“Learning and actioning general principles of cancer cell drug sensitivity” is out on Nature Communications. A monolithic study brilliantly crafted by Francesco Carli which offers a new strategy to predict potential anti-cancer drugs based on patients’ transcriptomics data.
We trained a model (called CellHit) to forecast sensitivity of cancer cell lines to thousands of drugs from large datasets such GDSC and PRISM. We leveraged Large Language Models (MixtralAI and ChatGPT) to annotate drugs with biological pathways (Reactome) likely representing their Mechanism of Action (MOA). We used this information to, on one hand, demonstrate that biologically relevant information is learnt by the algorithm to predict sensitivity to the drug and, on the other, to fine-tune and improve the performance of our models (called MOA-primed).
We then combined our models with Celligner, which is used to align RNAseq data of patients and cancer cell lines. We used this pipeline to process the transcriptomics data of over 10k patients from TCGA, demonstrating that our model is able to predict mono- and combination-therapies prescribed for specific cancer types. In collaboration with the groups of Gioacchino Natoli (IEO Istituto Europeo di Oncologia) and Chiara Maria Mazzanti (Fondazione Pisana per la Scienza), we validated our approach on data from patients affected by two highly letal solid tumors, i.e. pancreatic adenocarcinoma and glioblastoma multiforme. In both cases, we carried out predictions on bulk RNAseq from patients, which were profiled for drug sensitivity with CellHit. We identified subgroups of patients with predicted distinct responsiveness to drugs, and experimentally validated our predictions on derived cell lines.
The CellHit pipeline is freely available at this URL (cellhit.bioinfolab.sns.it) and enables rapid profiling of input bulk RNAseq data from cancer patients to generate new hypotheses for drug repositioning. This is our two cents to help accelerate the path to discovering new anti-cancer therapies.
We are indebted to all our fantastic collaborators without whom this work wouldn’t have been possible, an in particular: Pierluigi Di Chiaro, Miquel Duran-Frigola, Gioacchino Natoli, Chiara Maria Mazzanti, Mariangela Morelli, Anna Luisa Di Stefano, and of course all the other co-authors and members of the Bioinformatics group of the BIO@SNS of Scuola Normale Superiore who participated to this work.
It was a fantastic scientific ride, and a truly team effort!!!
A huge thanks also goes to our sponsors for their fundamental support to our research: Fondazione AIRC per la Ricerca sul Cancro ETS, THE Tuscany Health Ecosystem and the “Dipartimento di Eccellenza” of Scuola Normale Superiore.
