WEBVTT

00:00:05.160 --> 00:00:08.730
Have you ever wanted to scream or just pull
out your hair because you couldn't get an

00:00:08.730 --> 00:00:12.650
experiment to work? I'm going to share a
bit of unpublished data from my PhD. And

00:00:12.650 --> 00:00:16.820
the reason it wasn't published is because
it didn't work. For context, I was at the

00:00:16.820 --> 00:00:20.160
very tail end of graduate school, and I'm
just trying to get out the door. I

00:00:20.160 --> 00:00:23.390
discovered some really interesting
circuitry in brain slices for

00:00:23.390 --> 00:00:26.710
bi-directionally controlling dopamine
neurons. But to get that elusive nature

00:00:26.710 --> 00:00:30.270
paper, I needed to figure out what the role
of the circuitry was in behavior. So I ran

00:00:30.270 --> 00:00:33.100
an experiment in mice. It's the technique
called in vivo optogenetics. And it

00:00:33.100 --> 00:00:38.600
didn't work. The data analysis alone took
me several hours, and then I spent three

00:00:38.600 --> 00:00:42.170
days just trying to figure out what went
wrong. I was also writing my

00:00:42.170 --> 00:00:45.230
dissertation, so I never went back and
redid the experiment. Here, I'm going to

00:00:45.230 --> 00:00:49.880
use Potato to automatically analyze that
data and troubleshoot the experiment. So

00:00:49.880 --> 00:00:53.700
I'll upload the raw data in a research
paper looking at the same circuitry.

00:00:53.700 --> 00:00:58.150
First, you can see that Tater, our AI
co-scientist, uses a data analysis agent

00:00:58.150 --> 00:01:04.220
to write code. It's going to first check
that data and preprocess it. It'll do some

00:01:04.220 --> 00:01:08.090
statistical analyses and graph the
results. And I can use this code in my own

00:01:08.090 --> 00:01:12.130
Jupyter notebook to replicate the
findings or tweak the figures. As you can

00:01:12.130 --> 00:01:15.860
see, there doesn't appear to be a
difference between the control and the

00:01:15.860 --> 00:01:20.580
treatment mice. But Potato also graphs
the data across experimental sessions.

00:01:20.580 --> 00:01:24.310
Again, not really in effect. It even did
some analyses I never even thought of

00:01:24.310 --> 00:01:29.030
doing. Now the data analysis agent is
going to hand its findings back to Tater,

00:01:29.030 --> 00:01:32.870
which gives me a report of the findings.
Here it gives me the bad news. The result's

00:01:32.870 --> 00:01:36.720
not statistically significant. It also
points out some trends. Interestingly,

00:01:36.720 --> 00:01:40.280
it suggests that there is a consistent
effect with one animal. I actually

00:01:40.280 --> 00:01:43.910
remember that mouse. Now it's going to go
and do a critical review of the paper I gave

00:01:43.910 --> 00:01:47.950
it in light of my hypothesis. Tater is
going to compare the findings from my raw

00:01:47.950 --> 00:01:53.120
data and its analysis to the results in the
paper. Then it points out the reason my

00:01:53.120 --> 00:01:56.920
experiment didn't work. The Yang paper
did extensive validation of targeting,

00:01:56.920 --> 00:02:01.600
so I may have injected the virus into the
wrong location. The final report TATER

00:02:01.600 --> 00:02:05.270
produces is structured similar to a
scientific paper. It goes into more

00:02:05.270 --> 00:02:09.100
detail and gives me specific suggestions
on how to rerun the experiment. We've

00:02:09.100 --> 00:02:13.020
deployed TATER to a number of labs. If you
want to use it to do analysis and

00:02:13.020 --> 00:02:16.210
troubleshooting, instead of repeatedly
banging your head on the lab bench, reach

00:02:16.210 --> 00:02:16.670
out to us.