Predicting personalized drug combinations for cancer treatment with deep learning and patient data

• Applicant: Dr. Michael Strasser
• Subject Area: Epidemiology and Medical Biometry/Statistics ; Bioinformatics and Theoretical Biology
• Term: from 2016 to 2019
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 324453217

Targeted treatment of cancer is currently hampered by two major obstacles: emerging drug-resistance and tumor heterogeneity. Tumor heterogeneity complicates patient stratification and design of treatment plans because each patient carries a unique tumor that is sensitive or resistant to different drugs. Resistance is the main reason for ultimate treatment failure. Combination therapy, where multiple drugs are applied simultaneously, can overcome both tumor resistance and tumor heterogeneity. However, designing successful drug combinations is difficult as standard high throughput screening is highly inefficient for drug combinations. Therefore, computational methods to predict successful drug combinations are highly desired. Existing computational prediction methods do not consider the dynamics of the cell's underlying gene regulatory network, i.e. how this network responds to a drug perturbation, and neglect patient specific information, e.g. their genotype. With the amount of patient-specific molecular profiling data rapidly increasing, computational methods that can take into account and exploit this wealth of personalized data are currently missing.In this project I propose novel, data-driven approaches to identify drug combinations for cancer treatment, which incorporate available patient specific information. The goal of this project is to generate computational predictions of drug combinations for efficient and resistance-free cancer treatment which can subsequently be tested experimentally. The concepts of dynamical systems and cancer attractors will provide theoretical foundations, while state of the art deep learning methods allow to leverage the information contained in large, publicly available datasets on cancer research. Particular emphasis is put on personalized drug combinations based on a patient's genotype, as it has strong influence on drug response and patient-specific genomic information is already routinely collected. Specifically, I will develop a method based on the dynamic systems theory of cancer and transcriptomics data to predict new drug targets for differentiation therapy, i.e. reprogramming cancer cells into a non-proliferate cell state. Next, drug combinations which modulate this set of predicted targets in individual patients are identified using deep learning algorithms, which are trained on the wealth of available cancer genomics and transcriptomics data. To account for tumor heterogeneity, I will use existing, imaging-based single-drug screening assays to identify tumor subpopulations resistant to certain drugs and show how a combination of multiple low efficiency drugs can be combined to target the entire tumor.The approaches developed in this project will contribute significantly to the future development of cancer combination therapies as they provide a data-driven framework for personalized cancer treatment in the upcoming age of big data medicine.

• DFG Programme: Research Fellowships
• International Connection: USA
• Host: Professor Sui Huang, Ph.D.