1O - Uncovering new immunotherapy targets with machine learning and ex vivo validation

Background

Immunotherapy has revolutionised cancer therapy but current immune checkpoint inhibitors (ICI) produce low response rates in most cancers, indicating that new therapeutic options are needed. Conventional immune-oncology (IO) discovery uses preclinical models with limited translation capacity as they do not fully recapitulate human tumour complexity. We use multimodal patient molecular data with modern machine learning (ML) methods to identify new IO targets with improved clinical potential.

Methods

Phrasing target identification as a binary classification problem, we built an ML framework using hits that entered at least stage I clinical trials. We trained our model on multimodal patient data from single cell atlases of n=350 samples, bulk exome and transcriptome profiles of n=1,317 patients treated with ICIs, and immunopeptidomic data from n=60 donors. We augmented this with biological knowledge from gene regulatory networks, protein-protein interactions, and disease-gene links. Finally, mechanistic information was provided through n=7 genome-wide CRISPR tumour-T cell co-cultures and n=15,442 SNP-phenotype associations.

Results

Our ML framework achieved test set ROC-AUC>0.75. To assess model generalisation, we collated targets not present in our training data as they entered stage I IO clinical trials after we froze our database (at the end of 2019). We found model predictions enriched for these targets (p<0.001). Auxiliary validation tasks were used to further test generalisation. We demonstrated that our approach could rank hits by clinical phase (p<0.001) and identify genes associated with ICI response in unseen lung cancer clinical trials (p<0.001). These results support that our model learned broad aspects of cancer-immune interactions. We reviewed 41 hits for biological, pharmacological, and commercial tractability and began targeting four in patient-derived explants and organoid co-cultures. Initial data show that testing them results in macrophage repolarisation.

Conclusions

Our target discovery approach analyses multimodal patient data with ML to uncover new IO hits. Its success on multiple benchmark and validation tasks supports incorporating ML pipelines in target discovery.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

UCL.

Funding

Cancer Research UK.

Disclosure

C. Swanton: Financial Interests, Personal, Invited Speaker, Activity took place in 2016: Pfizer, Celgene; Financial Interests, Personal, Invited Speaker, October 26th 2020: Novartis; Financial Interests, Personal, Invited Speaker: Roche/Ventana, BMS, AstraZeneca, MSD, Illumina, GSK; Financial Interests, Personal, Advisory Board, November 12th, 2020: Amgen; Financial Interests, Personal, Advisory Board, Current - since 2018: Genentech; Financial Interests, Personal, Advisory Board: Sarah Canon Research Institute; Financial Interests, Personal, Advisory Board, Joined October 2020, and stock options: Bicycle Therapeutics; Financial Interests, Personal, Other, Consultancy: Medicxi; Financial Interests, Personal, Advisory Board, Member of the Science Advisory Board, and stock options until June 2021: GRAIL; Financial Interests, Personal, Other, Consultancy agreement: Roche Innovation Centre Shanghai; Financial Interests, Personal, Advisory Board, 29 November - 1 December 2022: Novartis; Financial Interests, Personal, Invited Speaker, Oncology Collective - 2nd Nov - 4 Nov 2022 - Atlanta, USA: Roche; Financial Interests, Personal, Advisory Board, ctDNA Advisory Board - 24th March 2023: AstraZeneca; Financial Interests, Personal, Invited Speaker, Pfizer Oncology 'Leading the revolution for the future: Pfizer; Financial Interests, Personal, Full or part-time Employment, Chief Clinician since October 2017: Cancer Research UK; Financial Interests, Personal, Ownership Interest, Co-Founder of Achilles Therapeutics, and stock options: Achilles Therapeutics; Financial Interests, Personal, Stocks/Shares, Stocks owned until June 2021: GRAIL, ApoGen Biotechnologies; Financial Interests, Personal, Stocks/Shares: Epic Biosciences, Bicycle Therapeutics; Financial Interests, Institutional, Research Grant, Funded RUBICON grant - October 2018 - April 2021: Bristol Myers Squibb; Financial Interests, Institutional, Research Grant, Collaboration in minimal residual disease sequencing technologies: Archer Dx Inc; Financial Interests, Institutional, Research Grant: Pfizer, Boehringer Ingelheim; Financial Interests, Institutional, Invited Speaker, Chief Investigator for the MeRmaiD 1 and 2 clinical trials and chair of the Steering Committee: AstraZeneca; Financial Interests, Institutional, Research Grant, Research grant from Oct 2019 - July 2023 - Genetics of CIN and SCNAs for Targeted Discovery (SCEPTRE): Ono Pharmaceutical; Financial Interests, Institutional, Research Grant, Research Grants from 2015: Roche; Financial Interests, Personal, Other, Co-chief investigator: NHS-Galleri Clinical Trial; Financial Interests, Institutional, Research Grant, from October 2022: Personalis; Non-Financial Interests, Personal, Principal Investigator, Chief Investigator for MeRmaiD 1 and 2 clinical trials: AstraZeneca; Non-Financial Interests, Personal, Member of Board of Directors, From 2019-2022: AACR; Non-Financial Interests, Personal, Other, Board of Directors: AACR; Non-Financial Interests, Personal, Advisory Role, EACR Advisory Council member: EACR. K.R. Litchfield: Financial Interests, Personal, Invited Speaker: Roche Tissue Diagnostics; Financial Interests, Personal, Other, Consulting work: Kynos Therapeutics, Monopteros Therapeutics, Tempus; Financial Interests, Personal, Invited Speaker, Invited speaker: Ellipses Pharma; Financial Interests, Institutional, Research Grant: Ono/LifeArc; Financial Interests, Institutional, Research Grant, Research Funding: Genesis Therapeutics; Non-Financial Interests, Institutional, Proprietary Information, Collaboration on data analysis: BMS. All other authors have declared no conflicts of interest.