How will AI transform precision medicine? – Ava Amini

Summary

The speaker from Microsoft Research shares a vision for AI-driven precision health, emphasizing the uniqueness of human biology and the need for personalized medicine. They highlight the limitations of current cancer treatments and propose leveraging AI to understand the complexity of biological data, collaborating with the Broad Institute for precision oncology. The presentation introduces 'evoi', a generative AI for designing novel protein therapies, showcasing its ability to create proteins with new functions, aiming to revolutionize treatment by personalizing discovery and experimental design.

Takeaways

• 🤖 The speaker emphasizes the potential of AI in revolutionizing biological discovery and precision health by leveraging the unique genetic makeup of individuals.
• 🧬 Human health is described as both robust and delicate, with a call for personalized medicine that takes into account the nuances of each person's biology.
• 🔍 Current medical practices are criticized for not fully utilizing the rich and complex biological data available, often treating patients based on population averages rather than individual differences.
• 💡 The speaker introduces a new vision for precision medicine using AI, aimed at understanding disease mechanisms and designing new treatments tailored to individual patients.
• 🧬 The script highlights the opportunity to 'unlock the language of biology' by using AI to interpret the vast amount of data stored within each person's biological makeup.
• 🧬 The speaker shares a concrete example of how AI can be applied to cancer treatment, noting the current limitations in targeted therapies and the need for more personalized approaches.
• 📈 The potential of AI is underscored by the ability to analyze nearly 50 million data points from a single tumor biopsy, offering a rich dataset for developing new AI models.
• 🤝 Collaboration with the Broad Institute of MIT and Harvard is mentioned to develop AI systems for precision oncology, integrating molecular measurements into AI training for personalized treatment recommendations.
• 🛠️ The speaker introduces 'evoi', a generative AI model developed by Microsoft Research, designed to create new protein sequences with potential therapeutic applications.
• 🔬 The script discusses the importance of reasoning across different scales and types of data to fully understand and utilize the language of biology for developing new treatments.
• 🔁 The concept of a 'data and learning flywheel' is presented as a method to continuously improve AI models by integrating new data from lab experiments, aiming to close the loop on precision medicine.

Q & A

What is the main vision being discussed in the script?

The main vision discussed is building AI systems to enable and accelerate new biological discoveries towards a vision of Precision Health, aiming to personalize medicine and better understand the mechanisms of disease.

Why is it important to consider individual differences in medicine?

Individual differences are important because each person has a unique genetic makeup and history of experiences. Current medicine often does not account for these nuances, leading to less effective treatments that are not personalized.

What is the current state of targeted therapies for cancer patients?

Fewer than 40% of cancer patients have access to targeted therapies that are specific to changes in their cells. Of those, less than 5% respond effectively to the current treatments, indicating a need for improvement.

How does the speaker define the problem with the current approach to treating cancer?

The speaker defines the problem as treating cancer in an ad hoc, reductionist way where the individual's richness and complexity are lost, often resulting in treatments that work on average but not personalized.

What is the scale of biological data that can be generated from a single tumor biopsy?

From a single tumor biopsy, nearly 50 million individual data points can be generated, considering various levels of biological resolution such as DNA mutations, transcriptome, proteins, and cell interactions.

What is the opportunity presented by this scale of biological data for AI researchers?

The opportunity is to develop new AI systems that can process and analyze this vast amount of data to unlock new biological insights, design new treatments, and improve personalized medicine.

What is the role of AI in the vision of precision oncology discussed in the script?

AI is central to the vision of precision oncology, where it is used to analyze molecular measurements from patients, train AI systems to produce personalized treatment recommendations, and design new molecular therapies.

What is the generative AI model 'evoi' mentioned in the script, and what does it aim to do?

Evoi is a new generative AI model developed to design new molecular therapies, such as proteins, by learning from large-scale evolutionary data sets. It aims to produce proteins with new functions that have never been seen before in nature.

