Welcome to Impact Quantum, the show that peeks under the hood of

quantum computing to reveal what's emerging and why it

matters. Today's episode is an absolute masterclass

in quantum inspired data science, with a guest who

quite frankly makes the rest of us feel like we're still stuck figuring

out long division. Joining Frank and Candice is Dr.

Marvin Weinstein, emeritus at Stanford University,

bona fide particle physicist and co creator of

dynamic quantum clustering, a method that

sounds like science fiction, but delivers real world,

potentially life saving insight. Marvin takes us on

a thrilling journey through brain cancer research, data

agnosticism, and how a physicist wandered into

biology and found patterns that even seasoned

researchers had missed. This isn't quantum computing per

se, it's quantum mechanics inspired analysis applied with

surgical precision minus the surgical gloves.

Whether you're a curious technologist or just here for the

intellectual thrill ride, this one is for you. And

no, you don't need a PhD to follow along. Just

curiosity and perhaps a cup of strong tea. This episode

is rated 5 Schrodingers. So buckle up and let's get

into it.

Hello and welcome back to Impact Quantum, the podcast where we explore the

emergent fields of quantum computing and

the upcoming ecosystem that is going to spread around it.

So you don't need to be a quantum physicist, but you do need to be

curious and curious about quantum computing. And with me today,

as always, is the most quantum curious person I know, Candace Kahuli.

How's it going, Candace? It's great. Thank you so much for asking. I'm really

excited about today. Yeah. So I think today you actually have

an honest to goodness physicist here on

as a guest. Absolutely. Amongst many things that

he's, that he's doing, I can say that he is a particle physicist at Stanford

University as well as,

as well as the CSO co founder at

Quantum Insights Incorporated. He's got a lot, a

lot of experience and a lot of great knowledge to share

to our audience. I think everyone's going to find him as fascinating as I do.

Cool, I hope.

But I am a genuine quantum mechanic. That's right. There you

go. So please welcome everybody, Marvin

Weinstein to the show. How's it going? It's

going well. As I was telling Candice, you got me in a very

excited state today, so I hope I'm coherent.

Awesome. Yeah. In the virtual green room, you had said you kind of

uncovered something very interesting. So we can start there if you

like. Well, yeah, I mean,

basically the thing we were talking about

during the previous interview was a tool that was Developed that was

what the company was founded for, to apply the

various problems. And it's called dynamic quantum clustering.

And it differs from other clustering

algorithms, other data mining tools, in that it

is completely unbiased. You can take a first look at data with

making no assumptions about if there is anything to be found in the data,

cleaning the data or

labeling it in any manner, shape or form. You just look at the raw

data. So

for personal history reasons, I mean, Candace was telling me about somebody

she knew who partner had

died or father had died of a brain tumor. But

my first wife also died of glioblastoma.

So when Quantum Insights decided

to close its doors, I was sitting with all of this data from the Cancer

Genome Atlas, including all of its glioma

data. That means all low grade gliomas and

glioblastoma data. So I had RNA sequencing data

for all of those tumors. And basically

I decided first thing I want to do with it is take a look

at it and see if there's anything to see in that

data that I mean people have looked, this data set's old,

so it's been around for a long time, has been heavily studied.

People were totally sure that everything that there

was to be extracted from that data set had been extracted from

that data set. And so basically

I said, well, nobody's looked with our tool.

And so what did I do? Well, the first thing was, as

I promised you, I simply loaded up the data. I did

restrict the gene expression from the 60,000 genes

that it comes with down to what everybody believes is

the 20,000 so called protein coding genes.

Not all of them code for proteins, but they're the list

that various tools

restrict to. So I wanted to stay within what other people were doing.

So I looked at those 20,000 genes. Well, that's a lot of data. I mean

that's a lot of noise. It's actually not a lot of data. It's only 600.

I mean that's always a misconception. People say biologists have huge data

sets. They really don't. I mean, for example, all of the

cancer data is 692

tumors, brain cancers.

That's not a big data set. There's only 692 pieces of

information. I don't care that there's 20,000 genes

because all you're seeing is the effect of 692

combinations of the expression levels for those genes. The whole

data set can be reproduced from those 692 pieces

of information. So

not like a physics data set which has Millions of

samples and stuff to look at. This is a biology data

set and typically restricted to a specific disease.

