more sophisticated, more interpretable models with richer predictive capacity.

Agentic AI is also transforming the way that humans interact with knowledge bases and ontologies — supporting queries with natural language, making this accumulated knowledge accessible and actionable in ways that were previously impossible.

This workshop brings together researchers from the communities who need to collaborate to realize this vision: biologists and bioinformaticians working with single-cell and spatial omics data; computational biologists and machine learning researchers developing foundation models and interpretability tools; ontology and knowledge base developers; literature mining and NLP researchers; and agentic AI researchers and practitioners.

Program draft agenda — timings may change

Attending virtually? Join via Zoom: sanger.zoom.us/j/96259448879

Panel discussion Key questions

• What role does curated, structured knowledge about cell types have as input to ML models, and in making ML models more interpretable and useful?
• How do we reconcile data/ML-driven cell type clusters with curated classical cell types?
• Keeping humans in the loop: how can we use agentic workflows to construct cell type knowledge bases and ontologies that support rapid human review of assertions, evidence, supporting literature and data?
• Can agentic AI scale cell ontology and knowledge-graph curation without sacrificing rigor?
• How can ontologies and knowledge graphs improve the performance of LLM-based natural language processing and knowledge extraction?

This is a draft agenda and timings may change. Each talk is 30 minutes including questions.

Confirmed speakers

• David Osumi-Sutherland — Wellcome Sanger Institute

  Agentic Workflows for Cell Ontology Curation: Evidence at Scale

  Cell Ontology as a use case — grounding cell-type knowledge for interpretable ML and agentic workflows.

• Hanchen Wang — Stanford University & Genentech (Leskovec & Regev labs)

  Recipes for Building Agents for Biomedical Discovery

  This talk will share our experience building AI agents for biomedical discovery, focusing on integrating language models with biological knowledge and domain-specific tools to support research workflows. I will discuss what worked, what failed, what we learned about reliability, evaluation, and human-agent collaboration, and where the field may be heading in the near future.

  Bio. Hanchen Wang is a postdoc at Stanford University and Genentech, working with Jure Leskovec and Aviv Regev. His research develops computational methods to understand biological systems and advance discovery. His leading work has appeared in Nature, Science, Nature Biotechnology, and NeurIPS. He received his CS Ph.D. from Cambridge, UK, working with Joan Lasenby at Trinity College, and his Physics B.S. from Nanjing.

• Harry Caufield — Berkeley Lab (Mungall group)

  Interchangeable Parts: Agentic AI for Data Modeling, Curation, and Validation

  Biological and biomedical knowledge curation is expert-intensive work. The experts it depends on cannot afford to spend their time using tools they don’t trust or must continually double-check. This talk will describe a set of agentic AI workflows and reusable tools developed across multiple projects. As a running example, I will use the DisMech project, a curated knowledge base of disease mechanisms in which every claim ties back to the primary literature. I will discuss how workflows can absorb routine effort while keeping domain experts in a position of oversight. Central to this approach is the treatment of curation as an assembly of interchangeable parts: a LinkML schema serves as a shared, machine-readable specification into which each component tool fits. Agents draft structured entries, validators confirm ontology terms (such as HPO, MONDO, and GO), and a separate step verifies that every quoted snippet genuinely appears in its cited source. I will detail how these steps are orchestrated across the full curation lifecycle, from AI-assisted pull-request review, issue triage, and literature scanning to knowledge-gap detection, compliance checks, and automatically rebuilt documentation. Agents are also employed to research, draft, and self-validate contributions before a human review. The talk will contextualize this work within a broader effort to apply agentic AI to biological and biomedical data curation in ways that deliberately center domain expertise and human review. The value of automation is not derived from replacing curators but from directing their judgment to where it matters most.

  Bio. Harry Caufield is a Staff Data Scientist at Lawrence Berkeley National Laboratory, where he works in the Mungall group on data modeling, knowledge harmonization, and information extraction from biomedical literature. He is a core contributor to multiple data resources and consortia, including the Monarch Initiative, Bridge2AI, and BioPortal. He has led the development of SPIRES and OntoGPT, widely-adopted methods for using large language models to extract structured, ontology-aligned knowledge from scientific text. His work spans the LinkML data modeling language, the Biolink Model, and biomedical knowledge graphs, with a consistent theme of pairing automated curation with domain expertise and human review. He holds a PhD in Integrative Life Sciences from Virginia Commonwealth University and completed a postdoc in biomedical data science at UCLA.

