stGPT-Xenium L3 CNN v0.1

stGPT learns the morpho-molecular tissue representation; spatho turns it into auditable spatial pathology evidence.

Release Status

• Public research artifact.
• Not a diagnostic model and not for clinical decision support.
• First checkpoint in the stGPT-Xenium series, intended for reproducibility, interface testing, and transparent model-development tracking.
• This is not a universal foundation model yet; stronger multi-case and slide-holdout models will be released as later versions.

What This Checkpoint Is

• Model family: stGPT-Xenium
• Release: L3-CNN-v0.1
• Backend: lightweight CNN contour H&E encoder
• Training unit: XeniumSlide contour/region-level image-gene-spatial learning
• Representative checkpoint case: Xenium_Prime_Human_Prostate_FFPE_outs
• Selected from a 39-case L3 morphology-ready A100 pilot where all cases completed successfully

A100 Pilot Summary

• Uploaded/trained L3 cases: 39/39
• Completed metrics: 39/39
• Failed cases: 0
• Checkpoint records: 273
• Error rows: 0
• Mean final validation loss across 39 cases: ~0.180
• Mean final validation gene loss: ~0.00162
• Mean final validation neighbor loss: ~0.0317

• last_val_loss: 0.1554184090346098
• last_val_gene_loss: 0.0013655137008754
• last_val_neighbor_loss: 0.0067438320256769
• last_val_image_gene_loss: 3.0473387241363525

Files

• checkpoint_best.pt: representative best checkpoint.
• config.yaml: stGPT config used for this case on A100.
• metrics.json: per-step training metrics for this checkpoint.
• stgpt_model_manifest.json: provenance and checksums.
• training_telemetry_summary.json: pilot-level telemetry summary.
• case_metric_summary_all_cases.csv: aggregate training metrics for all 39 pilot cases.
• checkpoint_manifest_all_cases.csv: checkpoint provenance for all pilot cases.

Evidence and Data Policy

• no xenium_slide.zarr
• no H&E patch PNGs
• no OME-TIF or raw image files
• no raw expression matrices

Intended Use

• testing the stGPT model/package interface
• reproducing the v0.1 checkpoint loading path
• comparing against future stGPT-Xenium releases
• research-only spatial transcriptomics and computational pathology method development

• diagnosis
• treatment decisions
• clinical reporting
• claims of universal tissue or organ generalization

Limitations

• Single representative checkpoint, not a multi-case global model.
• CNN image encoder baseline; Virchow-precomputed and ablation evidence are tracked separately and will inform later releases.
• Validation is pilot-level and per-case; slide-holdout and cross-case generalization are future release targets.
• Public 10x source data must be accessed from original sources, not from this model repository.

Version Series

• stGPT-Xenium-L3-CNN-v0.1: first public pilot checkpoint, this repo.
• stGPT-Xenium-L3-Virchow-v0.2: planned frozen pathology-encoder checkpoint.
• stGPT-Xenium-Foundation-v1.0: future multi-case/slide-holdout foundation release.