Ultra-Fast Combinatorial Analysis for Biomarker Discovery

Rocketml, Inc., Lake Oswego OR

Investigators

Abstract

Project Summary/Abstract Significance: One of healthcare?s largest challenges is that many patients do not respond to treatment. By recent estimates, around 90% of drugs are effective for less than 50% of patients. This causes enormous physical, social, and economic suffering. The annual cost of ineffective treatment is estimated at $350 billion per year in the US alone. Moreover, variable treatment response contributes to the rising cost of drug development, currently around $2.6 billion per drug. The fundamental reasons that treatment responses vary is that there are many different cell types in people, and cell types vary person-to-person. In the computational world this problem of complexity is called ?dimensionality curse?. To avoid the enormous computational burden of high dimensionality, researchers developed deep learning and dimensionality reduction tools like T-SNE, PCA. However in the process of solving computational challenges, crucial biological interpretability is lost leading to a bigger issue: reliability of these methods. Hypothesis : We hypothesized that faster combinatorics algorithms combined with massive parallel computing techniques would make computation solutions for single cell high dimensional data tractable. Preliminary Data: We have demonstrated our approach/algorithm on a publicly available AIDS dataset. Our method produced biologically interpretable, accurate results at an ultra-fast speed (1000x faster vs. current methods) making the approach tractable. The approach was further validated for biological significance and accuracy through collaboration with an NYU Medical Center researcher, who has now joined the team as Chief Scientific Advisor. Specific Aims: This project entails Aim 1 - Building a production grade, computationally efficient and scalable system for creation and storage of all combinatorial phenotypic signatures, including frequency of each cell type, by patient, with associated metadata. Aim 2 - Integrating an automated system for comparing phenotypic frequencies across patient groups, incorporating standard biostatistical analyses and machine learning-based modeling. Aim 3 - Applying our new tool to multiple existing, high-parameter cancer and infectious disease datasets to demonstrate the value of combinatorics over dimension reduction for accurately predicting clinical outcomes at ultra-fast speed. Together, these studies will demonstrate a new computational approach to high dimensional single cell data without losing biological information or interpretability at an ultra-fast speed, revolutionizing precision medicine.