• W. Jim Zheng — McWilliams School of Biomedical Informatics, University of Texas Health Science Center at Houston

  From Ontology Fingerprints to BRAINCELL-AID: Literature-Driven Genome and Cell Annotation

  Biomedical literature contains a vast and rapidly expanding body of knowledge, making it both a valuable resource and a challenge to systematically extract information relevant to gene function and cellular identity. Our early works introduced Ontology Fingerprints, a framework that represents genes as interpretable embeddings derived from Gene Ontology terms enriched in literature. This approach captures gene functions through statistical associations while preserving interpretability, enabling applications such as gene prioritization and pathway discovery through a self-attention-like mechanism. Building on this foundation, we extend literature-driven annotation to the cellular level using modern deep learning and large language models. We developed BRAINCELL-AID, an agentic AI framework that integrates fine-tuned LLMs, literature mining, and retrieval-augmented generation to annotate brain cell marker gene sets and cell types at scale. BRAINCELL-AID provides literature-grounded annotations for more than 20,000 gene sets and over 5,300 brain cell types, demonstrating strong cross-species performance and enabling biologically meaningful hypothesis generation. Together, this work establishes a continuum from knowledge-based embeddings to agentic AI systems, highlighting the role of literature-integrated representations in advancing gene function characterization, pathway discovery, and large-scale brain cell annotation.

• Tom Gillespie — UC San Diego (Martone group)

  The Neuron Phenotype Ontology in the Peripheral Nervous System

  We start with a brief introduction to the Neuron Phenotype Ontology (NPO) knowledge representation and knowledge base. We then cover the extensions to NPO that have been developed to support the SPARC Connectivity Knowledge base of the Autonomic Nervous System (SCKAN) as part of the NIH SPARC initiative. We discuss how we populated SCKAN, including running NLP pipelines across a large portion of the neuroscience literature. We provide an overview of the open source tools that have been developed to interact with SCKAN for query, curation, visualization, and exploration including natural language query interfaces. Finally, we touch on some of the ontology modelling and usability challenges that we have encountered along the way.

  Bio. Tom Gillespie is a postdoc in the FAIR Data Informatics Lab at the University of California, San Diego. He is one of the maintainers of the NIF standard ontology that covers the domain of neuroscience and is the designer and maintainer of the Neuron Phenotype Ontology.

• Richard Scheuermann — National Library of Medicine, NIH

  NLM Cell Ontology Knowledge Graph

• Sourav Sarkar — New York Genome Center (Satija Lab)

  Pan-human Azimuth for organism-scale annotation (virtual)

  In Pan-human Azimuth, we present a first-of-its-kind platform to harmonize human cell annotation under a single hierarchical pan-human ontology. Single-cell atlases have grown rapidly in volume and tissue coverage, but their fragmented annotation schemes have limited our ability to use them as a unified reference. As part of the NIH HuBMAP program, we curated a large corpus of human transcriptomic profiles with uniform, interpretable annotations, prioritizing volume, diversity, and strict quality control, and trained a deep neural network classifier to project new data onto this hierarchical annotation scheme. The scalability of our model allowed us to rapidly annotate tens of millions of cells from the scBaseCamp compendium, and annotating thousands of samples from across the body under a unified framework enabled previously inaccessible cross-tissue analyses. This revealed a striking degree of tissue localization in certain fibroblast substates, and allowed us to derive marker gene sets for a large number of cell types from signals coherent across a broad diversity of samples. We further demonstrate the accuracy and generalizability of our tool through its application to spatial transcriptomics, where we recover the spatial organization of canonical cortical structures in the human kidney from molecular annotations alone, also revealing a continuum of glomerular states whose extremes are independently validated by expert pathology.

  Bio. Sourav is a physicist turned Machine Learning and genomics researcher, working as a post-doc in the Satija Lab at the New York Genome Center. He did his PhD in Theoretical High Energy Physics from Humboldt University in Berlin. His current research interests are in Quantitative and Systems biology and applications of Machine Learning, in particular to single-cell and spatial genomics.

Organisers

• David Osumi-Sutherland Wellcome Sanger Institute
• Richard Scheuermann National Library of Medicine, NIH

Part of ICBO 2026 — Semantic Awareness in the Age of Generative AI — co-located with ISMB 2026 at the Washington Hilton, Washington DC, in the Monroe room (Concourse Level).