GENomics-AI: Are We Moving Too Fast?

00:01:08
Streeter: It was probably the weather, or where can I get free stuff on my birthday?

00:01:15
Dr. Kaylee Byers: I’ll admit, I’ve used it too. You look in your fridge, you say, “What can
I make with all these ingredients?” And ChatGPT says this thing, and you go, “I’ll make
that.”

00:01:23
Clip: Bon appétit.

00:01:25
Dr. Kaylee Byers: But AI isn’t just changing the way we go through our daily routines,
it’s changing what our world could look like. And while the potential of AI in genomics is
exciting in some ways, it also raises some questions, like, can we trust AI with the keys
to our genetic code? And how do we keep our sci-fi thriller plots from becoming real life
science fact? You’re listening to Nice Genes, the podcast all about the genomic stories
shaping our world, brought to you by Genome British Columbia. I’m your host, Dr.
Kaylee Byers, your human verified guide on this scientific mission. Most of us have
probably had an experience with an AI model where we ask a question and the answer
comes back sounding very convincing, but maybe not entirely accurate. Generative AI
can be useful, but it can also make mistakes, reflect biases, or confidently give us
information that isn’t quite right. So, what happens when we start using these tools for
something as consequential as understanding and even editing our genes?

00:02:34
Dr. Le Cong: My name’s Le and-

00:02:35
Dr. Kaylee Byers: That’s Dr. Le Cong.

00:02:36
Dr. Le Cong: Currently, I’m a professor at Stanford University School of Medicine.

00:02:40
Dr. Kaylee Byers: And he’s also one of the co-inventors of CRISPR gene-editing
technology. Hi, Dr. Cong, thanks for taking the time to chat with me today. So, to start
off, what drew you to this intersection of AI and genomics and health in the first place?

00:02:58
Dr. Le Cong: Yeah, I grew up in the 80s and 90s, and grew up playing computer games
a lot. And I was initially interested in becoming a game designer when I was a kid, but
then I got excited when, and many of us know, there was the global efforts on
sequencing the human genome. So, I got fascinated by the idea that we can now
decode human genome, and maybe instead of designing a game, we can design new
drugs, new medicines using a combination of both genes and genome, and then also
from computer science, because this I think is one of the most exciting intersection
between very human topic, which is genes and genome, and a very technology topic,
which is machine learning AI.

00:03:48
Dr. Kaylee Byers: Okay, let’s talk about CRISPR. So, one of the most major
developments in the history of genomics, and we’ve talked a little bit about CRISPR on
the show before, as this tool that can go in and edit your DNA, they’re like letters or
words, and CRISPR allows you to edit those words. But one thing we really haven’t
talked about is what’s actually going into using it in the lab. So, how simple is it to use?

00:04:12
Dr. Le Cong: Yes, there’s a lot of reasons why CRISPR is so powerful and the number
one reason is it’s very simple to use. So, when we are actually making a change to the
genome, the only thing you need is basically what sentence you’re looking to change.
You type that sentence into the search box in a word processor, and then you will
quickly locate the sentence, and you can make changes. So, very simple. And that’s
why people, even though there are some other tools, CRISPR is really the most widely
used one.

00:04:43
Dr. Kaylee Byers: And up until recently, how much specialized knowledge would
scientists need to be able to use CRISPR?

00:04:49
Dr. Le Cong: So, I’ll give you a specific example. Before CRISPR, to do a single gene
change for a PhD student, you need to spend probably something like $ 20,000 just to
make a single edit. And then, usually it takes a few weeks or a month, and then
oftentimes it doesn’t work, more than half probably. With CRISPR today, you just need

$20, $30 to order a CRISPR system online, and then you can basically get it usually
within a week and then most of the time it works. But even though you can easily order
and get a CRISPR system that targets a specific gene, what is much harder is to have
all the equipment, all the expertise to actually use CRISPR to basically run the
experiments. And this is I think the barrier. Where, buying a CRISPR is easy, but
actually using it in the real lab to use it in a cell, these are expertise that needs a lot of
training.

00:05:51
Dr. Kaylee Byers: Well, where does AI come into all of this? So, you’ve worked on this
CRISPR-GPT, at the highest level, what is that doing?