How does the AI model 'evoi' utilize the functional part of a protein to design new proteins?

Evoi uses the functional part of a protein as a prompt, learning from its training data to design a new protein sequence around that part, creating a brand new protein that can be linked to a specific biological function.

What is the ultimate goal of developing these AI systems in the context of precision medicine?

The ultimate goal is to close the loop on precision medicine by generating a data and learning flywheel that can personalize discovery, experimental design, and ultimately lead to more effective, tailored treatments for patients.

Outlines

00:00

🧬 Embracing AI for Personalized Medicine

The speaker expresses gratitude for the opportunity to discuss the integration of AI in advancing biological discoveries towards precision health. They emphasize the uniqueness of human health, which is both robust and delicate, and the current inadequacy of medicine to account for individual differences. The speaker argues for a new approach to unlock biological discoveries using AI, particularly in understanding and treating diseases like cancer. They highlight the need for AI systems trained not just on human language but also on the 'language of biology,' which is rich in data due to evolutionary processes. The speaker introduces Microsoft Research's vision to develop AI for personalized medicine, focusing on understanding disease mechanisms and designing new treatments, with reliability and safety as priorities. The example of cancer is used to illustrate the current challenges in targeted therapies and the potential for improvement through AI-driven personalized approaches.

05:03

🛠️ Innovating with AI in Precision Oncology and Molecular Therapy Design

The speaker shares their excitement about Microsoft Research's collaboration with the Broad Institute to create a new vision for precision oncology, utilizing AI to tailor treatments to individual patients. They describe the process of taking molecular measurements from patients to train AI systems that can recommend personalized experiments and treatments. The speaker introduces 'evoi,' a new generative AI model capable of designing novel proteins with unprecedented sequences, expanding the range of functionalities for chemical, biological, and therapeutic applications. An example is provided where the AI model is trained on a specific protein function and then designs a new protein sequence from scratch, which can be linked to a measurable biological function. The speaker concludes by emphasizing the importance of reasoning across different scales and data types to create a comprehensive understanding of biology, which is essential for realizing the vision of precision medicine and more effective treatments.

Keywords

💡AI Systems

AI Systems, or Artificial Intelligence Systems, refer to the computational models and algorithms designed to perform tasks that typically require human intelligence. In the context of the video, AI Systems are envisioned to accelerate biological discoveries and contribute to the field of Precision Health by analyzing complex biological data and aiding in personalized medicine.

💡Precision Health

Precision Health is an approach to healthcare that tailors medical decisions, practices, and/or products to individual variability in genes, environment, and lifestyle. The video discusses the vision of using AI to enable Precision Health by understanding and leveraging the unique biological characteristics of each individual for better health outcomes.

💡Genetic Makeup

Genetic Makeup refers to the complete set of genes in an individual organism, which determines many of its characteristics. The script emphasizes the uniqueness of each person's genetic makeup and how it contributes to the complexity of human health, underscoring the need for personalized approaches in medicine.

💡Biological Discovery

Biological Discovery pertains to the identification of new biological processes, mechanisms, or entities that can lead to advancements in medicine and health. The video script discusses the role of AI in unlocking new biological discoveries by learning patterns that differentiate individuals and applying this knowledge to develop personalized treatments.

💡Personalized Medicine

Personalized Medicine is an approach to medical treatment that tailors therapy to the individual characteristics of each patient. The video highlights the goal of using AI to develop personalized medicine, which takes into account the unique genetic and biological profile of a patient to deliver more effective treatments.

💡Cancer

Cancer is a term used to describe a group of diseases characterized by the uncontrolled growth and spread of abnormal cells. In the script, cancer is presented as a use case to illustrate the importance of personalized medicine, given the vast differences in how individuals respond to treatments.

💡Targeted Therapies

Targeted Therapies are a type of cancer treatment that is designed to target the specific genetic mutations or proteins within cancer cells. The video script points out that even though targeted therapies exist for some cancer patients, the effectiveness is limited, emphasizing the need for more personalized approaches.

