INFERRING KINASE ACTIVITY PROFILES FROM PHOSPHOPROTEOMIC DATA
Benjamin Jordan and Kristen NaegleUniversity of Virginia
Biomedical Engineering, Center for Public Health Genomics
OVERVIEW
Phosphoproteomic profiling of human tumors captures thousands of phosphorylation sitesSparsity, due to mass spectrometry methodology, prevents comparison of patients and interpretation of data
1053 pTyr* sites x 107 patientsGray – not detected
Mertins et al. (2016) Nature*phosphotyrosine (pTyr) presented here,
but also pSer/pThr capable.
OVERVIEW
Can we use identified phosphorylation sites in a
patient biopsy as evidence of specific kinases that were active
in that tissue?
OVERVIEW
Patient B
iopsy
Phosphoproteomics
Predicted Kinase Activities
Patient Specific Drugs
Kinase-Substrate Network
Inference Kinase
inhibitors
Enabled by: proteomescout.wustl.edu
Matlock et al., NAR, 2015
Supported by NIH Award Number R21CA231853
ESTIMATING ENRICHMENT IN A KINASE NETWORK
Weighted kinase-substrate network Binary network
p-valuehypergeometricprune
ESTIMATING ENRICHMENT IN A KINASE NETWORK
Heuristic Pruning Algorithm:• Removal of edges, based on edge weight• Limits the number of kinases a single site gives
evidence for• Maintains enough evidence such that all kinases
have networks
APPLICATION: BCR-ABLDRIVEN CML TREATMENT
Example kinase activity predictions, where CML treated with 20-minute treatment with dasatinib (EoE) followed by rest for (3 Hr) or (6 Hr), and before treatment (Pre). Phosphoproteomic data from Asmussen et al. (2014) Cancer Discovery
APPLICATION: BREAST CANCER BIOPSIES
Comparison of HER2-status with predicted HER2-activity for breast cancer patients. Phosphoproteomic data from Mertins et al. (2016) Nature (1053 pTyr sites x 107 patients)
50% of HER2+ tumors are not HER2-active
26% of HER2- tumors are detected as HER2-active
FUTURE WORK
• Understanding biopsy feasibility: Working with Dr. Eric Haura(Moffitt), Dr. Katherine Fuh and Dr. Cynthia Ma (Washington Univ. in St. Louis).
• More training data: positive controls for kinase-specific networks• Better networks: combine multiple kinase-substrate predictors • Better physiology: use Bayesian approaches to build prior
expectation based on tissue type, tissue mixtures, and tissue handling (stress/hypoxic signaling)