00:05:59
Dr. Le Cong: CRISPR-GPT is like any AI system. We have AI systems that help you
write your email or help you summarize a book. What we’ve found that even though
CRISPR itself is very low cost, very fast today, what is really hard and expensive is
actually before and after. Before meaning… For example, I’m studying cancer, which is
actually one of our collaboration projects. And what they want to do is to specifically
target some of the genes that are mutated in lung cancer patients. But then the problem
is there are many genes, and every gene in the human genome has over sometimes
30,000 bases, you can imagine 30, 000 letters within that chapter. But in order for us to
really understand cancer, we need to decide which part of that long chapter of text we
should be making changes to.
So, using CRISPR-GPT, you can basically type… Let’s say, I’m interested in studying
patient mutations in lung cancer. What are good potential genes that I should be using
CRISPR to engineer? You can type in natural language. And why I think people are
excited about CRISPR-GPT is that in the past, before we have AI, what you can do is
find maybe a lung cancer doctor and to schedule a call with them, and then having a
call, and explain to a physician, a lung cancer expert what I’m interested. And I have to
go through all these processes, even probably explaining what is CRISPR to them. But
now I have this virtual expert of CRISPR, I can just talk to this AI assistant. And so, this
is the part that CRISPR-GPT works out in the beginning, which is to help you with
design experiment. So, CRISPR-GPT will actually help me to pick the step-by-step
instructions of how I could use CRISPR, and finally, it can also help you analyze the
data. So, basically, even though CRISPR itself is a very good technology, that’s very
low cost, very simple and fast, all the steps that before and after is actually very
important for people to really have expertise. And CRISPR-GPT basically gives you the
expertise, and this is what we think are really exciting to us.

00:08:15
Dr. Kaylee Byers: So, what makes CRISPR-GPT different from simply putting a
question into ChatGPT and asking it sort of a CRISPR question?

00:08:26
Dr. Le Cong: So, actually we started the project by thinking, oh, maybe we’ll just use
ChatGPT, right? But what happens when my students and myself test it, ChatGPT will
give you oftentimes the answer, but sometimes it hallucinates. It even makes up an
answer that looks correct. We actually asked ChatGPT to design a CRISPR. It
answered very confidently, but then when we search for design, it’s a fake design. The
general purpose ChatGPT is not reliable enough in science, at least today. And two, in
order for us to make an AI that hopefully is more reliable, what we did is we take about
11 years of scientific discussion and trained an AI model across over 2, 000 scientists
around the world. And then, all this knowledge helps to train the AI to really understand
how scientists are actually using CRISPR in the real world. And this model actually,
when we tested it, performs better than any AI model we tested from OpenAI, from
Anthropic. So, this is how we think that AI can truly understand, for example, biology
and medicine is by treating them with real scientific knowledge from human experts. But
again, we still are testing this, I think there’s still limitations, but the nice thing is if you
have enough expert knowledge, I think we can actually start teaching AI how really
science should be done.

00:09:59
Dr. Kaylee Byers: So, basically think of CRISPR-GPT as an AI research assistant for
gene editing. It can help scientists figure out what to edit, how to edit it, and then make
sense of what happened afterward. And that could be a pretty big deal, especially when
it comes to medicine.

00:10:17
Dr. Le Cong: So, we have been working with Dr. Matthew Porteus, who’s a physician at
Stanford who have been working on rare pediatric disease treatment. So, his group
wanted to treat kids who have this rare disease where because they have a genetic
mutation in their blood stem cell, they cannot have proper immune response. So, these
kids will be born with very little ability to defend against infection. Every one of these
kids have different mutations. This is why it’s very hard to design a single drug that treat
all of them. But now you say all this real scientific knowledge from human experts
turned into an AI that can quickly design a CRISPR for all these 100 kids with different
mutation, if a human does it, takes a week. CRISPR-GPT can do it within an hour. So,
that becomes useful because this is one, a rare disease, so it’s not easily addressed by
existing human efforts. Human efforts is not sufficient. And two, we have enough

knowledge to train an AI that can hopefully at least provide an initial draft that human
can review and make decisions. We actually have tested in four different mutations and
it works really well. So, this is a long way of saying, I think we have to check multiple
boxes in order to know, oh, this is an AI system that should be used and is producing
value.

00:11:46
Dr. Kaylee Byers: This might be familiar if you tuned in last time when we talked about
a new gene editing tool called Prime Editing. The major dilemma? Even though this
technology could help treat all kinds of rare diseases, each one can require its own
unique treatment. So, could AI and systems like CRISPR-GPT be part of the answer?
It’s definitely promising, but for all that promise, does the efficiency also come at a cost?
Something that you’ve mentioned, part of this process that gets shorter is this beginning
and planning phase, but isn’t having the expert also really valuable to the process of
understanding the problem and the interpretation? Sometimes is there a quality check
on that information somewhere? You know what I mean, so that you are getting that
trained expertise?