💡Biological Data

Biological Data refers to the information obtained from biological systems, such as genetic sequences, protein structures, and cellular interactions. The script discusses the vast scale and complexity of biological data, which AI systems aim to analyze and interpret to advance our understanding of diseases and treatments.

💡Evolution

Evolution is the process by which different kinds of living organisms develop and diversify from earlier forms during the history of the Earth. The video mentions the role of evolution in creating the rich and complex biological data within each individual, which AI can learn from to understand the 'language of biology'.

💡Generative AI

Generative AI refers to AI models that can create new content, such as text, images, or protein sequences, that follow the patterns they have learned from existing data. In the script, a generative AI model called 'evoi' is mentioned, which is capable of designing new proteins with potential therapeutic applications.

💡Protein Function

Protein Function refers to the specific role or activity performed by a protein in a biological system. The video script describes how AI can learn the language of protein function by analyzing large datasets of protein sequences, enabling the design of new proteins with novel therapeutic capabilities.

Highlights

The presentation discusses the vision of building AI systems to accelerate biological discovery towards precision health.

The uniqueness of human health and life, being both robust and delicate, motivates the work on personalized medicine.

Current medicine largely fails to account for the nuances and complexities of individual health.

AI can unlock new biological discoveries by learning the language of biology, which is rich in data.

Microsoft Research aims to develop AI for personalized medicine to better understand disease mechanisms and design treatments.

Cancer as a use case highlights the need for precision in treatment based on individual biological changes.

Less than 5% of cancer patients respond effectively to current treatments, indicating a significant problem in the approach.

The reductionist approach to treating cancer overlooks the individual's biological complexity.

A single tumor biopsy can generate nearly 50 million data points, presenting an opportunity for AI.

Microsoft Research collaborates with the Broad Institute to create AI for precision oncology.

AI systems are being developed to produce personalized recommendations for patient-specific experiments.

A new generative AI model, Evo, is introduced to design new molecular therapies like proteins.

Evo learns from large-scale datasets to produce proteins with expanded functional capabilities.

An example demonstrates how Evo can design a new protein sequence based on a functional part.

The designed protein by AI shows a structure linked to measurable biological function.

The need for reasoning across scales and integrating various data types to understand biology fully.

The vision of closing the loop in precision medicine by generating a data and learning flywheel.

The goal is to realize more effective, personalized treatments through AI and data integration.

