AI Reads Tumor Gene Activity, Potentially Making Immunotherapy Patient Selection More Precise

The most exciting aspect of immunotherapy is also the most difficult one in clinical practice: the same class of drugs may keep cancer under long-term control in some patients, while having almost no effect in others. If physicians can determine more accurately before treatment who is likely to benefit, they can reduce blind waiting, and patients may spend less time undergoing ineffective treatment.

According to Medical Xpress, COMPASS, an artificial intelligence model developed by a Harvard Medical School research team and collaborators, aims to predict how cancer patients will respond to immune checkpoint inhibitors (ICIs). The findings were published in Nature Medicine. These drugs work by releasing immune brakes such as PD-1, PD-L1, or CTLA-4, allowing T cells to recognize and attack cancer cells again, but responses vary widely across cancer types and individuals.

COMPASS does not analyze a single gene mutation, but rather the activity of nearly 16,000 genes related to immune cell states, interactions with the tumor microenvironment, and signaling pathways. The research team first trained the model using data from more than 10,000 tumors across 33 cancer types in The Cancer Genome Atlas, then fine-tuned and tested it using results from 16 immune checkpoint inhibitor clinical trials covering 7 cancer types.

For evaluation, the researchers used a leave-one-clinical-trial-out design: data from one trial were temporarily removed, the model was asked to predict which patients in that trial would respond, and the predictions were then compared with the actual results. The report noted that COMPASS improved predictive performance by about 8.5% to nearly 10% on average compared with the best existing methods, and maintained its advantage across different cancer types, drug regimens, transcriptome sequencing platforms, and biopsy sources.

One meaningful aspect of the study is that the model design emphasizes interpretability rather than merely outputting a high or low score. The report said COMPASS can identify biological clues behind its predictions. For example, some tumors that appear to have sufficient immune cell infiltration still fail to respond, possibly because of gene-expression programs that suppress immune function. Conversely, some “immune desert” tumors, despite lacking typical infiltration features, may still carry molecular signals that promote other immune activities.

If these results hold up in future prospective clinical trials, COMPASS could become an auxiliary tool for physicians choosing immune checkpoint inhibitors, and could also help clinical trials recruit suitable participants more precisely. For researchers, signals generated by an interpretable model may also point to new drug targets or hypotheses for combination therapy.

However, the current evidence still comes mainly from existing data and retrospective validation, and cannot be directly equated with a clinically usable diagnostic tool. Before it truly enters care settings, several questions must still be answered: how stable the model is across different hospitals and populations, whether the required gene-expression testing can be standardized, how prediction errors would affect treatment choices, and how regulators will assess this type of AI medical tool that evolves as data are updated.