00:12:39
Dr. Le Cong: Yeah, I think this is a great question. What we’re realizing is all the AI
systems is very powerful today, but at the same time, like you said, what is the way we
should be using AI systems in science and even other type of task? We think that we’ve
been promoting the idea that AI is always going to work with humans in the lead. And
there are a lot of people saying that AI system is so powerful that maybe we can
outsource or can just let AI do a lot of work.
I think for some type of work, maybe, but I think for scientific work, for work that have,
number one, impact to human, two, requires really human oversight, I think we have to
really have humans be in the lead where key decisions are made by human, and also
the final outcomes, final designs should be checked by human rather than the AI
making the decision. And I think this is one way that we think that AI should be used in
science, which is human in the lead, but I do think that there’s always a question of, how
do we really have almost like a standard? I think this is a really new topic, I don’t know if
we have an answer, but I think we should have a standard or a discussion of this.

00:13:57
Dr. Kaylee Byers: And hold onto your butts.

00:13:59
Clip: Hold on to your butts.

00:14:00
Dr. Kaylee Byers: Because that discussion we will have.

00:14:04
Dr. Mohammad Hosseini: So instead of science investigating things, we bring science
to the investigation table.

00:14:09
Dr. Kaylee Byers: Coming up, what happens when technology moves faster than the
ethical conversations around it? You are listening to Nice Genes, a podcast all about
the fascinating world of genomics and the evolving science behind it, brought to you by
Genome British Columbia. I’m your host, Dr. Kaylee Byers, and we want to get more
people to listen to the genomic stories that are shaping our world. So, if you like Nice
Genes, hit follow on Apple Podcasts or wherever you get your shows. If this episode
has your circuits firing, leave us a review and prompt your pals to join the chat. Artificial
intelligence is advancing at lightning speed, and with that, ethical debates are heating
up too.

00:14:53
Clip: We think it’s simple to get to alignment when machines are encoded with the
same values as ethical humans. But imagine that we tell the machine to be persistent
and it may refuse to give up when it should.

00:15:07
Clip: Tech leaders are warning about the catastrophic harms posed by artificial
intelligence if left unchecked by government oversight.

00:15:15
Clip: The people building AI earnestly believe that it could kill us all by the end of the
decade.

00:15:20
Dr. Kaylee Byers: Throughout history, major scientific breakthroughs have opened up
incredible new possibilities. And then, sometimes we’ve had to grapple with
consequences we didn’t fully anticipate. Take the automobile. Suddenly people could
travel further and faster than ever before. Amazing. But as cars became part of
everyday life, we also had to reckon with traffic fatalities, air pollution, and all the
infrastructure that comes with millions of vehicles on the road. Or nuclear science.
Unlocking the energy inside atoms opened up possibilities for electricity and scientific

discovery, but it also made nuclear weapons, bringing a whole new set of questions
about how that power should be used. And genomics has raised some of these same
questions. Our growing ability to study and alter genetic material has transformed
opportunities in fields like medicine, agriculture… The list goes on.
While it’s also forcing us to confront difficult questions about safety, eugenics,
capitalism, discrimination, and who gets to benefit. And now, AI is helping researchers
tackle complex biological questions, analyze enormous amounts of data, and explore
possibilities that were once totally out of reach. But just because we can do something
doesn’t automatically mean we should. If you’ve listened to the show before, you know
how much it pains me not to just throw in another Jurassic Park quote here. So, how do
we decide what we should use these technologies for? Who’s accountable when things
go wrong? And how do we protect the people whose data and lives are part of that
research? Running back, getting my keys, panicking… Anyway, hi.

00:17:13
Dr. Mohammad Hosseini: Monday vibes.

00:17:14
Dr. Kaylee Byers: It’s Monday vibes, here with me now is Dr. Mohammad Hosseini.

00:17:20
Dr. Kaylee Byers: I’m an assistant professor at Northwestern University, Feinberg
School of Medicine, Department of Preventive Medicine. What I do mostly is I focus on
the ethics of doing research and research integrity, and in particular, I look into the
ethics of using AI, and I also teach different things ethics-related.

00:17:41
Dr. Kaylee Byers: So when you think of AI and genomics, what’s sort of the first thing
that comes to mind for you?

