Background

Limitations and challenges of immune checkpoint inhibitors

Immune checkpoint inhibitors (ICIs) have revolutionized cancer treatment by activating the patient's immune system to attack cancer cells. While demonstrating high efficacy in inducing long-term survival in some patients, only 20-30% of patients actually respond, which is a significant limitation. Therefore, identifying biomarkers to predict drug response before treatment is a critical task.

Previously, tumor mutational burden (TMB) and PD-L1 expression were used as predictive tools. However, these markers have limitations in terms of consistent accuracy across different cancer types and drugs. This is because the tumor microenvironment and immune regulation mechanisms vary greatly and are complex in each cancer type. Consequently, there is a need to establish a universal predictive model that comprehensively reflects individual genomic characteristics and functions regardless of cancer type.

Key Findings

AI prediction pathways made transparent with 44 immune concepts

A team led by Professor Marinka Zitnik at Harvard Medical School, in collaboration with Roche, developed COMPASS, an artificial intelligence (AI) foundation model that predicts treatment response regardless of cancer type or treatment. The research results were published in the international journal 'Nature Medicine'. The model analyzes response based on bulk tumor transcriptome data.

To overcome the limitations of black-box AI, COMPASS applies a 'Concept Bottleneck Transformer' architecture. It first extracts scores for 44 biological immune concepts, such as immune cell status, intercellular interactions, and signaling pathways, from the input patient's transcriptome data. Subsequently, it combines these scores to calculate the final response. This allows clinicians to transparently verify the biological basis of the prediction.

Outperforms 22 existing models in pan-cancer prediction performance

The research team trained the model using 33 cancer types and 11,084 tumor data from The Cancer Genome Atlas (TCGA). Subsequently, it was validated using 16 independent clinical cohorts, including 7 cancer types and 6 immunotherapies. The analysis showed that COMPASS improved the average accuracy by 8.5% compared to 22 existing prediction methods. The area under the precision-recall curve (AUPRC) also showed a 15.7% increase compared to existing models.

Significance and Prospects

Integrated analysis opens a new era of precision medicine

This study overcomes the limitations of existing methods that rely on single biomarkers and pioneers a new approach that encompasses tumor and immune system interactions. The model has demonstrated high generalizability by maintaining stable predictive power even in rare cancer types or new treatment conditions not exposed during the learning process. The ability to provide detailed patient-specific drug resistance factors through 44 immune concept scores is also a strength.

However, there are still challenges to be addressed before full implementation in clinical practice. Although COMPASS has achieved excellent results in multiple cohorts, its effectiveness needs to be demonstrated through prospective clinical trials to establish it as a tool for making actual treatment decisions. The time and cost required for transcriptome analysis of patient tumors are also barriers to be overcome. If accessibility to genomic data is expanded and the analysis process is simplified, it is expected to become a key to personalized precision medicine.