It's not huge. What it is, is

complicated and really hard to see what's going

on because there's so much noise.

So first thing I did, as I said, was restricted

the raw data. It's a matrix after all, rows

and columns, okay? Every row is the expression level

for 20,000 genes. I'm rounding the numbers off. You don't

want the 20,312 all the time.

So it's that. And there are

692 rows in all. You feed that into DQC, it's

made to ingest that quickly and you do, you just

simply run the first analysis and surprise. The first thing you see

is there's a whopping big signal.

In fact that data, raw, unprocessed,

unlabeled, untreated in any way, no

training set, separates into two clusters. One

very large cluster which is mostly the lower grade

gliomas, and another cluster which is

almost all of the

glioblastomas. Well, that's pretty

cool. There's already a signal. It's not the best classification

in the world. Maybe it's very good, it's competitive.

But DQC has a standard trick which is

you can pick out a smaller number of

genes to look at, in this case a smaller number of features in the

fancy language which

give the same information. And so first run at that

produced 544genes and exactly

the same picture. So I didn't have to look at 20,000, I

had to look at 544 which were doing most of the heavy

lifting, produce the same two clusters,

same, not wonderful, but pretty good classification

scheme. Then there's another DQC based trick

which is using the information in the two clusters, now

I can order the genes that I'm looking at, the

544, in order of their

importance to the signal.

Then I look the first 10 genes, the first 20 genes, the first

30 genes, and I did those analyses over and over. Each time I did

it starting from 10, I got a pretty good

classifications. 20 made it better, 30 made it better

until I got up to 90 genes and then at

100, 110, 120, everything

stopped getting better and started to get worse. Interesting. So the

cutoff interesting was I wanted to look at the 90 gene signal

because the cleanest information was going to be in the 90 gene

signal. Did that and sure enough I find

four clusters. So what are the four clusters? Three of

those clusters are all low grade gliomas.

100% low grade gliomas. They

capture all of the low grade gliomas

except for four tumors. The

fourth cluster is all of the glioblastomas and

those four that were not captured. Now remember,

there were 692 tumors, I was missing four.

So when you look at them plotted in the space,

let's call it PCA space. You like that word? And it is the

PCA space for the tumor expressions. Those four

lie right next to the gliomas, whereas all the other data lie far away from

the gliomas. Just for folks that may not know

what PCA is, is it Principal Component

Analysis? Principal component, yeah. PCA

is a way of rotating the data so that

the dimension of the data in which

the data is most spread out is the first dimension. The dimension

in which the data is next most spread out is the second.

It tends, if you're lucky in low dimensions

to show you what you need to see in order to

try to do clustering. Because most clustering algorithms

deteriorate rapidly as the dimension of the data goes

up. So they like to do a hard dimensional reduction, they

call it to two or three PCA directions

and then try to cluster based on what they see there.

There are algorithms which work in higher dimension, but,

but still there are things they struggle with.

DTC doesn't really care. It doesn't start with a hard dimensional

reduction. It simply works

with, with what is showing the most information. If it's 6,

if it's 10, if it's 16, if it's 50, that's fine, I

don't care. I'll, I'll work in that. The only impact

and price I pay is the time it takes to run the algorithm.

But, and so the usual trick is you work in the lowest number

of dimensions that appear to be noise free,

which you can tell by looking at the spectrum that you see in pca

and then work your way up to twice that number of

dimensions and look again and if you see the same information,

well, it's quicker to run everything in the lower dimension, but

you don't stop if you see a change. So

at any rate, Granite, I got these four clusters.

Now you notice the only misclassification out of

692 tumors is four tumors.

So a considerably less than 1%

failure. Of doing close to like

1/7 of 1%, would you say? Yeah, yeah, yeah.

So I mean that's, that's, I'm trying. To quote the worst possible

statistic that I can imagine, but less than 1%. We can all

agree on. So that would put it at over 99%. If I tell you

from this analysis you have a low grade glioma, I'm

100% accurate. Right. If I tell you you have a

glioblastoma, I might be as much as

2 1/2% inaccurate on just glioma question

that's better than world class. Let's say what's current state

of the art is closer to like 80, 20. 19 around trying to

say how do we compete? And I haven't succeeded yet. My

collaborators, one bioinformaticist at

Wisconsin and a cancer doc at Stanford,

are going to have to help me with that. In searching the literature, what

I find are statements like most schemes for

doing this, unsupervised from

the raw data and then moving on from an internal

analysis. Still working, starting with the raw data,

what they call the area under the curve. So the likelihood you're right

is 70 to 80%.

