Development and real-world validation of a multi-modal machine learning model to predict response to immune checkpoint inhibitors in cancer patients.

Abstract

e14532 Background: Immune checkpoint inhibitors (ICIs) have transformed cancer care, offering the promise of durable, life-changing responses for cancer patients. However, these benefits are realized in only about 20% of cases, with most patients failing to respond. This disparity underscores the critical need for more advanced predictive tools to refine patient selection and fully unlock the potential of ICIs. We present a machine learning (ML) model, called ioBERT, that predicts ICI response using multi-modal data, including clinical information, drug target and structure features, and genomic alteration data commonly captured by commercial NGS panels in real-world clinical settings. Methods: We curated a clinicogenomics dataset of ~2,900 cancer patients treated with FDA-approved anti-PD1 and anti-CTLA4 therapies. Inclusion criteria for ioBERT development included real-world ICI outcomes (rw-OS), clinical data (e.g., disease type, subtype, sex, tumor stage), and genomic alteration data for ~200 genes commonly assessed by commercial NGS panels. NGS profiles were filtered to biopsies collected within 2 years before ICI initiation. Samples with whole-exome data were subset to these ~200 genes. Multi-modal inputs were processed using custom and publicly available encoders, including BioBERT and ProteinBERT, to generate numerical embeddings, which were integrated into a neural network to model relationships with rw-OS outcomes. ioBERT predictions were validated on hold-out and independent rw-datasets. Results: ioBERT outperformed top-performing models from the Anti-PD1 Response Prediction DREAM Challenge in predicting ICI response, as measured by the concordance index (C-index). Its performance was independently validated in a real-world dataset, outperforming existing ICI biomarkers including tumor mutational burden, and PD1 expression where available. Direct comparisons in real-world datasets were limited by other models’ reliance on RNAseq, which is not routinely available in clinical settings. Conclusions: Our study demonstrates the value of integrating multi-modal data to achieve a comprehensive view of patient and tumor biology, enabling more accurate ICI response predictions with ML. Our approach has the potential to improve patient outcomes by identifying those most likely to benefit from ICI treatments and reducing unnecessary therapies. Furthermore, ioBERT can be integrated with off-the-shelf NGS assays, facilitating rapid clinical translation without the need for bespoke assays or costly sequencing methods.