Skip to main content
Strand AI predicts spatial protein marker channels directly from an H&E whole-slide image using Lattice, our spatial-proteomics imputation model. Each marker below has a model checkpoint trained against paired multiplex ground truth, predominantly from tumor biology. Predictions outside an oncology context are exploratory; see the intro page for what this means for non-cancer tissue.

​
Public panel

The public panel for new general-access predictions is the 8-channel set below: a nuclear counterstain plus seven biological markers spanning the T cell, B cell, macrophage, pan-leukocyte, epithelial, and proliferation compartments. Existing delivered predictions remain readable under their original marker names. Published cohorts have separate release panels; this request catalog does not claim that CD4 is already available on every public sample or change the historical TCGA-COAD Hugging Face release. Additional markers are available to contracted partners under agreement. The expanded panel changes as checkpoints are validated, so we do not publish a fixed count: GET /markers is the authoritative list for your account. Contact support@strandai.com to discuss an expanded panel for your cohort.

​
Discovering markers programmatically

The GET /markers endpoint returns the exact set your account may request, so agents and scripts can discover valid names upfront instead of guessing. The list matches what a prediction will accept: a self-signup account sees the public panel; a contracted account sees its expanded panel.
The MCP server exposes the same list as the list_markers tool.

​
Reading the predictions

Predictions are returned as a multi-channel OME-Zarr aligned to the slide’s pixel grid. Each requested marker becomes a separate channel; the SDK helpers convert it to AnnData (Python) or SpatialExperiment (R) so you can work with it using your existing multiplex tooling. See the Quickstart for an end-to-end example and the Python / R SDK reference for the conversion helpers.

​
Accuracy caveats

  • Predictions are model outputs, not ground truth. Use them as a hypothesis-generation surface, particularly on cohorts that look out-of-distribution to the training data.
  • Per-marker confidence varies. Lineage and structural markers (e.g. PanCK, CD8, CD45) generally calibrate better than functional / activation markers in the expanded catalog (e.g. PDL1).
  • We do not yet publish per-tissue or per-organ accuracy numbers. Those will be added as benchmarks become available.
Reach out at support@strandai.com if you want pre-publication benchmarks on a specific tissue or marker.