PredictIO: A Package For Meta-Analysis of Immunotherapy Clinical Trials in Cancer
Author(s): Benjamin Haibe-Kains
Affiliation(s): University of Toronto
Social media: http://twitter.com/bhaibeka
Clinical profiling studies have shed light on molecular features and mechanisms that modulate response or resistance to immunotherapy but their predictive value remains largely unclear. We (Bareche et al., Annals of Oncology 2022) and others (Litchfield et al., Cell 2021) have recently curated a compendium of public datasets of DNA, RNA and clinical profiles of patients treated with immunotherapy. These multimodal datasets have been processed transparently and reproducibly using our ORCESTRA platform (Mammoliti, Nat Commun 2021) and have been made available to the community (21 datasets to date; orcestra.ca/clinical_icb). Leveraging our compendium of immunotherapy clinical datasets, we developed the PredictIO R package implementing a meta-analytical pipeline to assess the predictive value of molecular predictors robustly. We first used PredictIO to compute the association between immunotherapy response and established biomarkers (eg, tumor mutation burden or CD8 gene expression) or published gene signatures (eg, mutational or gene expression signatures). We have also implemented a de novo RNA signature discovery pipeline built upon our meta-analysis pipeline and developed a new predictor of immunotherapy response which seems to outperform state-of-the-art biomarkers and signatures; these findings have been published in (Bareche et al., Annals of Oncology 2022). These meta-analyses identified a subset of signatures that were significantly predictive in a pan-cancer setting, while observing that their predictive value depends on the cancer types. While promising, these initial studies suffer from severe limitations in terms of data availability for specific cancer types and the lack of frameworks to develop and validate multi-omics predictors of immunotherapy response in a collaborative and scalable way. We are currently working towards integrating more datasets in our compendium. In addition, we are working toward the development of predictive models integrating DNA, RNA and clinical profiles of the patients treated with immunotherapy. All new functions will be integrated into our PredictIO R package and into our predictio.ca web-application. All the code is made publicly available under the open-source MIT license.