Visual omics · spatial evidence · atlas retrieval
From routine histology to spatial molecular evidence
HistAgent reads local and contextual H&E morphology and predicts an ordered list of genes for each tissue location. The rankings support spatial analysis, atlas search and tissue-specific questions.
Three ways to use HistAgent
Learn, analyze and retrieve
Follow the runnable tutorials, generate ranked molecular readouts and discuss measured evidence, or retrieve related spatial molecular states from the reference evidence bank.
Work through five applied notebooks
The notebooks cover gene ranking, spatial findings, standard ST analysis, clinical prediction and atlas retrieval.
Open tutorials → 02 / HistAgentGenerate and interpret molecular evidence
Run HistAgent on paired H&E views or discuss the evidence card of a measured reference spot with adapted Qwen3-8B.
Try HistAgent → 03 / Atlas ExplorerSearch the reference evidence bank
Use a natural-language query to retrieve related molecular states measured by spatial transcriptomics.
Open Atlas Explorer →
Method at a glance
A ranked molecular readout from H&E
HistAgent receives a spot-centered H&E crop, a larger contextual crop and tissue metadata, then produces an ordered list of genes for downstream spatial analysis.
- Dual-scale morphologyLocal and contextual views share a GigaPath encoder and retain distinct spatial roles.
- Autoregressive generationThe decoder produces up to 50 non-repeating gene symbols in rank order.
- Downstream analysisRanked genes support cell-state scoring, pathway analysis and spatial neighborhood comparisons.
Tutorial sequence
Five tutorials
Each notebook combines explanation, runnable code and outputs. Pretrained models are provided where training would take too long for a tutorial.
Generate and evaluate ranked molecular readouts
Calculate spot-level gene-ranking recovery and gene-level spatial recovery across five held-out slides.
Hit Rate@50 · mAP@50 · PCC T02Analyze spatial biological findings
Compare predicted and measured expression, localize an RCC TLS-like niche and review cross-study finding recovery.
RCC TLS case study T03Run standard spatial transcriptomic analyses
Run SVG detection, spatial domain identification, deconvolution, differential expression and pathway enrichment.
Five standard ST analyses T04Interpret whole-slide clinical predictions
Inspect tissue regions associated with slide-level predictions and compare patient-level risk groups.
WSI · prognosis T05Search the spatial transcriptomics atlas
Explore measured tissue locations and run representative natural-language and H&E image queries.
Text and image retrieval