24.03 Liontrust Global Innovation Report - The Rise of AI 04.24 - Flipbook - Page 30
MODERNA
ANDREW GIESSEL, DIRECTOR OF AI AND DATA SCIENCE
Dr. Andrew Giessel is Director of AI and Data Science at Moderna, where
he oversees the team applying machine learning and statistical techniques to
improve the company’s mRNA medicines, including the Covid-19 vaccine.
Prior to Moderna, Andrew served as Director of Data Science at Sense
AI, a startup developing a platform for geospatial mobile sensor data. He
holds a PhD in Neuroscience from Harvard Medical School and a BS in
Biochemistry & Computer Science from the University of Kansas
Where in the healthcare industry is the greatest
potential for AI to accelerate innovation and create
customer value?
The space of potential opportunities is so big.
Traditional, more physics-based approaches to problem
solving require a good understanding of the rules of the system. But
often the things that matter in biology are combinations of different
factors that can be very hard to describe, making it difficult to come
up with a simple set of rules for the system. AI and Machine Learning
(ML) allow for a different, more data-driven approach, helping make
sense of such a complicated domain.
understood. The traditional approach is to use a combination of rules
and heuristics, which can get you quite far, but we are increasingly
seeing benefit in pursuing a more data-driven approach.
The impact that LLMs are having is incredible and substantial,
especially on the research side of things where we have been working
on trying to understand the relationship between the sequences we
use for our mRNA and how effective they are in the body. This is a
huge, dual level problem because mRNA encodes for proteins and
proteins are the molecule that your body then makes, leading to a
therapeutic effect in a person.
As a platform company, this is compelling. If we make a model that is
general enough to predict mRNA stability, for example, we can then
reuse it for the next mRNA we make. These models generate better
data, which can be used to improve the model further, in a virtuous
circle. And this is why we are leaning into this approach, as these
models can identify complicated relationships that would otherwise
be difficult to capture. This is helping overcome fundamental design
problems and accelerating the drug discovery process.
There’s LLM-driven optimisation in picking the right protein sequence,
which is important as this determines effectiveness. In the case of a
vaccine, you’re picking a protein sequence for part of a virus – but
as the virus evolves, the protein changes, so if you have picked
the wrong protein then the vaccine won’t be as effective. There’s
also optimisation in picking the right mRNA sequence. There are
effectively an infinite number of mRNAs that could code for a given
protein, and these can vary in terms of the amount of protein they
make and how stable they are. It is hard to approach this process
from first principles as these are difficult rules that are not yet well
30 - The rise of AI: Technology and Innovation Report
This starts with “experimental screening”: testing a range of variants,
applying rules to generate examples, and then testing these. This
generates experimental data that is used to build a model, with
the input being the exact mRNA sequence and the output being
performance. This modelling allows us to repeat the process multiple
times to pre-screen further rounds of possibilities, helping constrain
the design space which is mindbogglingly large.
What is the rate of progress in AI that you are seeing
currently, and how should we expect the healthcare
industry to transform going forward?
Modelling approaches have been around in fields like
small molecule drug discovery for a while. It’s a very
similar concept – instead of designing a protein via mRNA, you’re
screening many small molecule compounds. What’s newer and more
revolutionary is leveraging LLMs that can understand and generate
natural language.