Transcripts

thank you so much Peter it's an absolute

honor to be here with you all to share

with you our vision of how we can build

AI systems that can enable and

accelerate new biological Discovery

towards a vision of Precision

Health our work is fundamentally

motivated by the fact that human health

and human life is a Marvel on one hand

incredibly robust and on the other hand

incredibly delicate and to me the most

amazing thing about all this is that

each and every one of us is unique we

all have different bodies different

genetic makeups a different history of

experiences that makes you you and me me

and yet somehow it all seems to work in

concert and yet shockingly in medicine

today the fact is that these Nuance

differences these richnesses and

complexities are largely not taken into

account to preserve our health in a way

that's truly personalized

we don't just need a better way to

detect and treat disease or discover new

drugs I'd argue that we need a whole new

way to think about how we unlock new

biological discoveries and learn those

patterns that differentiate us all and

deploy them

forward through the use of powerful

tools like AI now it's no secret that

today we're experiencing a revolution

with a new generation of AI systems

train trained on Words and text the

language of us as

humans but yet there is tremendous

opportunity to unlock and learn the

language of

biology because of the fact that biology

through the hand of evolution naturally

stores an incredible scale richness and

complexity of data within each and every

one of

us our vision at Microsoft research is

to leverage this opportunity to now

develop new AI towards the vision of

personalizing medicine to help us

discover the mechanisms of disease

better to design new treatments and

ultimately deploy these systems into the

real world in a way that's reliable and

safe let me start by showing you why

this matters and a concrete example of

why we should

care let's consider cancer as a use case

in the United States nearly 40% of the

population would will develop cancer in

their lifetime and while we've known for

a while that cancer is driven

fundamentally By changes in the biology

of our cells we yet don't understand how

to condition the therapies based on

those changes to deliver the right

treatment to the right patient at the

right time indeed for fewer than 40% of

cancer patients today there exists what

we call targeted therapies that are

therapies specific to changes in their

cells but what's even more shocking is

of those 40% less than 5% of patients

even respond effectively to today's

treatments so clearly there's a problem

and clearly we can do better but why are

we

failing the answer to that is that

because by and large cancer is treated

in this ad hoc reductionist way where

that richness and complexity of the

individual is lost because they're

viewed as rather an instance in a

population given a treatment that works

on average over large scale population

wise

studies and so I've been speaking a lot

about this richness this complexity this

scale of biological data what exactly do

I mean by this and how can we make this

concrete I'd like you to consider now

suppose we have one cancer patient and

we take a biopsy from one tumor from

that patient

at just this level thinking about the

scales of biological resolution from the

level of DNA mutations to the

transcriptome to proteins to the

interactions of cells in their

neighborhoods and how they communicate

with each other we can generate nearly

50 million individual data points from

just the single tumor biopsy using our

measurement and experimental techniques

that we have in the lab

today now let's pause for a second and

consider this number right when I look

at this as an AI researcher and as a

computer scientist I see tremendous

opportunity and if I look at this number

as a

biologist I see Power that comes at the

hand of this richness and

complexity what's beautiful is we can

put those together by using this ability

to directly measure biology at its

natural scale the Nano scale as it's

occurred in in real time this gives us

opportunity to unlock the development of

new Ai and today I'm very excited to

share with you two concrete ways in

which we at Microsoft research have

brought this to

life first in collaboration and in close

partnership with the broad Institute of

MIT and Harvard we're collaborating to

create a new vision for precision

oncology closing the loop on this Vision

that we been talking about by putting AI

front and Cent and Center such that we

can directly tail take these molecular

me measurements from the patient level

train and build AI systems that can now

produce recommendations about what

experiments are best to test for that

patient in a personalized way we can use

this as a flywheel to generate the new

data that we produce in the lab to then

iterate and improve our AI model and and

ultimately recommend or design new

treatments that are specific and

tailored to the needs and biology of

that

person we can dive deeper on one of the

components of this Pipeline and ask what

does it take to actually design new

molecular therapies new treatments that

have functional

ability and to this end we've built and

are building powerful new generative AI

systems to design new molecular

Therapies like

proteins by learning the language of

protein function learning from large

scale evolutionary scale data sets of 50

million unique protein

sequences we've trained a new generative

AI model that we call evoi that can now

produce brand new instances of proteins

that have never been seen before in

nature with the goal of expanding the

functional capabilities whether they be

chemical biological or therapeutic

available to

us let's bring this capability this

generative capability to life through an

example what I'm showing you here is an

example of one protein where in green

I've isolated the functional part of

that protein that is responsible for its

function binding to calcium in our cells

we can take that part and isolate it and

use this as a kind of prompt to our

model Evo diff which can then take that

prompt and learn from the information

that it's been trained on and see to now

design a brand new protein sequence

around that part step by step from the

bottom up creating a brand new protein

sequence that's never existed

before and what's amazing is that this

new protein designed by our AI model

shows a structure that we can then

explicitly link to biological function

measured in the lab in the real world

now I'd be remiss to say that it stops

there at our ability to generate new

molecules and proteins that's just the

start to actually learn and understand

this language of biology we need to be

able to reason across scales across

different types of data and put all

these components together from the

molecule level all the way up to that of

the

patient we see this as a foundation by

which we can now close the Loop and

generate a data and learning flywheel to

close this vision of precision medicine

for once and for all to be able to

personalize Discovery experimental

design and ultimately hopefully realize

more effective treatments thank you

[Applause]