The challenge

Spatial-omics studies combine whole-slide images, sparse molecular measurements, cell geometries, transcript coordinates, segmentation results, and biological hierarchy across donors, specimens, sections, slides, batches, and technologies. Existing tools can store or query many of these objects, but method developers still build bespoke pipelines to turn them into synchronized, model-ready samples.

That glue is scientifically consequential. Coordinates from different specimens must never be joined accidentally. Gene panels require an explicit vocabulary policy. Raw counts, normalized values, residuals, and imputed values cannot be silently interchanged. Thousands of cells from one donor do not constitute thousands of independent biological replicates.

The work

Spatial ML Data Engine is an AI For Tomorrow collaboration with Brown University. It proposes a biology-native, PyTorch-compatible runtime for streaming larger-than-memory spatial-omics data into models. A query could return aligned image context, cells or bins, transcript points, sparse molecular features, a dynamically constructed local graph, assay metadata, and donor or slide identity—together with enough provenance to reproduce the batch.

The system would borrow mature coordinate indexing, raster windows, point neighborhoods, spatial joins, caching, and distributed sampling from geospatial, point-cloud, graph, and medical-imaging systems. Its contribution is the semantic and execution layer that composes those primitives safely for biology: specimen-scoped coordinates, assay vocabularies, molecular-layer identity, segmentation provenance, multimodal sample contracts, and biologically meaningful sampling policies.

The architecture is not predetermined. The first phase will compare a custom runtime against the strongest practical thin stack of existing tools. If that composition offers comparable performance, usability, and semantic guarantees, the project will narrow to a semantic adapter and sampler layer rather than build an unnecessary storage engine.

Why it matters

Data access can determine which spatial methods are practical. Repeatedly materializing image tiles, graph variants, or training corpora makes scientifically meaningful choices—such as graph radius, image scale, support shape, segmentation view, or donor split—expensive to vary.

A reusable runtime could turn those choices into ordinary experiment configuration, reduce duplicated preparation, improve accelerator utilization, and allow several unrelated model families to share one validated data path. More importantly, it could make biological correctness executable instead of leaving it as metadata every downstream model must interpret on its own.

Success will be measured end to end. The project must show that it reduces total experiment cost or user-code complexity while strengthening semantic validation—not merely that one isolated loader microbenchmark is faster.

Timeline

Completed milestones are highlighted; muted entries show the work still ahead.

  1. July 2026
    Complete

    Product thesis and falsification criteria defined

    The design review established the biology-native sample contract, competitive baselines, performance hypotheses, and explicit go, narrow, or stop conditions.

  2. Phase 0
    Future work

    Run the thin-stack bake-off

    Compare the proposed runtime with a strong composition of SpatialData, specialized matrix loaders, spatial databases, PyTorch Geometric, and ordinary samplers.

  3. Six-month target
    Future work

    Build the minimum viable runtime

    If the bake-off supports the thesis, deliver indexed multimodal queries, dynamic local graphs, biological sampling policies, deterministic execution, and complete provenance.

  4. Public beta
    Future work

    Validate reuse across model families

    Demonstrate image, cell, transcript, graph, and donor-aware workloads across several model consumers, including at least one non-GNN use case.

People involved

Project leads, contributors, and collaborators named in the public project record.