other-cancer

AI trained on spatial tumor images can distinguish between immune structures that help versus hurt cancer outcomes

Tertiary lymphoid structures (TLS) are organized immune cell aggregates that form within tumors, and they're attracting enormous clinical attention — immunogenic TLS correlate with better outcomes and checkpoint therapy response, while tolerogenic TLS can actually help tumors evade immunity. But standard bulk gene expression analysis can't distinguish between these two functionally opposite states.

This study developed a hierarchical graph neural network (GNN) that operates on spatial transcriptomics data (Visium), analyzing the architecture of TLS at multiple scales simultaneously. The AI model successfully classified TLS as immunogenic or tolerogenic based on their spatial cellular composition — a task previously impossible at scale.

Accurate TLS classification could transform how we predict immunotherapy response and could guide decisions about which patients to treat with immune checkpoint inhibitors.

Key Findings

A hierarchical graph neural network classifies TLS as immunogenic or tolerogenic from spatial transcriptomics data
Immunogenic TLS harbor germinal center reactions; tolerogenic TLS contain regulatory T cells and suppressive myeloid cells
Bulk transcriptomics cannot distinguish these functionally opposite TLS states
Spatial cellular architecture at multiple scales is key to classification accuracy
Model trained on 10x Visium spatial data

Implications

Accurate TLS classification at scale could become a standard component of immunotherapy response prediction. This would help identify patients most likely to benefit from immune checkpoint inhibitors and potentially guide TLS-targeted interventions to convert tolerogenic to immunogenic states. Clinically relevant for most solid tumor types.

Caveats

Preprint — not peer reviewed. Based on abstract only. Spatial transcriptomics remains expensive and not routinely available clinically. GNN performance on prospective patient cohorts needs validation. Clinical application requires integration with pathology workflows.

Source: bioRxiv — 2026-04-08

Landmark DCIS trial, CRC microbiome & AML combos

Two themes cut across this week’s 236 posts. The first is de-escalation: researchers are building rigorous evidence that patients can safely do less — skip lymph node surgery, forgo radiotherapy, watch pre-invasive lesions instead of cutting them out. The landmark result here is a randomized trial confirming that active surveillance

Colorectal Cancer Microbiome Discovery Platform

Analysis of 9,000+ patients revealed that only colorectal cancer — not other cancer types — consistently harbors distinct microbial communities, suggesting the gut microbiome has a unique relationship with colorectal carcinogenesis that could be exploited for diagnostics and treatment.

GBM Tumor Microenvironment Drug Target Atlas

Non-cancerous brain support cells were found to promote glioblastoma growth through paracrine signaling, and blocking this communication dramatically slowed tumor growth — revealing that GBM's microenvironment contains targetable vulnerabilities that are being systematically overlooked by drug…

BET Inhibitor Target Selectivity Database

BRD2 and BRD4 have fundamentally distinct functions — BRD2 prepares genes for activation while BRD4 triggers transcription — meaning that pan-BET inhibitors have been disrupting gene regulation in counterproductive ways, and the field now needs selective BRD4 inhibitors.