COMPASS, a novel AI model that combines transcriptomic data with tumor-immune information, predicts immunotherapy response more accurately than existing biomarkers across different cancer types, shows data published in Nature Medicine.
The model also offers insight into why a patient may or may not respond to treatment, giving clinicians a better understanding of why responses might not be as expected.
To date, predicting response to immune checkpoint inhibitors (ICIs) has largely relied on tumor mutational burden (TMB) and PD-L1 immunohistochemistry. In addition, researchers have developed transcriptomic signatures that capture T cell dysfunction or score immune checkpoint activity, while network based and machine learning approaches have tried to incorporate gene interaction data, but each of these approaches has limitations because people undergoing ICI treatment often respond in unexpected ways.
“In our benchmarking against 22 of these methods, previous approaches showed inconsistent performance once tested across different cancer types and treatments, which is the core problem COMPASS addresses,” said study senior author Marinka Zitnik, PhD, associate professor of biomedical informatics in the Blavatnik Institute at Harvard Medical School.
She told Inside Precision Medicine that “most AI models for predicting immunotherapy response are validated within a single cancer type or a single drug class, which limits how much they generalize once you move to a new tumor type, a new therapy, or a different hospital’s data. This is significant because it means we can have a model that predicts response in one population of cancer patients, but fails when applied to a different population, a different tumor type, or a different treatment regimen, which is exactly the kind of gap that keeps promising biomarkers from ever reaching routine clinical use.”
Zitnik explained that “COMPASS is different because it is pretrained on transcriptomic data from over 10,000 tumors spanning 33 cancer types before it ever sees a clinical outcome, so it learns broad, biologically grounded patterns of tumor immune biology.”
The model uses a “concept bottleneck architecture” which means that rather than going straight from gene expression data to a prediction, it forces the information through an intermediate layer, or bottleneck, before making the final prediction. In this case, the bottleneck is biologically defined immune concepts, like T cell exhaustion, macrophage activity, or transforming growth factor (TGF)-β signaling. This means clinicians and researchers can see which biological programs are driving a given patient’s predicted response or resistance.
Zitnik and team tested the performance of COMPASS in 1133 patients from 16 clinical cohorts spanning seven cancers and six ICI regimens. The found that, compared with the second-best performing biomarkers of the 22 tested (these varied by cohort), COMPASS improved accuracy by 8.5 percentage points and area under the precision-recall curve by 15.7 percentage points, on average.
“In a setting where response rates to immunotherapy are already low, a meaningful jump in prediction accuracy translates directly into better patient selection,” said Zitnik. “The improvements we saw mean fewer patients would be steered toward a treatment unlikely to help them, and more responders would be correctly identified for therapies that could extend their lives. This is important in cancer where checkpoint inhibitors carry toxicity risks and cost, so gains in identifying who is likely to benefit can meaningfully change treatment decisions and clinical trial design at scale.”
In survival analyses, the team found that individuals with a COMPASS response probability (PR) of 0.5 or higher had a 1-year overall survival rate of 86% compared with 40% for those with a PR below 0.5, yielding a hazard ratio for survival of 4.7.
Performance varied by cancer type and cohort size, but the researchers note that COMPASS was still able to accurately predict response for cancer types it had never seen.
“When we excluded lung adenocarcinoma entirely from training and tested only on that held out cohort, COMPASS still achieved 76.5% accuracy,” said Zitnik. “We saw similarly strong cross cancer generalization for urothelial carcinoma, melanoma, and other tumor types.”
Furthermore, COMPASS achieved 85.3% accuracy for predicting a response to combination therapies (ipilimumab plus pembrolizumab) when trained only on monotherapy cohorts.
In addition to predicting treatment response, the tool also generates a personalized response map for each patient that traces exactly which genes and immune programs are shaping their prediction.
“This explainable AI feature means a clinician is not just given a predicted ‘yes’ or ‘no’, they can see whether a patient’s tumor looks inflamed but is failing due to TGF-β driven suppression or whether it looks immune desert but still shows residual cytotoxic activity that could respond to treatment,” said Zitnik.
“This kind of mechanistic insight could help with patient stratification in clinical trials, indication selection when a drug is being tested in a new cancer type and generating hypotheses about resistance mechanisms that could point toward combination therapies.”
The investigators are now planning to move from using bulk RNA sequencing data for training to single cell and spatial transcriptomic data. They are also exploring how to connect COMPASS with AI agent systems, to help translate the mechanistic insights into more autonomous research workflows.
The post Predicting Cancer Immunotherapy Response Better with COMPASS AI Model appeared first on Inside Precision Medicine.