Forget Moonshots Biomedicine Needs an Air Traffic Control System - - PowerPoint PPT Presentation

Forget Moonshots Biomedicine Needs an Air Traffic Control System

Jeff Shrager

Cancer Commons xCures Stanford Symbolic Systems Program (Adjunct)

Hard AI: Cancer

Dynamic “rules” (biology doesn’t change but everyt rythi hing ng else e does) s) Physica sical l simula latio tions ns is is extremely emely expensiv sive, , and every y experime iment nt kills s people le or anima mals ls in horrible le ways! Simula ulati tion

• n is essent

entiall ally y imposs ssible e (The immun une e system stem is as comple lex x as the brain!**) n!**) Data is essen entia tially lly non-exi xiste stent Feedbac back k can take e years s and is very y noisy; y; There e are NO EXPE PERTS! TS!

Easy AI: Self-Driving Cars

Smal all, l, local decisi sion

• n environ
• nment

ent, in both space ce and time Mostl tly y static tic, , mostly tly well-und nder ersto stood

• d

rules s and principl iples es Comput uter er simula latio tion is nearly ly trivia ial Data is plentiful tiful Expert t guidan ance ce is is insta tant ntaneo aneous us, , cheap, , and nearly ly perfect ct Physica sical l simula latio tion n is easy Extremel emely y broad d decisio sion n environ

• nment,

ment, in both space ce and time

** Immune system: l trillion T cells, l trillion B cells, all circulating 50x/day, plus 10 billion antigen-presenting cells. Human Brain: 100 billion neurons, trillions of synapses, and 1 billion glial cells. And it pretty much doesn’t move. AND BOTH LEARN!

Global Cumulative/Coordinated/Continuous Treatment Analysis

Why is Cancer so Hard?

We’re treating an extremely high dimensionality, low data density, problem the same way that ants search for food!

Phenotypes: Lung, Breast, ... Treatments: Chemo1, Chemo2, ... ~1 Million patients/year ~100 cells =~10,000 patients/cell Plenty for Classical Clinical Trials

pre-OMIC OMIC era: Tissue sue x Chemo mos

Millions of molecular features Thousands of drugs in combination

Now: w: Featur ures s x Targeted ted Comb mbos

~1 Million patients/year ~11Z cells =~0 patients/cell Need a new paradigm (and “big data” won’t cut it!)

This is the State-of-the-Art Statistical Model for Adaptive Trials: And this is the State-of-the-Art Algorithm:

GCTA is a redesign of clinical research in the image of a “Learning Air Traffic Control System”, where entropy- minimizing reinforcement learning routes hypotheses in real time to the patients who at the same time have the most to gain, and the most information to offer to the overall system.

GCTA is NOT a fantasy! The VA just hasn’t put AI Engineers on the problem yet!