Bio-AI Seedling Program

What if you could simulate genome dynamics before you ever touched a pipette? Virtual Genome merges Michigan’s 3D genome architecture model (ARCH3D) and the Hardwired Genome with NVIDIA Cosmos and Geneformer, fusing live-cell imaging, high resolution microscopy, RNA-seq, and Hi-C data into one unified framework. It's a shift from descriptive biology to life engineering — compressing the discovery timeline for cellular reprogramming and drug interventions from years to days, and laying the groundwork for autonomous, AI-guided laboratories.

ATLAS is MXBio's generalist clinical AI: a lightweight, 1-billion-parameter model trained across a secure, 20-petabyte data commons that spans whole-body radiology, pathology, EMR text, lab results, and longitudinal outcomes. Rather than the single-task tools that dominate clinical AI today, ATLAS is built to reason across modalities at once — automating diagnosis, prognosis, and retrieval, and surfacing novel disease biomarkers along the way.

This pilot integrates 20 terabytes of linked imaging, pathology, and clinical records to address a fundamental question: “What happened to patients like this one?” The system is designed to retrieve clinically relevant prior cases in real time, with longitudinal follow-up documenting treatment responses, adverse events, and long-term outcomes. If validated, this approach could support clinical decisions, inform operational planning, and accelerate clinical research.

Whole-slide pathology imaging holds more biological signals than any one clinician can see. This platform trains foundation models on petabyte-scale whole-slide imaging alongside clinical and genomic data, building toward a multimodal disease atlas — trained inside secure, HIPAA-compliant infrastructure.

Intracranial EEG data has always been siloed, clinic by clinic. This pilot aims to fix that at scale — pairing a petabyte-scale, multi-institution seizure database with predictive AI models that guide targeted surgical and stimulation therapy. Built in partnership with Stanford, Penn, Duke, ETH Zürich, and others, it's designed to put centralized computational infrastructure that is large enough to provide robust analysis, yet also within reach of smaller clinical sites that could never build it alone.

Cancer genomics has generated a vast but fragmented record, scattered across institutions, pipelines and reference genomes over two decades. The Canon of Oncology reconciles massive global cancer data within a unified, versioned framework, enabling discovery of hidden associations, rare molecular events and shared mechanisms across cancer types. Continuously revised as methods improve and evidence accumulates, the corpus preserves the provenance of each result, provides a reproducible foundation for discovery and reveals gaps in knowledge that guide future research and investigation.

KHORANA is a generative cancer genomics model trained directly on raw RNA sequence data — no reference genome, no traditional bioinformatics pipeline. “Raw reads in, biology out.” Scaling from a 12.9-billion-parameter proof of concept to a population-scale model trained across 10,000+ tumor samples, KHORANA is built to power non-invasive liquid biopsy diagnostics and, ultimately, to be released as an open-weight model any lab in the world can query.