Interesting. So we're not talking anything like the same. There

are some special biomarkers. If they're found

on a glioblastoma, then people are pretty sure

it's a glioblastoma at maybe the 1% level.

Okay, but separating

glioblastoma from low grade gliomas, blind,

they're nowhere near that good.

So at any rate,

that's what I found. I now have the world's best classifier.

In 90 genes, I plot the gene expression levels

for each one of those clusters. And for most of the genesis

I see the genes either fall into the category, the

expectation for the expression of that gene

for the either goes systematically up through

the four clusters, going from the lowest grade glioma to the

glioblastoma, or systematically goes down. That's

what you want to see. Those genes are involved in what's happening,

but there's still 544 genes.

And I can't see the forest for the trees.

Interesting. So does this inform treatment options?

Well, that's the problem. Treatment options, or at

least my understanding. So remember, I have to be

very upfront. I'm not a biologist. Right. Everything

you will hear me talk about, I learned by looking at this data. I have

nothing formal training in biology whatsoever,

so you're dealing with a novice. It reminds me of the

the original Star Trek show where Bones would always like, I'm a doctor, not an

engineer. Like you're like, I'm a physicist, not an engineer. I mean a doctor, you

know. So what, what initially inspired you to take

all of your quantum mechanics Knowledge and, and, and

apply it biological data. Yeah, so remember I'm the

co inventor of this algorithm. The other inventor is David

Horn, Tel Aviv University, a frequent visitor to

Slack. He collaborated on many physics

papers. And Slack is not the

messaging app. It's Stanford Linear Accelerator. Is that right?

No case Stanford. It used to be called the Stanford Linear Accelerator

Center. So I'll tell you out of school the story which

reflects wonderfully on the doe. At some point the

DOE wanted to put its name on everything

and trademark it.

Well, Slack said, because Stanford said you can't trademark

the Stanford Linear Accelerator Center. It's us, right?

We run the place. So DOE made

SLAC change its name to SLAC S L A C

and call it the SLAC National Accelerator Laboratory. So I guess

as an abbreviation we're now Snal, not Slack.

Slack sounds better. I grew up with it as slack for

42 years. To hell with the DOE. I don't intend to

listen to what they want. But it is now officially the SLAC

National Accelerator Laboratory. Right. So at any rate,

David Horn came into my office and life went as normal. He said, oh, I

have something interesting to show you. Because he kind of had left high energy

physics about eight years earlier and was looking

into data mining. And he said there this cool idea

that grows out of

something done by somebody called Emanuel Parsons, called the Parsons

estimator. And I figured out I should think about it as a

quantum potential. I already was very

suspicious. It sounded like a very strange idea.

And so we did our usual thing. We stood at the blackboard and

yelled at one another for three or four hours. And

then we came to a meeting of the minds and said, this really isn't the

stupid idea. It's kind of cute. And

you know, David said, well, he showed me some simple problems

having to do with classifying crabs. It's the standard

old problem that people did

and seemed to be very interesting.

He said, but the problem is in order to understand who's

so the what, the idea behind it is very simple. You take

all the data, you create a function. The properties of this

function are wherever there's more data

than it is in the surrounding, there should be a peak. And

wherever there's less data, there should be a value. The problem is,

of course, the way you create that data is very sensitive to a parameter that

you introduce. Okay, I don't want to get too

messy in this. It's all published so it can be

read and the sensitivity is hard to

deal with. So what we finally

understood was if we treated this. So this

is just a professional deformation.

Because we're particle physicists and quantum mechanics,

we think of everything as having something to do with particle physics.

Quantum mechanics. This problem has nothing to do with particle

physics or quantum mechanics, except we want to get rid of the sensitivity of

that function. And we said, if you think of this as the solution to

a problem in quantum mechanics, that problem has a

term having to do with the particles moving around and another

one having to do with the landscape it finds itself in.

