Cellular Level Pilot: Predictive Modeling for Pre-Clinical Screening
The cellular level pilot of the Joint Design of Advanced Computing Solutions for Cancer (JDACS4C) program focuses on developing predictive models, both computational and experimental, to improve pre-clinical therapeutic drug screening.
The pilot team uses advanced computational technologies including machine and deep learning to rapidly develop, test, and validate predictive pre-clinical drug efficacy models, in collaboration with NCI’s Patient-Derived Models Repository.
The goal of this pilot is to accelerate the identification of promising new treatment options for precision oncology.
With shared expertise across the JDACS4C collaboration, this pilot is jointly led by:
- Dr. James Doroshow, NCI, Director, Division of Cancer Treatment and Diagnosis
- Dr. Yvonne Evrard, NCI, Frederick National Laboratory for Cancer Research
- Rick Stevens, Department of Energy, Argonne National Laboratory
Aims of the Pilot
- Develop reliable machine-learning-based predictive models of drug response that enable the projection of screening results from and among cell lines, cell-line xenografts, and patient-derived xenograft (PDX) models
- Integrate uncertainty quantification and optimal experimental design to assert quantitative limits on predictions and to recommend experiments that will improve predictions
- Develop Hybrid Predictive Models that support the graded introduction of mechanistic models into the machine-learning framework for incorporation of molecular pathways and gene regulatory networks
The multidisciplinary team is developing machine learning-based predictive models trained on experimental data from many sources, including:
- Cancer cell lines
- Patient-derived xenografts (PDXs)
In the past two years, the team has established a large database, focused primarily on RNAseq data, to train the deep learning models with over a dozen integrated NCI data sources including:
Using gene expression data from the integrated database, the team has developed deep learning models for predicting tumor response to single drugs and drug combinations.
Looking forward, the pilot team plans to focus on developing hybrid computational models, those employing both data-driven and biological/mechanistic understanding. This integrated approach will enable:
- Continued improvement and validation of the computational models
- Increased potential to advance risk identification, pre-clinical drug screening, and treatment selection for patients