Predicting personalized drug combinations for cancer treatment with deep learning and patient data
Bioinformatics and Theoretical Biology
Final Report Abstract
Cancer therapy is moving from general purpose chemotherapy agents to targeted drugs and drug combinations, promising less side effects and less drug resistance. However, drug combinations are currently determined by expensive experimental screening, drug resistance is still the main reason for treatment failure and the origin of resistance is still being debated. Within this fellowship, machine learning methods were combined with a dynamical systems interpretation of drug perturbation to predict drug combinations based on transcriptional response to single drug perturbation screens from the LINCS consortium. While predicted combinations overlapped significantly with candidates found in independent large-scale experimental drug combinations screens, it also revealed the limitations of the used data, in particular its lack of single cell resolution. However, the developed deep latent variable models are general and this project formed the foundation for an application to single-cell resolved perturbation screening data e.g. CRISPR screens. To investigate the action of drugs on a single cell level, we combined single cell RNAseq with lineage tracing via genetic barcodes on naive and fludarabine treated HG3 cells. In this nominally homogeneous cell population, two transcriptionally distinct subpopulations showed different response to fludarabine treatment as determined by genetic barcodes. This pilot study unveiled the importance of single-cell analysis in drug response and established the use of genetic barcoding in drug perturbation experiments.
Publications
- Cancer cell population growth kinetics at low densities deviate from the exponential growth model and suggest an Allee effect. PLOS Biology 17(8): e3000399.
Johnson K.E., Howard G., Mo W., Strasser M.K. , Lima E.A.B.F., Huang S., Brock A.
(See online at https://doi.org/10.1371/journal.pbio.3000399)