That's called a potential function. It turns out

that potential function always has sharper

features, more pronounced dips

than the solution, has peaks,

and turns out to be much less sensitive to the parameter that goes into

building that function. So, literally, by

saying, what problem is this picture,

the solution to which turns out to be

trivial to solve what potential function,

you get a sharper picture. And the sensitivity, the parameter used

to build what's called the kernel function, that potential function

goes down by a factor of 10. So you have a pretty

unique answer. It's easy to arrive at. You don't have to be careful about

picking your parameters. The problem is if

you think of this as things living. So we have these

valleys now where the bulk of the heavy

concentration of data is, or we have stream

beds, but the data is up along the

walls as well as being down in the

valley. So the question is, which data belongs to which

valley, which stream bed, et cetera.

And so you want to move the points down the sides of the valley

and have them collect in whatever structure is at the bottom.

Well, people try that. In fact, my

colleague had been trying it. And as the dimension goes up,

for reasons we understand, that

surface becomes rippled just due to noise.

And so basically, if you try to just move things

down using ordinary calculus, what's called gradient descent,

you're just moving points in the direction of the slope. They get

stuck in the ripples. Oh, I see. Because

it can't find the global minimum. It can't find the important

minimum. Right. If things move according to quantum

mechanics, all bets are off. It's a much nicer

story. And it's the uncertainty principle, which made the

solution wider to begin with. So we're going to

exploit the uncertainty principle. If I move points

according to the laws of quantum mechanics. The first thing is, unlike

gradient descent, the quantum wave function extends out to

where the valley starts going up again.

So points automatically start to slow down as they

reach the minimum. And they don't overshoot and

rattle around, they just stop because they see now

equal influence from Both walls and therefore no

force. Also they don't see ripples because

the uncertainty principle allows for quantum tunneling

and they simply go through those tiny ripples or ride above them.

So as a way of making the data move

and find the minima in the function in any number of

dimensions and as a way of speeding up the

analysis, because quantum evolution is done by matrix

multiplication, so it's enormously

parallelizable. Didn't say that very well.

Parallelizable, you get a very quick algorithm

that is using physics principles. But to solve a non physics

problem, just getting the points efficiently down to the bottom.

If there is a riverbed that tells

you something about the data, says there's some one parameter

thing, some regression on the data that you can do to

something that's very extended. It's a huge discovery.

It's much better than finding simple clusters.

But that's what this does. So DQC has

advantages. One, it doesn't require training sets.

So it's great for biology data because having annotated

training sets that are really good, hard to combine.

So this is interesting and what does DQC stand for?

Dynamic Quantum clustering. Meaning we're using quantum mechanics.

The find the minimum. Now do you need a quantum computer to do this

or this is just an algorithm? Interesting. I told you I'm here under

false pretenses. You asked me here to talk about quantum

computing. And I told you I don't do quantum computing. I'm talking about using

quantum mechanics to run on an ordinary

computer. Could it run on a quantum computer? Yes, if

they were really as fast and as good as they say they're going to be,

would even be better because it can handle bigger. I'm

focusing on biology, by the way. Way this algorithm is data

agnostic, right? It's not talking about

biology per se, it doesn't care. It just says that

there's something interesting. Data is not distributed with

equal density every place. Things that are more like one

another tend to be located in a more dense region.

Okay, so and this has been applied to many things. It's been

applied to

finding radioactive sources in the city of Chicago hidden

in a building. Okay, it, there's a paper that I wrote on

that it's been applied to. I guess

there's no paper on this, but it was a problem I did for somebody

finding

tanks in the desert that have been camouflaged, painted,

same thing using the data from a

multispectral hyperspectral camera.

So it doesn't care what the data is. It's data agnostic

it's feature agnostic. It is

unsupervised completely. That doesn't mean that you don't use the results

of a previous analysis to now supervise the next analysis

based on what you learned. You do do that.

But at any rate, that was it. So what's now

going on is, as I said, we have the world's best

classifier. But I don't know how to tell you what the

best drug for your tumor, the one that's most likely

to work on, the biology that's happening now, should

be. And that's why I need to go find a biologist and they're

not so great at doing it either. So witness how

many people go through many, many failed drugs. Yeah,

well, precision medicine is definitely, you know, one

of, one of the, you know, one of the biggest outcomes of using

this type of, this type of clustering that

we can, we can create. I mean there's so many, there's, there's just, there's

so much out there that needs this type of, you know,

this type of. Still in its infancy. It's got a place to go to

be precision medicine. Where do you. Oh, go ahead.

