Immunotherapy Choices Grow More Complex as COMPASS Puts Its AI Model Before the Data and Code

Immune checkpoint inhibitors have reshaped the therapeutic imagination for some cancers, but they have also left a difficult reality: with treatments that are similarly expensive and similarly capable of causing immune-related side effects, some patients achieve durable responses, while others receive almost no benefit. The truly difficult question is not whether immunotherapy can work, but how to determine before treatment which tumor environment has a chance of being reignited.

A study published in Nature Medicine on July 3 describes COMPASS, a pan-cancer foundation model centered on tumor transcriptomes and designed to predict patients' responses to immune checkpoint inhibitors. According to the paper's abstract, the research team tested the model across 16 clinical cohorts, covering 7 cancer types and 6 immunotherapy regimens, and compared it with 22 benchmark methods; the results showed that COMPASS performed better across cancer types and drug contexts.

The input for this kind of model is not imaging, nor a single mutation, but tumor gene expression data. Public documentation for the COMPASS code repository and PyPI package shows that users can make predictions using gene expression matrices in TPM format, with outputs giving probabilities for two classes: non-response and response. The same tool also supports feature extraction, fine-tuning, and re-pretraining. In other words, it was designed as a tool researchers can operate, not merely as a closed model inside a paper.

One aspect of COMPASS with greater clinical significance is that it attempts to connect prediction back to interpretable concepts in tumor immunology. The project documentation says the model can provide features at the gene level, gene set level, and at the level of 44 high-level concepts; the download page also lists gene encodings, cancer type encodings, concept details, and example data. This does not mean the model can already explain causal mechanisms, but it at least gives researchers an opportunity to examine whether predictive signals fall within a more plausible immunological or tumor biology context.

The public data also make the boundaries of the research clearer. The COMPASS website lists multiple immunotherapy cohorts and model weights, including a foundation model pretrained on TCGA, PFT and LFT models trained using data from 1,133 immunotherapy patients, and models for drugs including atezolizumab, ipilimumab, nivolumab, and pembrolizumab. This information helps external teams reproduce and compare results, but it also reminds readers that the model's strengths and weaknesses remain constrained by the composition of existing cohorts, sequencing platforms, definitions of clinical endpoints, and the representativeness of patient populations.

For physicians and patients, the most reasonable current positioning for COMPASS remains as a research and trial stratification tool, not as a clinical arbiter that determines medication use on its own. Retrospective cross-cohort validation can show potential for generalization, but it cannot replace prospective clinical trials. If the model is to enter real-world care in the future, it will still need to answer questions about specimen handling, gene expression standardization, data drift across hospitals, responsibility for misclassification, and how regulators should review updateable AI models.

The point of this study, then, is not only the phrase "AI predicts immunotherapy response" itself, but that it puts the data, model weights, and software package in a position where they can be questioned. As cancer treatment increasingly depends on molecular stratification, the next threshold will lie not only in algorithmic accuracy, but also in whether models can withstand stress tests across different populations, different workflows, and real clinical decision-making.