Building the Infrastructure for Open, Trustworthy Science
by Solly Sieberts, Ann Novakowski, and Aditi Gopalan
This July, our team is heading to ISMB/ISCB 2026 to share what we are doing to make biomedical research more open, reproducible, and AI ready. Our team will present thirteen talks and posters, spanning the full pipeline of open science including AI benchmarking, data governance, and shared research infrastructure.
Below are highlights of our work. If any of this overlaps with what you're working on, we'd love to meet with you in person!
KEYNOTE from Our President
We're especially proud to share that our president, Luca Foschini, is delivering an invited keynote at this year's DREAM Challenges COSI. Join us at Benchmarking Automated Scientific Enquiry, Sunday at 12:05 in International Ballroom East. As AI systems increasingly take on tasks once reserved for human scientists, from generating hypotheses, to designing experiments, and interpreting results, the question of how to rigorously evaluate "automated science" itself has become one of the most consequential open problems in the field.
HTAN Team to Give a Congressional Briefing
The Human Tumor Atlas Network (HTAN) team will present their work at a congressional briefing on Capitol Hill as part of ISMB’s Bioinformatics in the USA track. The lunchtime event will introduce ISCB and the importance of computational biology to congressional staffers under the theme “Computational Biology: Past, Present, Future.” They will highlight how computational biology has advanced U.S. science, industry, and medicine and why continued investment in the field matters.
ISMB attendees can learn more about HTAN during their poster presentation:
Building the Human Tumor Atlas Network: A Foundational Multi-omic Data Resource for Cancer Research (Poster D-360): This poster showcases the resources available through the HTAN Data Coordinating Center. HTAN maps the cellular and molecular architecture of human tumors in support of precision oncology, and our DCC standardizes and distributes the multimodal data generated across contributing centers so it can have a lasting impact on the field.
How We Benchmark AI in Biology
A recurring theme across our sessions this year, and shared across the field, is: Now that AI agents can attempt the same scientific problems as human researchers, how do we know if they're actually any good?
TALK: Prospective Benchmarking of Generalist AI Agents Across Biological Prediction Challenges | Sunday 14:20 (DREAM COSI): We’ll share results from evaluating generalist AI agents across seven biological prediction challenges, including the CAGI 7 missense variant challenges and the SEA-AD DREAM Challenge. Early findings: agents performed well on tasks that reward mechanistic reasoning (e.g., correctly explaining drug resistance through a specific molecular mechanism) but lagged behind both human teams and specialized deep learning methods on high-dimensional regression problems. We also found real variability in results across repeated runs of the same prompt, a reproducibility concern the field will need to address as agentic participation grows. We’ll walk through what this means for how challenges are designed moving forward, including real-time leaderboards, which work well for human teams, but may actually disadvantage agents.
Unbiased Evaluation of Computational Biology Models in the AI Era (Poster A-012): This poster makes the broader case for why community-driven challenges, in the tradition of DREAM and CASP, remain essential as AI reshapes computational biology, and how these must adapt to the new landscape. As coding agents and specialized "co-scientists" become more capable, avoiding the field's long-standing "self-assessment trap" (where methods are judged by their own developers) matters more than ever. The recently completed SEA-AD DREAM Challenge, which ran the first-ever dedicated agentic track alongside traditional human submissions, gives us an early framework for how that evaluation can work going forward.
BioArena: A Community-Driven Platform for Trusted Benchmarking of LLMs in Biomedicine (Poster B-011): Check out our open-source, community-driven platform at BioArena.io for evaluating LLMs in biomedicine through blinded, pairwise human comparisons across 30+ frontier models. It integrates with Synapse for verified user authentication and uses AI-assisted validation to keep votes relevant and well-categorized. Our goal is to provide a trustworthy resource for judging LLM performance in the biomedical space.
OpenChallenges for Biomedical Benchmarking
Community challenges have quietly powered some of the biggest breakthroughs in computational biomedicine, see AlphaFold. But the challenges themselves have long been scattered across dozens of platforms with no shared metadata standards or central registry.
TALK: OpenChallenges: A FAIR-Aligned, Centralized Platform for Crowdsourcing in Biological Research | Sunday 15:30 (DREAM COSI): OpenChallenges.io is the first FAIR-aligned registry for biomedical benchmarking challenges. Our team indexed more than 800 biomedical challenges into a single, searchable registry organized around a common vocabulary, collectively representing publications yielding over 28,000 citations. By treating community challenges as durable scientific infrastructure rather than one-off competitions, OpenChallenges aims to make this scattered landscape discoverable for researchers, journals, and funders alike, and to help accelerate the next wave of AI-driven breakthroughs in medicine. Also, check out our recently published manuscript in Patterns.