Oh, please don't let me. What you have to say. I was going to say

based on dynamic clustering, quantum clustering,

you know, where do you see it evolving in the.

So I'll finish telling you this story about why I'm excited because

I think it's evolving to a really. I, I've

seen something today I never thought I would

see. So last night it showed up at 10 o' clock

in the evening and I'm still digesting what I saw.

What I show you, you should take with a grain of salt. But there's no

question, there's zero chance that I'm wrong in terms

of what you'll see. Okay, so the way

docs like to look at the problem or cancer

researchers is they talk about so called

biological pathways. Biological

pathways are sets of genes

which carry out some process. In the end, all processes

are making proteins, but we're not looking at the proteins being

made, but we know these sets of genes are

functioning together to produce an interesting

output. So if I can

take the information I have and find

a way of saying, oh, so in fact what I'm

seeing is actually predicted by the following

set of genes. And I can assign

meaningful coordinates to each tumor

based on where they are and what that set of

genes is doing together. I mean, biospace,

a point in that space depending on how many

things still I'm producing. One axis in biospace

and it's representing a process which is

happening in the patient where a bunch of genes are

telling me something, not one. And

that bunch of genes I can look at and ask what are their

properties? What are their common properties?

So I will share something with

you. So at any rate, did that.

Okay. Went to biospace using DQC

methods again. Remember I told you I had four clusters. So

there are six pairs of clusters which

differ in how the genes are being expressed in those clusters.

So I can find the most, the list of the most important ones between

1 and 2, 1 and 3, 1 and 4, 2 and 3,

2 and 4, 3 and 4. So

six possible axes

in biospace, the sets of genes that are most important.

And then using those axes which go from minus something

to plus something, I can assign a coordinate to every one of the

tumors. So I have points in a six dimensional space.

Okay. The way that's done,

it's done in a way such that zero on

that axis means that for that set of genes,

that point is consistent with what the value for

all of the genes in that thing. The average value of those genes

is. Plus means you are moving

x standard deviations away from

being at the average expression. So I don't need to know what the

normal expression of a gene is. That's always one of the

problems. You rarely have data for normal

cells of the same type as the tumor.

And so you don't know where to set your zeros. Here I'm doing it by

the average and I'm saying how far from a standard deviation am

I out one way and how far out am I the other way?

And so you plot the same tumors.

Now I have to remember what I do. I go to share

share the screen. So you plot the same set of

tumors. Now you see my background. Yes. And I am going to

switch over to the computer in my basement and show you a fun thing.

So the axes you see here are

DQC's plotting of

the cancers in a six dimensional biospace.

But I want you to see, blues are glioblastomas,

reds are the lowest grade gliomas,

magentas are the next lowest grade gliomas.

And the goals

are closest to the glioblastomas.

Interesting. Now I told you this is an animation. We're going to start the points.

This is how the QC works. Okay. So we're moving the points

downhill.

You like that? Yeah. So it's all. What's

happening, they're all converging into one like a

regression, right? Right. There's a one dimensional

shape. The healthiest tumors, they're not

healthy, but they're the healthiest. They're not the least awful.

Yeah, the least awful. So if I look at for this.

We already saw that when we analyzed the. So the colors

here are the clusters that I discovered

in RNA sequencing space in what we call

gene space.

They've just been arranged in a line from best to

worst. So blue is the worst. The

glioblastomas over here. Okay. Okay. So

for those listening, don't worry, we're gonna link in the show notes to a video

representation of this. Interesting. At

any rate, this is what you see.

So the. This is the plot in bio space. Now that's very

interesting because these have the dimensions of the bio coordinates

and those coordinates have a meaning.

Okay. In fact, I'll tell you what the meaning is. And this is based

upon a set of data of patients.

Yes, this is 692 patients.

That data was submitted to the cancer genome project. Okay. Oh, so this

is open source data that you're pulling. This is absolutely open source.

What my company did when we existed, because we had various

projects going and things like this, we downloaded all of

that data for the RNA sequencing data and as much

as we could find about each of those tumors, which was

not a hell of a lot. But there's something, It's a good

database. As I said, it's been studied for years and years and years.

So this is results obtained by starting from no information

and just relooking at the brain cancer data

and saying, people have been studying this forever. Did they ever

find anything like this? And the answer is no.

