HistoGen: Enabling Tumor Boards with AI Assisted Pathology-Derived Data

Integrated molecular profiling from H&E data.

PEAT-Nucleate BioHack 2026 • TCGA Lung Cohort

The Clinical Problem

Mutation panels are necessary — but not sufficient

Therapies fail across modalities

NSCLC Patients can fail targeted therapy, chemotherapy and immunotherapy – even when the mutation profile looks actionable.

Limited median benefit

Initial targeted responses can be strong, yet resistance emerges. Median PFS in most trials: 6–12 months.

Resistance is not just DNA

Secondary mutations matter, but so do pathway rewiring, lineage plasticity, stress programs and spatial niches.

Averages mask heterogeneity

Readouts report means for 'EGFR+' — concealing the subset that responds vs. the subset that doesn't.

Key question: What differs between responders and non-responders with the same mutation?

HYPOTHESIS

Treatment is slowly improving, but survival is still low

PEAT-Nucleate BioHack 2026

HYPOTHESIS

Why modern therapeutics still fail

Sensitive Bulk

Most visible tumor may respond initially, but response averages hide resistant subpopulations.

Plastic/Stress-Tolerant

EMT, hypoxia, lineage-shift, or drug-tolerant persistent programs dominant survival.

Myeloid / Treg-rich

Spatial pockets of stroma, macrophages, poor drug penetration, necrosis, or pre-existing resistant clones can blunt targeted therapy, chemo, and IO.

HistoGen: Increasing visibility into resistant tissue ecosystems?

PEAT-Nucleate BioHack 2026

THE GAP

Spatial transcriptomics is the gold standard

— but costs $$$ and rarely runs at large clinical scale.

Three 2025–2026 papers changed this.

Phoenix

Helmholtz / TUM

Virtual single-cell spatial transcriptomics from H&E. 1.2B-param flow model on 22.2M cell pairs. TCGA atlas released.

GigaTIME

Microsoft / Providence

Virtual 21-channel mIF from H&E. Trained on 40M cells, precomputed on 10,200 TCGA patients. Open weights.

Haiku

Penn / Enable

Co-embedding of H&E + clinical text. Retrieves real matched mIF tissue from a reference cohort.

Nobody chained them — until now.

Virtual Molecular Tumor Board

H&E slide + clinical note + traditional DNA mutation data → ranked, citation-backed molecular hypothesis report

INPUT

H&E Slide

+ Clinical Note

+ DNA Mutations

Phoenix

Virtual spatial Tx

Cell types & niches

GigaTIME

Virtual 21-channel

mIF protein maps

Haiku

Patient retrieval

Real mIF matched

KB

Claude Agent — Orchestrator

Input H+E + Clinical Data

​

Predicts spatial RNA and protein expression

Clusters similar patient groups

PEAT-Nucleate BioHack 2026

Impact

Enrich future trials

Stratify by TME archetype — not mutation alone — to identify responder subpopulations a priori

Explain trial failures

Post-hoc: why did mutation-positive EGFR/KRAS cohorts underperform? Map their TME states

Use existing biobanks

No new biopsy protocol needed — run on archived H&E slides already in every hospital

Bridge to deep biology

HistoTME screens → spatial transcriptomics validates on the interesting subset

Mutation is the lock. TME is whether the key turns.

Key References

HistoTME

Patkar et al. NPJ Precision Oncology 2024 — HistoTME for NSCLC / ICI response prediction

HistoTMEv2

Preprint 2025 — Pan-cancer extension of HistoTME (29 → extended signature set)

Phoenix

Helmholtz/TUM 2026 — Virtual single-cell spatial Tx from H&E; 1.2B-param flow model; 9,544-patient TCGA atlas released

GigaTIME

Microsoft/Providence Cell 2025 — Virtual 21-channel mIF from H&E; Spearman 0.88; precomputed on 10,200 TCGA patients

Haiku

Penn/Enable 2026 — Co-embedding H&E + clinical text; H&E-to-mIF Recall@50 0.611

KEYNOTE-789

Yang et al. JCO 2024 — EGFR+ post-TKI; pembrolizumab + chemo vs chemo; no PFS/OS benefit

CodeBreaK 200

de Langen et al. Lancet 2023 — KRAS G12C; sotorasib vs docetaxel; PFS improved, OS not

Resistance

Tsai et al. JCI — KRAS G12Ci resistance mechanisms and TME remodeling

Review

Nat Rev Clin Oncol 2025 — NSCLC TME crosstalk and therapeutic implications

github.com/aonkondey01/PEAT-Nucleate-BIoHack-2026

Thank You

1,053 slides

TCGA LUAD + LUSC

29 signatures

HistoTME screening

3 AI models

Phoenix, GigaTIME, Haiku

1 agent

AI Tumor Board Agent

Questions?

Mutation is the lock. TME is whether the key turns.

PEAT-Nucleate BioHack 2026 • github.com/aonkondey01/PEAT-Nucleate-BIoHack-2026