Machine learning may help to uncover new immunotherapy targets

Results from a study presented at the ESMO Targeted Anticancer Therapies Congress 2024 (Paris, 26–28 February) revealed that a machine learning framework, built on hits entering early clinical trials, achieved an area under the receiver operating characteristic curve of more than 0.75 on a test set of multimodal molecular data (Abstract 1O). These data comprised 350 samples from single-cell atlases, bulk exome and transcriptome profiles from 1,317 patients treated with immune checkpoint inhibitors and immunopeptidomic data from 60 donors. The model was augmented with biological knowledge from gene regulatory networks, protein–protein interactions and disease links in addition to mechanistic information from CRISPR tumour T cell co-cultures and single nucleotide polymorphism-phenotype associations. Further investigation showed that model predictions could enrich for subsequently collated targets not included in the test set, and they could rank hits by clinical phase and identify genes associated with response to immune checkpoint inhibitors in unseen lung cancer trials (p<0.001 in each case). Initial testing of four hits identified showed evidence of macrophage repolarisation. 

“Diligent laboratory work and classical statistical methods were highly successful in the development of the first immune checkpoint inhibitors; however, to enhance response rates, we need new targets, for example, in the form of new tumour antigens or novel host immune checkpoints or both,” says Prof. Jakob Kather from TUD Dresden University of Technology, Germany. He continues, “To this end, we can use the huge amount of data available from clinical trials, the real-world setting and research laboratories as a valuable resource for discoveries, and realistically, machine learning is the only way to make the most of all this information.”

Using artificial intelligence (AI) to interrogate clinical data is already showing value in uncovering predictive biomarkers that can be used to tailor immunotherapy (Lancet Oncol. 2023;24:1411–1422; ESMO Daily Reporter, ESMO 2023). “Combining machine learning from clinical trial data with biological knowledge, as the researchers have done in the study presented in Paris, improves understanding of the processes involved in immunotherapy response. As well as aiding patient selection, we can take this one step further and look into the mechanisms responsible and develop new effective treatments,” comments Kather. He highlights multi-modal machine learning, as used in the presented work, as a recent technological advance that could unlock greater potential: “It is now possible to use one machine learning system to process all the different types of data – genomic and proteomic profiling, imaging and functional data – which increases our understanding of the interactions between these different systems and enables us to extract even more valuable information from diverse sources to build a clearer picture.”

AI has the potential to transform drug discovery and its use is increasing in oncology and other fields (Nat Rev Drug Discov. 2022;21:175–176); however, there are still questions to be answered around the extent of its capabilities (Nature. 2023;622:217). Kather concludes, “Rather than replace traditional preclinical methods, machine learning is likely to be most effective when used to complement tried-and-tested research techniques.”