This has never been discovered. This has never been discussed. So

using traditional analytical sources,

you could not. Whatever. You could not get at this information

without doing the. The dynamic

because you make a lot. Of assumptions about what you're supposed to look at. You

make a lot of assumptions about how you filter the data.

You end up throwing the baby out with the bathwater.

Go ahead. No, you also said you didn't do any cleanup of the data. Like

that's just the wrong. No. Well, I mean, they've cleaned it up obviously. Obviously at

some level. But we're not doing the post. Whatever they

did cleanup that people normally do where they filter out

genes, where they have this gene should be expressed at

least at this level. All genes that aren't expressed at that

level we're throwing out of the data set. Okay.

If I see a difference between two

clusters and the genes are expressed differently in the two clusters,

but what they call the fold value isn't big enough.

I'm throwing it out of the data. Well, you can imagine

if there's hidden information in the data and you're busy throwing things

away, the chance you throw the baby out with the water bath water

is very high. Exactly. And

that's exactly what this shows. The benefit of going in

unbiased, unfiltered,

completely agnostic. Look to see if there's a signal first.

And then when you see the signal, which I did. So stage one is,

wow, there's a signal. Stage two, what is making the

signal? EQC is built for solving those

problems. Right. So basically, and

that's where it differs from AI, okay, AI needs training

sets for the most part. There are

versions of AI now that claim not to, which

are real. They make up data in order to train

the data. There aren't enough training sets.

So what you do instead is you make up artificial data and then try

to teach it to reconstruct the real data.

Okay, by by picking the parameters in the artificial data

and then you try to classify existing data.

But it's a different story here.

Everything is understood. The algorithm is totally

prescriptive. I know exactly what's going on.

There's no mystery. Once I find something

and we ended up, I just showed you with this concept of

biospace, which is what

people in literature, it turns out that's where the idea came from

to look at it this way, what people were

talking about as latent coordinates in the data.

So there are people doing AI that say, oh, I'm going to keep feeding

AI from this and AI is going to reduce my problem to

some low dimensional manifold and I'll call that a latent

coordinate picture. But then I'm faced with the problem. I don't really know

what the coordinates mean. I am busy trying to interpret

them and I certainly don't know how to exploit them.

Different here. Right. So we started with no training data.

Am I looking at here? Shouldn't be showing

you this, but my, my collaborators say I can show it to you.

So here are the axes. So what do

you know from these axes? Well, the

genes in this axis have, as I

say, tumor associated fibroblast activation,

their immune checkpoint genes

signaling chemokine driven inflation, the pathways that are

being recruited for this or that. Basically

here it says if you want to

change overexpression or under expression, you want to look at

the drugs which do the following thing.

There's one such description for every one of the six axes

they have A meaning. And so if I simply look at the

coordinates and biospace and see which. Along which of these

axes the biggest signal lies,

that's the first set of drugs you try on the tumor.

So by looking at in biospace and how the tumor

evolves in biospace,

that's what this is, right? The evolution of the tumor

in biospace. Every one of these points, after all, is a

snapshot in time of the tumor at that point.

What this suggests is it's a continuous

evolution to glioblastoma through these

biological processes. And as they change.

So you're seeing the. So basically, what

have I learned? God is showing me, or biology is

showing me how the tumors evolved in

time.

Interesting. I don't know. That doesn't. So do they all start out

as like you showed that image again, but

the one where they're all on the same plane, the one that we're all

on the same plane. This is the

snapshot and survival term for the patient because that's what

is changing along this curve. We already saw that. Oh, I see. So

this had different survival times. So these are all

tumors. I don't know where healthy is.

Okay. So not everybody starts out, for example,

you know, in the red. And then basically

they probably do. Okay.

Glioblastomas have to be. If you look at them in terms of their gene

expression patterns, they're a mess. Okay.

They've undergone many mutations to get where they are. And the more mutations,

that's the different colors, basically, and they change. Okay. So

everyone starts out maybe with the red, but not everybody

goes all the way to the blue. Purple. Right. And they probably. Everybody probably

starts out to the left of the red. Right. Because these

tumors probably form at the single cell or small number of

cell levels. Okay. And take 10 years to grow.

Okay. The first show up and be seen. Okay. So

it's not. We don't have examples of the earliest

version. That's the beauty of what. What's blowing me away. Yeah.

