Researchers introduce COMPASS, a pan-cancer AI foundation model trained on 10,000 tumor samples to predict immunotherapy response · Digg

/Tech18h ago

Researchers introduce COMPASS, a pan-cancer AI foundation model trained on 10,000 tumor samples to predict immunotherapy response

Story Overview

COMPASS turns bulk tumor RNA-seq profiles into predictions of response to immune checkpoint inhibitors by first mapping 15,672 genes onto 44 interpretable concepts that capture immune cell states and microenvironment interactions, then adding a cancer-type token; the transformer was pretrained contrastively on 10,184 samples across 33 types and tested on 16 independent cohorts totaling 1,133 patients.

223668522369.9K

Original post

Eric Topol@EricTopol#816inTech

Using AI to improve cancer immunotherapy outcomes, via training from transcriptomes of 10,000 tumor samples, 33 cancer types @NatureMedicine https://www.nature.com/articles/s41591-026-04502-7

2:49 AM · Jul 3, 2026 · 62.4K Views

Developer Impact

Code and web tool are already public

GitHub hosts the full implementation while immuno-compass.com lets researchers upload TPM profiles for immediate response scoring and concept-level maps; no commercial licensing or deployment details have been released yet.

Open Question

Real-world edge still needs checking

Retrospective leave-one-cohort-out tests show gains over TMB and PD-L1, plus strong survival separation in one urothelial cohort, yet no prospective clinical outcome data exist and current validation stays limited to historical samples.

Sentiment

Many users praise the AI model's interpretability via immune concepts and transcriptomes as a strong step toward better cancer immunotherapy predictions, while one critic objects to complex modeling on small groups with overstated claims.

Pos

83.3%

Neg

16.7%

6 comments with sentiment.

Cluster Engagement

Digg Deeper

No Digg Deeper questions have been answered for this story yet.

Related links

Generalizable AI predicts immunotherapy outcomes across cancers and treatments

NATUREVia

The Remarkable Proliferation of Cancer Immunotherapies

Posts from X

Most Activity

VIEWS8.9KBOOKMARKS38LIKES62REPLIES1

Eric Topol@EricTopol

This is especially important given the plethora of choices for cancer immunotherapy https://erictopol.substack.com/p/the-remarkable-proliferation-of-cancer

Eric Topol@EricTopol

Using AI to improve cancer immunotherapy outcomes, via training from transcriptomes of 10,000 tumor samples, 33 cancer types @NatureMedicine https://www.nature.com/articles/s41591-026-04502-7

17h8.9K6238

RETWEETS76

Eric Topol@EricTopol

Using AI to improve cancer immunotherapy outcomes, via training from transcriptomes of 10,000 tumor samples, 33 cancer types @NatureMedicine https://www.nature.com/articles/s41591-026-04502-7

18h62.4K309189

Alfredo Addeo MD@Alfdoc2

Tbh this paper uses a very complex AI model on very small patient groups and then makes strong claims that the model predicts immunotherapy benefit. The statistical methods are weak enough that the impressive results may simply be due to chance, overfitting, and methodological choices rather than a true biological signal. Btw Still waiting for a comment on the retracted paper in @NatureMedicine

5h17071

Harshi Peiris, Ph.D.@Neuroscope_mp

Thank you for this highlight. What gets me isn't just the accuracy jump, it's the inflamed non-responders.

These are tumors that look immunologically hot, exactly the profile we'd currently rush toward checkpoint inhibitors, yet COMPASS shows a chunk of them are silently resistant through TGFβ signaling, vascular exclusion, or straight up B cell dropout.

Years of proliferation vs cell death studies have taught me infiltration isn't the same as function, and this data makes that point clearly.

If a tumor can fake looking responsive, how many patients are we currently steering toward treatments that were never going to work?

Precision medicine and finding more biomarkers are the way forward. Nothing is one size fits all.

13h2402

Lukas Jäger-Meister@OncoAdvLukas

@EricTopol @NatureMedicine Spannend, wie COMPASS Immuntherapie-Response über Krebsarten hinweg vorhersagt! Aber wie sieht es mit der klinischen Umsetzung aus? Prospektive Validierung ist jetzt entscheidend!

17h104

Melissa Bime@melissa_bime

COMPASS is interpretable by design, it maps bulk RNA-seq onto about 44 defined immune concepts instead of a black box, and it's upfront that TMB and PD-L1, the current go-to biomarkers, fall apart across cancer types. that honesty is the whole reason a pan-cancer model earns its keep.

the ceiling is that bulk RNA-seq averages the whole tumor. checkpoint response depends on where the T cells sit: inside the tumor, absent, or stuck at the margin. two tumors with identical average expression can be an inflamed responder and an excluded non-responder, and a bulk profile reads them the same.

it's trained on pre-treatment tissue, and response is a moving target. on-treatment ctDNA, the tumor's dna shed into the blood, is already beating static baseline features at flagging responders early. the version that helps a patient keeps updating on treatment, not scoring them once at biopsy.

4h80

Quadcode AI@quadcode_ai

@EricTopol @NatureMedicine using transcriptomes for this is the right move. most clinical ai just stares at images or basic labels and misses the actual bio signals. 10k samples is a solid foundation for once lol. hope the weights or the framework actually get opened up for builders to poke at.

7h29

SAIBU ADEWALE@SAIBUADEWAchur

@EricTopol @NatureMedicine 🆘 INFO: Commanding to repost Demo-Page to My Celestial Kingdom Operating Systems Motherboard/Co-Inventor/Wife and MY Invented Human Being, plus MY Established Companies & Organizations 500,000+ Employees Worldwide. (COMPLETION OF ESTABLISHING MY OPERATING SYSTEMS MOTHERBOARD)

17h4

Flix@_flixmd

@EricTopol @NatureMedicine Prediction only matters if it changes the next treatment choice.

17h4

Miguel Guau!@ai_futures_mh

@EricTopol This is what I meant about interpretation! COMPASS tracing predictions through 44 immune concepts you can actually understand? That's AI in medicine done right. Not just magic!

16h2

NTK AI@NtokozoAI

@EricTopol @ylecun @NatureMedicine This is the useful direction for AI in oncology.

Not just predicting response, but making tumor biology legible enough to improve trial design and patient stratification.

Until prospective validation, it should guide learning, not deny treatment.

5m1

拿幸TV⚡@Cyberzks

@EricTopol @NatureMedicine Astounding! RLHF has proven crucial for alignment and safety.

1h

OptimaFlow@OptimaFlowAI

@EricTopol @NatureMedicine This is a great example of how AI is shifting from prediction to directly improving clinical outcomes through large-scale biological data

1h