Illuminating the Dark Proteome: Integrating Machine Learning to Uncover Relevant Yet Understudied Cancer Targets

Overview

Traditional cancer research has largely focused on well-characterized proteins, leaving approximately 40–50% of the human proteome—the "dark proteome"—poorly understood and potentially rich in undiscovered therapeutic targets. To address this gap, this project developed a machine learning framework to systematically identify understudied proteins enriched in cancer that may represent novel therapeutic opportunities. Random Forest classifiers were trained on proteomic data from the NIH Clinical Proteomic Tumor Analysis Consortium (CPTAC), comparing tumor versus matched normal tissue across 8 cancer types. Models achieved over 95% accuracy and an AUC of 0.99 across all cancer types. Feature importance rankings identified the 200 proteins most critical to distinguishing tumor from healthy tissue per cancer type. These were cross-referenced with Pharos and the Human Protein Atlas to prioritize minimally researched proteins (<50 publications). Clinical validation via Kaplan-Meier survival analysis through cBioPortal confirmed statistically significant survival associations (p < 0.05) for 15 high-priority targets. From this pool, 3 proteins were selected as final candidates based on their upregulation in tumor tissue, favorable subcellular localization for therapeutic accessibility, and AlphaFold structural analysis confirming druggable binding conformations: GPRC5A in pancreatic ductal adenocarcinoma, proposed as an antibody target, MYO1E in pancreatic ductal adenocarcinoma, and GMIP in clear cell renal cell carcinoma, both proposed as small molecule inhibitor targets. Wet lab experimental validation further confirmed the expression and functional relevance of these candidates. This computational pipeline demonstrates how machine learning can systematically expand the therapeutic target landscape beyond well-studied proteins, revealing novel candidates inaccessible through conventional approaches.

Illuminating the Dark Proteome: Integrating Machine Learning to Uncover Relevant Yet Understudied Cancer Targets

Overview

Competition history

Related projects

Protein Assassins: Screening for Improved Degraders to Reverse Tumor Immortality

Copiloting Predictability with AI and ML to Beat the Second Hit in Colorectal, Breast and Lung Cancers

Integrating Machine Learning with 3D Organoid Modeling to Identify Biomarkers to Combat Drug Resistance in Cancer

Detection of Premalignant Pancreatic Cancer via Computational Analysis of Serum Proteomic Profiles

Structure-Guided In Silico Engineering of PROTACs for Targeted S100A16 Proteasomal Degradation via E3 Ligase Recruitment

Integrating Mathematical Modeling with Machine Learning for Cancer Driver Gene Identification

Using Machine Learning Techniques to Detect Mutant p53 Transcriptional Activity

Synthetic Engineering with ImmunoGenAI: Multi-Modal Deep Learning Integration for Predicting and Optimizing Immune Evasi

Browse more like this