Don't need to know any of this. I don't

have to know. I only need the gene expression pattern

and I only needed the information about survival time to

interpret the axis. Everything else came after

I found the axes when I had to interrogate

pathway databases to find out what they do.

And truth be told, I asked an AI to give me the

information about that because it's a pain in the ass to

go through those things yourself. So we could use. And I just

wanted to know what I might see. This is not to be taken

seriously. Okay. Because My, my

biologist and my doctor friend are going to have to do the job

of vetting what these interpretations. I only trust

AIs a little

bit. It's sort of fun to do that. Okay.

But what I wanted to give you was a feeling

for the difference between biospace information

and simple single gene information. Okay.

And it's awesome what the difference is. And

it's awesome that there's a progression in

biological processes that lead you to

glioblastoma. I can't tell

you this actually represents evolution,

but if it looks like evolution and it smells like

evolution and it wax like evolution, it's

evolution, okay? I mean that's just my feeling.

Now I've already given you all the I don't know any biology,

do know a lot of physics, do know how DQC works.

Okay. I know that better than anybody. But this

business that you can take the information that you learned in

the genetic right, in the single gene

basis and convert it to biological

process basis and learn entirely new things

more suited to advising doctors who are treating

cancer patients. Because I can take a new

tumor stuff and put it on that plot, see where it

is, see what its access definition is

and see what the likely best drug is to start with. And

then if that doesn't work, drop down to the next most likely. The next

most likely. So

basically that sort of we can stop sharing actually

now, which he says I can stop sharing.

Okay, great. So you know why I'm

in this befuddled state at the moment? Because I am still

absorbing what this is telling me. I certainly never expected

when I thought of trying that because people talked about these latent

variables and hidden dimension, hidden coordinates

and describe ways that might work. I didn't see any

examples actually worked out. This is the

story from beginning to end

genetic coordinates to discovery

to the world's best classifier to changing that into

bio coordinates discovered from the genetic side.

The treatment options, a tool for helping doctors treat,

for suggesting to cancer researchers new experiments to

do to verify what they're seeing on this.

Lots of suggested. I alone with no knowledge can think

of 10 things people should explore based on this. And

drug companies want to know what the next set of things

to target should be for a given disease.

Wow. I think that's pretty cool. That is

impressive. So it really is. You know,

DQC is telling me the data is whispering

to you. I'm the tool

that'll teach you how to listen.

That's the way I feel about it. Since it's my baby, it's grown

up, I really think it's grown up and

I'm very impressed with where it got. So you're getting me in my

very biased statement for it. Oh, we can tell it's super, super humble. But

no, it's really. It's really exciting. But also to see where it can

be taken from there, you know, like, this is just the beginning.

The. There's so many scratching the surface. First place, those

axes could be improved because there's more than one

set of genes that give similar information

how to exploit it, how to do the bench experiments.

That's not me. I don't know that stuff. And I'm

83. I'm not ready to start learning how to be a bench

biologist. Okay. But

it's. It. It's just so cool. I mean, you know, it's

like you've seen the underbelly of what's happening in the biology.

At any rate, I don't know if you agree with me, but I think it's

really cool. No, that is really cool.

There's a lot to take in. I'm sorry about

that. No, no, I mean, you know, we have a scale system for these

shows, right? Like five. Five. What is the five Schrodinger.

Schrodingers, yeah. So we have, like from zero to five Schrodingers. This is definitely gonna

be a good five Schrodinger show. Like, and I was able to follow on because

I was a d. Data scientist before this. So, like, when you

said pca, like, I knew what you were referring to at least. I.

But like, so, like, it was like, this show is really geared towards the

quantum curious. Some of which will be data scientists, some of these will be

marketers, some of those will be, you know, kind of traditional software engineers,

et cetera, et cetera. Marketers. Right. Because it's our thesis that

when the quantum computing ecosystem comes around, and

indeed, I think what you've proven today is you don't really need

quantum computing to take advantage of the

innovations in quantum science. Right. Like. Right.

I think that was an assumption I think Candace and I had. I don't want

to speak for Candace, but I know I certainly did. But I know that there

is a field called quantum inspired algorithms, which is probably.

