Precision Medicine: Where Small-Sample Clinical Trials Meet Big Health Data

Overview

Motivating case: NSCLC

 

Precision therapy challenges

 

Incorporating larger data sets for drug/population discovery

 

FOCUS ON TARGETED CANCER THERAPY

Subject Profile - 46 Yr Old Male NSCLC

- First diagnosed in November 2013

- 30 pack-year smoking history

- Required supplemental O2

- Significant pain and dyspnea due to widely metastatic disease

- Staging head CT also revealed numerous (15 to 20) asymptomatic brain metastases measuring up to 1.7 cm that had not been previously treated

- In hospice

Ignyta Inc. Entrectinib Study: Before

Ignyta Inc. Entrectinib Study: After

This example May extreme but shows what is possible

Precision Therapy Challenges

Traditional:

  - Develop the drugs

  - Evaluating dose

 

Current challenges:

 - Identifying respsonsive subpopulations 

 - Powering efficacy trials with smaller sample sizes

 

Future challenges:

  - Transitioning to a residual disease model

IDENTIFYING RESPSONSIVE POPULATIONS

Pre-clinical: cell line Experiments

  - grow cancer cells and expose them to a potential therapy

  - Measure half the maximal inhibitory concentration (ic50) - how much do I need to affect biological function?

 

Post-Trial: Subgroup analysis

Cell line experiments create data

REgressors are Biomarkers - Mutation, expression level, etc.

 

IC50 level is the response - Can we affect function with relatively low dose?

REgressions are high-dimensional

How do we prioritize investigations of HUndreds of samples and tens of thousands biomarkers per sample?

 

How do we define "response" vs. "non-response" when ic50 is a continuous measure?

 

How do we resolve results when data-driven analyses give one result and clinical investigators hypothesize another?

How can this be scaled to more markers and models?

PIRLS: Penalized iteratively reweighted least squares

 

STILL IN DEVELOPMENT BUT AVAILABLE ON GITHUB AT HTTPS://GITHUB.COM/KANEPLUSPLUS/PIRLS

 

Models supported:

  - OLS, GLM, Ridge, LASSO, Elastic net

  - need a little more infrastructure for "m" regressions

 

Plan for out-of-core

Going further: the interactome

Biomarkers are data summaries

  - They are derived measurements from the nucleic acid sequence

  - Regressions do not take into account the relationships between biomarkers

 

Create a graphical model from the correlation matrix (preprint here)

  - Responsive subpopulations for a given therapy have shared structural differences in their interactomes when compared to a reference group

  - Structural differences correspond to the genetic profiles of responders

  - We derive specialized biomarkers for individual therapy/genome combinations

Interactome challenges

How can we quantify structural differences between interactomes?

 

How should we incorporate pathway information?

Thanks

Slides are available at http://slides.com/michaelkane/deck-14

 

My email is michael dot kane the at symbol yale dot edu.