Dr. David Fajgenbaum

And so now within this giant graph of every disease, every gene, every protein, you would find Castleman's with lines or edges to these two concepts and then lines or edges to serolimus.

And you would see a connection between them.

And so now imagine doing that for every disease, every gene, every protein, basically what the world knows about all of medicine.

That's exactly right.

This leads to this, leads to this, and this reverses this, which reverses this, reverses that.

Everything is there.

It's everything.

Well, now what we do is we train machine learning algorithms on all of those known treatments.

So like the serolimus for Castleman's, sildenafil for pulmonary arterial hypertension, insulin for diabetes.

Imagine training this algorithm, because machine learning algorithms are so good

at finding patterns.

And so we're giving the machine learning algorithm lots of information about known treatments.

And we're saying, this is an example of when a drug works for a disease.

And we do it thousands of times with all of the treatments that are out there for all the diseases that are out there.

And then we say, okay, algorithm, now go and score how close of a pattern the connection is between a known treats relationship for every other drug versus every disease.

So if a toenail fungus drug works

looks like there's no way it could work for pancreatic cancer, you need to give it as close to a zero as possible, 0.0001, right?

But if leucovorin looks really promising for a subtype of autism, because the pattern of connections are there and there's a clear intermediary between that subtype of autism and that metabolite, give it a high score, so you get a 0.99.

And so now what we do, we do all 4,000 drugs, all 18,000 diseases.

So it's about 75 million scores that we generate, that our machine learning algorithms generate.