00:17:46
Dr. Mohammad Hosseini: Two words come to mind, possibility and risk. Possibilities
because with AI, you can do whatever you used to do at scale. And when it comes to
genomics, it’s very important to be able to reduce the amount of tedious and mundane
research tasks so that you can do what you are hoping to do faster. When it comes to
data analysis, genetics data is enormous. For example, we know that the number of
possible synthetic organic combinations, like how many organic combinations you can
synthesize is somewhere around 10 to the power of 60, that’s how many possibilities.

And just for reference, the number of grains of sand in the world is 10 to the power of
18.

00:18:38
Dr. Kaylee Byers: Oh, wild, what a fun number.

00:18:40
Dr. Mohammad Hosseini: Yes. So, AI is like a multiplier of whatever we used to do.
And when it comes to the context of genomics, it’s basically a game of statistics. And in
that sense, AI is an absolutely great tool. Now, the flip side, the risks is that yes, it
enables us to do many things, but are we in a good position to try everything? I wrote
this op-ed for The Chicago Tribune where I use this analogy of imagine AI is like a
master key. Would you try your master key on every door? It may open many doors, but
are you ready to see what is behind it?

00:19:21
Dr. Kaylee Byers: You wrote a paper called, A Plea for Caution and Guidance, about
using AI in genomics. And I’m curious because in the paper you get into a few different
examples, and you’re imagining future responsibilities… And actually a line that you
have in here that I really love is that you say, you’re pushing us to deliberate, “Duties to
future sentient beings,” we’ve got these… We’re pushing things onto the future. And one
example you talk about is, what if we brought back the Neanderthal, right? What if that
is something that we decided we were going to put our time and money and AI
genomics power into? Why is it useful to think about this kind of a scenario when we’re
muddling through the ethics of AI and genomics?

00:20:07
Dr. Mohammad Hosseini: Even if this sounds like sci-fi at the moment, it is useful
when you think about ethics to think about these scenarios because these scenarios
actually bring science to the investigation table. So, instead of science investigating
things, we bring science to the investigation table, and are like, okay, what is this
science going to do and how is it going to do what it’s going to do, and what is going to
be the impact and consequences, and so on? Let’s just move with the example and
think as sci-fi as we can. Suppose we can come up with all kinds of required methods
and tools that enables us to come up with the very first Neanderthal without any
casualty. What are we going to do with it?

00:20:51
Dr. Kaylee Byers: Underneath the extinct human thought experiment is a pretty serious
consideration. If we eventually have the ability to recreate something or someone that

no longer exists, what do we owe the thing or person we bring back? We’ve already
talked about de-extinction for animals on this show, so it’s not completely absurd to ask
where these technologies could take us next. And this is where genomics starts running
headfirst into ethics about human rights, responsibility and biosecurity. I want to come
back to DNA a little bit here. So, as AI tools help us work with DNA and genomics and
they become more powerful and easier to use, what changes and what do we need to
think about in terms of who is able to access and use those tools and the kinds of
oversight that we might need?

00:21:40
Dr. Mohammad Hosseini: Great question. And I think that’s one of those questions
that touches biosecurity. There’s different theories around that, there’s people who say,
let’s regulate what kind of data can be used to train what kind of AI models so that we
can regulate what kind of capabilities come with different AI tools. So, let’s try to do our
regulation at the source, and make sure that data that can cause risks can be a
biosecurity concern is not being used to train future AI models. There’s that camp and
there’s another camp that says, well, regardless of what you know and what kind of
knowledge is fed into these tools, you still need hands and labs and infrastructure to do
the research. So, let that kind of information be publicly available… That’s good, it’s like
open science and so on. But let’s try to regulate the tools and infrastructure and the
expertise that is needed to do what your AI might suggest. But at the end of the day, I
think enforcing them is the bigger challenge. And one major risk is bad actors. We are
not operating in a world where everybody who is interested in science is interested in
extending certified knowledge and discovery of the truth. And this is, I think one of those
instances that very quickly can become political. And once you have an actor that can
do whatever it wants with impunity, even if you have guardrails and guidance, you’ll go
nowhere because they will do it anyway.

00:23:22
Dr. Kaylee Byers: So, what to do? It’s a bit of a balancing act, risk on one side,
possibility on the other. Researchers like Dr. Cong can build that balance into the
technology from the start. I do want to ask you this question, because a tool like
CRISPR-GPT to heal people, to improve health outcomes for things that would’ve
thought decades ago or even two years ago, we couldn’t. And at the same time, I
recently had to fill out a risk assessment form, and one of the things that I have to fill out
is dual use. Could the potential things that you create have this dual use that could
harm people? And so, when you’re working in a platform like this, what do you think are
some potential ways to address those sorts of concerns for dual risk use of such
technologies?

