Biomni and Sage Partner on AI-Powered Access to Biomedical Data
Sage Bionetworks and Biomni have announced a collaboration that connects researchers to approximately 3 petabytes of curated biomedical datasets through natural language queries. The integration is designed to simplify discovery across fragmented repositories and reduce time spent on manual data searches.
Sage Bionetworks, a nonprofit known for open science governance, operates Synapse, a cloud-based platform hosting datasets in oncology, neuroscience, rare diseases, and digital health. Biomni, an open-source AI system for literature and data discovery, now integrates Synapse’s API to make those datasets searchable in plain language.
The partnership centers on two technical components:
- Physical access to Synapse data (~3 PB), governed and structured for reproducibility.
- Biomni’s query layer, which parses user prompts, translates them into Synapse queries, and retrieves datasets or files by relevance.
Biomni supports multiple query modes for Synapse data: free-text prompts are parsed to identify key entities like disease or modality, with the system selecting the appropriate query type and returning datasets annotated with descriptions and governance status; keyword-based filtering applies AND logic to require multiple terms, narrowing results to datasets where specified metadata co-occur; and file-level retrieval targets specific formats such as VCF, BAM, or FASTQ, enabling direct access to data objects within larger studies for downstream workflows such as variant calling or secondary analysis.
Access controls remain in place: while discovery spans both public and restricted datasets, downloading governed files requires Synapse approval and a personal access token. This maintains participant privacy and compliance while allowing researchers to rapidly locate relevant resources.
In a test query on neurofibromatosis type 1 (NF1), Biomni simultaneously searched across Synapse, GEO, cBioPortal, and recent literature to return datasets spanning patient-reported outcomes, imaging, and genomics. The results included:
- Synapse: A multi-tumor RNA-seq cohort (syn53140231) covering NF1-associated tumors such as malignant peripheral nerve sheath tumors and neurofibromas, linked to a Johns Hopkins biospecimen repository. Data files support gene expression analysis but require governed access approval.
- GEO: Several open datasets, including single-cell transcriptomics of neurofibroma development, bulk RNA-seq of NF1 patient-derived bone cells treated with MEK inhibitors, and mouse models studying environmental influences like maternal obesity on optic pathway gliomas. Together, GEO listed 386 NF1-related studies across tumor biology, treatment response, and molecular mechanisms.
- cBioPortal: Open cancer genomics datasets highlighting NF1 mutations and copy number alterations across multiple cancer types, suitable for comparative genomic analysis.
By combining sources, Biomni produced a structured overview of data ranging from mechanistic single-cell studies to large patient cohorts. Notably, the same NF1 tumor data (syn53140231) could be explored both as raw RNA-seq files in Synapse and as processed results in cBioPortal, underscoring the value of cross-platform discovery for tailoring datasets to different research applications.
Researchers can access Synapse data directly through Biomni in two ways: via the web platform, where natural language queries return governed datasets from Synapse’s open science collection, or through the open-source integration, which allows cloning the Biomni repository and embedding Synapse search into custom workflows with full control over parameters, access, and data integration pipelines.
The collaboration extends Biomni’s earlier integration with the Consensus academic search engine, which added a large, curated literature corpus and AI-based retrieval to its framework.
Additionally, many of the most valuable biomedical datasets remain access-restricted: they can be discovered through Biomni, but downloading requires approval through platforms like Synapse. This governance protects participant privacy but often slows reproducibility and collaboration.
Biomni notes a longer-term goal of extending their platform to help with this process, envisioning an assistant that could guide data access requests, track their status, and prompt follow-ups while respecting existing governance rules. Achieving this will depend on community input to design protocols, build infrastructure, and develop the supporting code needed for more seamless yet secure access.
To stay updated on developments like this, follow our newsletter where we track emerging trends in biomedical AI and data infrastructure.
Topic: AI in Bio
