AI in Life Sciences Discovery | Stanford's RAISE Health Symposium 2025

Nice to see everybody. Thank you so much

for coming back from from break. So I

have the I have the privilege of opening

this session. I'm Nicolola Fuzie. I'm a

general manager of Microsoft Research.

Um and it's been a a great uh day and I

feel like I'm going to take a little bit

of a different perspective um compared

to some of the things that have been

said this morning because I'm going to

focus a lot on models. Um and in

particular I'm going to focus on trying

to learn language of biology using

models and give a few examples. So I

guess the the framing for me is um are

we going to see the kind of same things

that we've seen in NLP natural language

processing happen in biology. in the

natural language processing. Uh if you

kind of look at the history of the

field, we went from a lot of people

working in a very linguistics-based

um way like figuring out which features

are important for models to learn from

and designing very bespoke architectures

to architectures that are much more

general in purpose um uh and learn much

better in terms of compute and data. So

they're much uh more suited for the

compute we have and they can ingest more

data more quickly. And in biology, we're

seeing a lot of models and we've been

seeing them over the years that encode a

lot of symmetries and a lot of

invariences that are important to learn

from. But are we going to see kind of a

similar shift to more general purpose

architectures?

And one good case study from from the

work my team does is on proteins. Um, if

you're familiar with proteins, you know,

the the the simplest way to describe

what's going on is, you know, you have

protein sequences and those sequences

determine a structure and that structure

determines in large parts what what the

protein does. And for the longest time,

if you want to design a protein that

does something that has a particular

function, you design the structure and

then you try try to solve for the

sequence that leads to that structure.

And the problem is that we have, as we

saw this morning, we have a few

character um experimentally

characterized structures. Um and we have

a lot of sequences. That's because

sequences are very cheap to acquire by

sequencing. Um and so for instance, you

can go to a puddle of mud in the middle

of nowhere. You can sequence your sample

from that and you're going to get

probably some good proteins. But if you

want to then experimentally characterize

their structure, you need to do X-ray

crystalallography or cryom which is

expensive and slow. And so uh with the

interest of putting ourselves into a

datari environment, we were asking we

asked the question can you go directly

from sequence to function skipping

structure entirely and treating it more

like a

nuisance. Um and that's what we did. We

and we took a pretty general purpose

architecture. So this is a diffusion

model. So, if you're familiar with

diffusion models, those are the things

that generate the good-looking images

you see on the internet that are

synthetically generated in most cases.

And they work by taking uh a set of

pixels that are completely like noise,

noisy looking, and then they iteratively

den noiseise them through a neural

network to look very natural. Um, and we

do the basically the same thing but just

over discrete sequences. So, instead of

having a continuous value of a pixel, we

have a discrete set of amino acids that

we can put in any position of a protein.

Um, and it turns out that if you do

that, you get proteins that kind of

recapitulate the diversity of natural

proteins very very well. So if you start

from an empty protein and you try to

flip and decide where each different

amino acid goes, you get a bunch more

proteins that are natural looking but

entirely synthetic. And then if you

prompt this model with particular

prompts, you can get it to do very

interesting things such as let's say

that you have a functional motive you

want to present. you can design a

scaffold around it using this model. Or

let's say that you want you have a

protein where you like most of it but

you want to edit a particular region.

You can condition on the part you want

to keep and just uh figure out the the

the region you want to edit. Um and then

there is another example which which is

particularly useful which is let's say

you have a set of related proteins that

do kind of the same thing but you want

different versions that have maybe

different properties like more thermally

stable. you can just keep generating

proteins u from those

examples. And the most important thing

about this work though is that it's it's

a general purpose model. um meaning like

there is nothing special um about the

model that we chose um you can it's a

disc other than being discrete but you

can extend it from just operating over

amino acids to working over general

letters and so you can use that to do a

language model like GPT and in fact we

have later models that we are going to

release soon that are based basically on

the same architecture and the point is

that once you move from models that bake

a lot of knowledge and a lot of

structure structure about biology to

models that are much more in general in

purpose. You actually do not have to

have models that are separate for every

different domain of biology you want to

model. So you don't have to have a

protein model or a single cell RNA seek

model or a pathology model being

different models. Once the architecture

is general enough, you can combine all

of these models under the same

architecture. Um and that's what's

happening in in the field. And and so I

just leave with this slide.

What's happening in the life sciences is

a gradual shift that started many many

years ago. So many many years ago we

used to have um basically one model per

data set. You were the expert in the

data set you acquired you baked all of

your knowledge into the architecture you

trained on it and then you got a good

model. More recently, we shifted to kind

of the era of foundation models where we

put a lot more data and a lot less

structure about the data into the model

and then we shifted to fine-tuning that

foundation model on our particular slice

of the data to make it work better. And

I argue that where we're going towards

is is a is a future where we have one

kind of general model architecture way

of modeling like we've seen in NLP with

the GPT kind of framework. Um, and

learning is not going to happen by

moving necessarily weights around in

weight space by fine-tuning or training,

but rather it's going to happen through

in context learning, which is what we're

seeing a lot in language. Um, and with

that, I think I'm almost out of time.

And so I'm going to introduce Brian up

to the stage. Thank

you. Hi everyone. Uh, so my name is

Brian He. uh I'm a faculty member here

at Stanford and chem and data science

and uh excited to share some of the work

that we've been doing on developing uh

genomic language models across all

domains of life. So all of you are

probably aware that machine learning has

revolutionized the study of molecules.

So for example, methods for protein

structure prediction like alphafold uh

methods for denovo protein design like

RF diffusion were awarded the Nobel

Prize in chemistry last year and this is

very welld deserved. You know molecules

are very complicated. We're still very

interested in molecules. But what we've

been working on is we're also interested

in how these molecules come together to

form say a complete uh organism. uh but

even you know a very simple

single-sellled organism is orders of