00:24:11
Dr. Le Cong: We actually made sure that if there’s any identifiable sequence within
CRISPR-GPT, for example, we can actually prevent the sequence itself from being
passed into a AI model. So, the AI model will never see the patient information or
identifiable biological sequence, it’s only helping with the logic of how we should be
thinking about a design problem rather than actually operating on the patient sequence
itself. So, I think making sure that information is protected for each of the patient or
individual is very important. And two is, if we did build a separate, we call it a defense
system within CRISPR-GPT, where, if the defense system detect that the person using
it are trying to, for example, use it to engineer a virus sequences or something
dangerous, it will immediately stop. So, I do think that this is almost what we call a
defense mechanism, is I think there’s always need to say, as these AIs become more
powerful and maybe more useful, we need to have a defense mechanism against what
we call security safety of the system. So, I think these two are very important.

00:25:21
Dr. Kaylee Byers: Absolutely, I think those are great steps. I was just reading a book
and they were talking about some of the history around avian influenza, and there were
some researchers that had found a way to make the virus more infective, and they went
to publish it and the government was like, yeah, that’s cool, but maybe no more.
Actually valuable research, stop immediately, you can’t do that anymore. And I think
about that. There’s the advances and then the potential harms, and sometimes we get
really caught up in the things that are possible and what they can give us without
thinking of those consequences. So, I think those are really important safeguards to put
on.

00:25:55
Dr. Le Cong: Yes.

00:26:00
Dr. Kaylee Byers: So, lastly, I just want to talk about this headline about CRISPR-GPT
that says, “CRISPR-GPT turns novice scientists into experts.” And it’s maybe a bit of a
clickbait, and it’s probably an exaggeration, but it also does point to this bigger question
that I think it’s really important to ask regardless, which is, if we democratize something
as powerful as gene editing, who gets to make those decisions?

00:26:25
Dr. Le Cong: My thought is one, we have been very careful in terms of not yet releasing
CRISPR-GPT publicly. Right now, people actually have to sign up and provide a verified
email address, nonprofit affiliation, verification, and what they intend use case for us to

open access. So, I do think that one idea is for a system that could have dual use
potential, we probably wanted to release it initially in a step-by-step fashion, we want to
be mindful that AI is still early in how it’s helping medicines and science, and the
potential implication can be big. So, we have to be really careful that without checking
all the boxes, we probably don’t want to say, oh, we can greenlight everything.

00:27:16
Dr. Kaylee Byers: In science, we’re often frustrated by how long things take, but then
speeding things up? Well, there’s consequences for that too. So, when the stakes are
this high, maybe taking our time isn’t such a bad thing. Here’s Dr. Hosseini again.

00:27:33
Dr. Mohammad Hosseini: AI is like putting science on steroids, and that may be good
for some scientific workflows, but not always. You cannot just run 100 miles an hour all
the time, sometimes you really need to slow down. So, I think something we should try
to advocate for is slower science because anyone who has done any kind of science
knows that sometimes you just need time. Sometimes you need to sleep on an idea.

00:28:06
Dr. Kaylee Byers: I could talk to you about this forever, but I do feel like that’s a good
point for us to end on because we often have this tendency to say, scientific
advancements take forever, but also maybe we want them to? Maybe we need that to
ensure that we’re doing things right and building trust. So, Dr. Hosseini, thank you so
much for coming on and chatting with me. Ethics is actually a real interest of mine, so
I’ve really enjoyed it.

00:28:33
Dr. Mohammad Hosseini: Same, thank you so much for having me.

00:28:37
Dr. Kaylee Byers: Our guests for today were associate professor of Pathology and
Genetics at Stanford University School of Medicine, Dr. Le Cong, and assistant
professor of Preventative Medicine at Northwestern University’s Feinberg School of
Medicine, Dr. Mohammad Hosseini.
You’ve been listening to Nice Genes, a podcast brought to you by Genome British
Columbia. If you like this episode, go back and check out some of our previous ones
wherever you listen from. Share us with your friends and leave us a review. You can
also DM the show on social media by going to @GenomeBC.

Next time, we are talking about pancreatic cancer, a disease that spent decades giving
researchers a very hard no, but now big changes are on the horizon.

00:29:18
Dr. Huffman: It’s a game changer. It’s an absolute game changer, and the elation,
when we see these articles come out and we see the reaction to it in some of the
research that we’re reading, it lifts you completely.

00:29:31
Dr. Kaylee Byers: Thanks so much for listening, and AI caramba, look at the time, got
to go.