MIT 6.S087: Foundation Models & Generative AI. BIOLOGY

goe all right uh well welcome to manolis

he's a professor at MIT he's doing

amazing stuff in computational genomics

and Ai and biology he's going to talk

about the AI Frontiers here super

excited to have so give him a warm

Applause

thank awesome welcome everyone so um

basically there's a lot going on in

biology and there's a lot going on in Ai

and my goal is to tell you a little bit

about both and how the field is

dramatically changing so um uh I'm going

to tell you primarily about health and

understanding biology and Medicine who

wants to live forever here good good who

wants to live at least live like until

next year yeah so so there's this joke

where like like oh who wants to live

till 100 well the person who's 99 and uh

yeah we never actually want to die we

just don't necessarily want to live

forever but anyway so the goal is how

can we use AI to truly understand the

mechanism through which human biology

works and how we can use that to

basically develop new Therapeutics that

put an end to disease as we know it

who's excited about that good so what's

our goal our goal is to understand

medicine and and medicine has truly come

a long way so this is I was just giving

a talk in Athens last week that's when I

made this slide and uh this is how uh

medicine used to work so you would have

some type type of uh it's unclear

whether here you're depicting a God or a

physician but you know even nowadays the

distinction is kind of blurred in many

ways um and then they're doing this

magic and the patient is sort of

subjected to that and you know there's

some peer review Comedia apparently um

but this is how it used to be done and

then eventually we started looking at

things closer and closer and this is

more than 100 years old the um structure

of neurons inside our cortex and our uh

different regions of the brain we can

basically now see the the fact that

we're based on smaller and smaller parts

and the first diagnosis of Alzheimer's

dates to a 100 years ago from Imaging

where we could actually see the plaques

and neurop fiary Tangles that are still

today the definition of Alzheimer's but

something dramatically changed in the

last few years and this something

is human genetics human genetics

basically tells us that there's

something playing a causal role here and

what that allows us to do is start going

Beyond just correlation to causation and

the last part that changed is a lot of

our own work in being able to gather

massive massive amounts of data for

integration so you can basically think

of this as the Next Generation

microscope where instead of gathering

four cells at single cell resolution we

are gathering 2 million cells at single

cell resolution and instead of measuring

whatever we can stain and we can stain

about like 5 10 20 things at best we can

measure the expression of 20,000 genes

for every one of these dots okay so this

is a 20,000 dimensional space with 2

million cells projected down into

something that we humans can visualize

in 2D okay so what are the Paradigm

shifts that are happening the first

paradigm shift is that we're going from

hypothesis driven to data driven instead

of just saying oh we have a very

specific hypothesis like gather a bunch

of data and then we're going to say yes

no answer we now have just massive data

we shoot first and we ask questions

later so we basically have systematic

data sets we're building resources

massive data sharing and really

comprehensive use of biology the second

as I mentioned we're going from

correlation to cation correlation means

the countries that eat more chocolate

also get more noble prices does the

chocolate need a noble prices do they

just buy uh I don't know more chocolate

with prices is it correlation causation

or reverse causation so that's what

epidemiology has always been fuzzy about

whereas with genetics we actually now

understand mechanism we know that this

if if there's a genetic difference then

eventually you can establish causality

and then the last step which is the most

relevant to this class is we're going

from classical data analysis where there

was a different methodology for every

problem we would just like come up with

a question develop a statistical test

and answering and and the humans did all

the thinking to now and where the goal

was to develop very few parameters and

very targeted models to understand so

that we're not overfitting the data

whereas now we're basically saying

billion parameters no problem bring it

on and we're building this Foundation

models that are very often multimodal

where we're learning representations

we're learning hierarchical deep

representations and we're truly

understanding concepts and yielding

insights everybody with me so far so

these are the major shifts

what does that mean that means that we

com we can combine now causality from

genetics and big data to truly

understand the mechanism of disease

genetics that means that we're starting

with causality because we know these

regions have something to do with

disease the problem is that we don't

understand the mechanism and that's

where massive data come in we can

basically say this correlates with

Alzheimer's let's go and find out what

changes in the brain of people that have

this genetic difference or the people

that have Alzheimer's or people that

have environmental exposure and then we

can figure out the specific genes and

proteins that are responsible and then

use those to understand mechanism so

we're gathering this massive data and

that's where the Deep learning comes in

where we can now go from sequence

information to a model that understands

the language of biology understands how

mutations are acting understands how

proteins are folding how chemicals are

resulting into their functions and then

eventually make predictions that we can

validate experimentally and that's

another amazing thing about biology in

society it's very easy to say well maybe

this causes that but intervention would

cost like I don't know billions of

people changing the way that they do

stuff whereas with Biology we can take a

cell and then change a gene and then see

what happens so they're much more

transparent those models everybody with

me awesome so what we need to do is

basically go from Simply there's

something going on genetic here to

here's the circuit here's the genetic

variance the differences in the letters

here are the motifs the sequence

patterns that these letters perturb here

are The Regulators that bind these

motifs here are the control regions or

enhancers and the cell types where they

become active and here are the target

genes that are controlled so effectively

a circuitry and then using that my lab

has basically worked on applying this

type of methodology to a dozen plus

disorders from cardiac disease obesity

cancer Alzheimer's addiction neurod

degeneration pathogenesis schizophrenia

psychosis bipolar Down syndrome autism

PTSD so every aspect of the human body

and the human brain we can now start

studying systematically across dozens of

cell types across hundreds of uh tissues

and millions of cells and hundreds of

individuals we can now start asking how

is the action of disease percolating

through to give you an example I like to

joke that we published a paper in the

New England joural medicine about one

bit of information in the human genome

this is about changing a t into a c and

what we showed is the mechanism through

which we can translate a region of

genetic Association the strongest

association with obesity to a mechanism

that basically tells us how is that