Upcoming Challenges
TALK: Digital Pen Feature Extraction Challenge | Sunday 11:45 (DREAM COSI): Speaking of open challenges, we’ll be teasing the upcoming Digital Pen Feature Extraction DREAM Challenge, prompting the community to develop open-source methods for extracting features from Digital Clock Drawing Tests (dCDTs), a common cognitive screening tool used in both the clinical and research setting. The challenge is slated for launch in Fall of 2026. Pre-register to stay abreast of updates.
Solving Research Fragmentation in Open Science
Multi-institutional biomedical research runs into fragmentation and poses a real bottleneck to open science efforts.
PANEL: Policies and Strategies for Resilient Open Science | Tuesday 15:20 (BOSC): We will discuss what it takes to keep open science infrastructure resilient amid a shifting funding and policy landscape. Our conversation will explore how we move from "open by default" toward active, coordinated advocacy for the digital commons. If you are wrestling with similar questions, we’d love to hear from you.
Governance as a Service: A GA4GH-Compliant Open Platform for Federated Biomedical Data Access (Poster A-045): GaaS is being designed as an open-source framework that decouples data governance from underlying data infrastructure. Built on GA4GH standards (including Data Use Ontology and GA4GH Passports/Visas), GaaS offers a single standardized access request process built around the Five Safes framework, portable credentials, and a matchmaking system that routes requests to the right committee automatically. Pilots are planned across several of our data portals with the goal of preserving institutional governance authority and giving researchers one coherent way to request access.
Powering Open Science for Aging Research
Core to our mission is supporting data infrastructure and resource sharing for the research community. Our aging portfolio allows researchers to understand protective factors which help people live long health lives:
The ELITE Portal: Multi-omic Data Resources for Studying Exceptional Longevity and Healthy Aging (Poster D-144): The ELITE Portal is our NIA-funded open-access ecosystem for studying exceptional longevity and healthy aging. It brings together harmonized multi-omic data from major longitudinal cohorts (e.g., the Long Life Family Study, the New England Centenarian Study, and the Arivale Wellness Study) alongside cross-species reference data and longevity-intervention models, giving researchers a unified place to study the biology of aging well.
Batch-Aware Data Integration in Rare Disease Research
Rare diseases like Neurofibromatosis type 1 (NF1) pose a particular data challenge with small patient cohorts requiring multiple datasets to gain statistical power.
Batch-aware Integrative Analysis of Transcriptomic Data from a Rare Disease (Poster D-198): NF1 predisposes individuals to several tumor types, including plexiform neurofibroma and malignant peripheral nerve sheath tumors, but the mechanisms driving these distinct, sometimes aggressive, disease trajectories aren't fully understood. Our team applied the tensor decomposition technique PARAFAC2 to bulk RNA-seq data from NF1 patients across four sample batches, comparing the resulting embeddings against standard PCA on both corrected and uncorrected data. PARAFAC2 produced well-resolved sample clusters with similar explained variability to same-rank PCA. The approach offers a way to preserve real biological signal in rare disease datasets where aggressive batch correction risks erasing it.
Connecting Neurodegenerative Disease Research Across Data Silos
Despite significant overlap in molecular pathways across diagnoses, researchers still lack the computational tools to query across disease boundaries. Cross-disease investigations currently take months of manual data harmonization.
TALK: Sage Brain: An Open Knowledge Graph Platform for Cross-Disease Neurodegenerative Research | Tuesday 17:40 (BOKR): We're building Sage Brain, an open knowledge graph platform designed to connect 16 biomedical research portals spanning over 3.9 petabytes of genomic, transcriptomic, proteomic, and clinical data. The platform uses a five-stage architecture that harmonizes heterogeneous data sources through a common data model and ontology alignment, then exposes them through an AI layer supporting natural language queries, cross-disease pattern detection, and hypothesis generation guided by a structured "Hypothesis Games" framework. Initial work is focused on shared AD-ALS biology, with ontology harmonization drawing on SysBio FAIRplex mappings across Accelerating Medicines Partnership programs. Come talk to us if you're working on cross-disease neurodegenerative research or knowledge graph infrastructure.
Synthetic Data Generation using Foundation Models
Rounding out our presentations, we will share methodological advancements in synthetic data generation using foundation model-based approaches.
TALK: From Graphical Models to Foundation Models | Thursday 12:40 (CAMDA COSI): Presented as a talk and a poster (Poster D-074), this work tackles a persistent challenge in genomics: data is heavily siloed and difficult to access due to privacy regulations. Synthetic data may make data more accessible to a broader range of researchers. Our team evaluated two state-of-the-art synthetic data generation methods — MIAV (a deep learning approach built on TabPFN) and TabSDS (a lightweight, non-parametric alternative) — both developed by researchers at Sage — in the CAMDA'26 Health Privacy Challenge. Both methods outperformed the baseline on fidelity, a promising sign for privacy-preserving genomic data sharing going forward.
Our engagements this year share what we believe good scientific infrastructure looks like - open, standardized, rigorously evaluated, and built to keep pace with how research is actually changing. If you're at ISMB 2026, we'd love to connect!