That's sort of what this falls. Yeah,

but it's just exciting

that innovation like this can come about in such a way that

it's going to improve people's lives. What you've discovered is

I'm not a biologist or a doctor, but I would imagine that a doctor or

pharmaceutical Researcher would look at that and say, oh, you know what this means? This

means xyz. I hope so. I mean, I mean

I'm pretty much at the limit of what I can do even

with collaborators on our own. The, the point

we're writing the paper now. This is, I've already blown

my collaborators out of the water because this was discovered last night and they

don't know about it yet. I have their permission to talk

about it though, so that's cool. It's,

you know, so I, I, I'm glad you liked it and the five

shortinger level because I'm only here because you guys refuse

statement. I don't know anything about quantum. Well, I, I do know something about quantum

computing but I am not a quantum computer person and

so I didn't belong on your show but you kept refusing to let me off.

But I mean, I think it's important that people think about like this is not,

I think one of the things that obviously you're, you're, you're a great

presenter and great teacher of these very

complicated topics but you've also something to figure it out. Plus I also think it's

important for people to realize that quite quantum physics and research in that

space is already improving people's lives or at

least already showing fruits of that. And

I think that your research kind of shows that. It's like, you know, you don't

have gen, you don't have the billionaires facing off over, you know,

Jensen saying it's going to take 20 years, Bill Gates saying it's going to take

less. Right. I mean this is pretty basement

and the data is free. So I think the other lesson here

is we have a wealth of data that's under explored

because looking at it in an unbiased fashion hasn't been done.

Right. So I have lots more diseases I want to look at

and I have all this TCGA data for

pancreatic cancer and various other

cancer and so

it's sort of fun, right? I like how you kind

of mix, you know, I know you say you weren't appropriate, but I think you

were totally appropriate for the show and you've got the physics

background when you're talking about quantum clustering,

why it's affecting the biological,

giving us biological data that we're able to move forward with

potentially for precision medicine. I love the

bridges that are being created all over

the place here that you're not just kind of stuck in one thing thinking you

can only do one thing because you have a certain amount of knowledge but how

you've bridged that to bring in all of this

biological data information, I think it's

fantastic. I'm very happy that you came and you joined us today.

I learned. I'm glad I didn't bore you and I hope I didn't get too

far into the weeds, which my wife accuses me of doing all the time.

Mine too.

Where can people find out more about you and what you're up to?

Me and what I'm up to? Well, I'm on LinkedIn. People contact

me through LinkedIn all the time.

I have a long history and you know, if you go look at

the archives, the physics archives. Physrev. Physrev A.

Physrev B. I, I mean my, my past history is a little

eclectic, even in physics, which I attribute to having a

short attention Spanish. But I started in

particle physics. In phenomenology means looking at data,

trying to understand what it's telling me. I moved into

pure abstract particle physics

and then I went into what's called lattice field theory and lattice gauge

theory, which is trying to learn stuff from

how to say this. Didn't expect to talk about this. So,

so let's talk about how we do physics, which is another

totally off topic thing. And you may be running out of time. I don't know.

You tell me when I have to shut up.

But the, the, the story is

physicists are smart, but there are very few problems we know how to solve

exactly. Only a handful.

Everything else is done by a process we call perturbation theory.

Mathematicians also call it perturbation theory. You say, well, this

problem that I know how to solve exactly kind of looks

a little bit like this other problem, but with some

modifications. So let me add the modifications to the problem

and try to calculate corrections to the answer

based on the modifications. So I have the original problem

set and forces involved and the changes in those

forces a little bit. And then I calculate

perturbatively what's happening. People do it in

celestial physics all the time. I have this

planet moving around the sun in an elliptical orbit. Oh well, but

there's the moon. So how does that affect the orbit?

Well, I can't solve that problem. That's already a three body problem.

And there's no exact solution to the three body problem by the time

it's also got Mars and Jupiter and

Saturn and Pluto and Mercury in the problem.

I can't plot orbits. But people do it all the time.

NASA plots orbits. How do they do it? They calculate

the original orbits and they Start calculating the effects of Mars

and this and that on that orbit, because we know what those

forces are if Mars is on its orbit. And through

successive corrections, successive iterations, you're

able to make the small perturbations in the orbit that get the answer

right for you and eventually lets you send something to the moon

and not miss. Okay,

so perturbation theory is, is what we use. But what is perturbation

theory based on? I have a solution, I know how to get

exactly. And I know how to make small corrections

to that solution. And then I can describe all kinds of

crap. So, for example,

condensed matter physics talks about matter.

So I ask you, has anybody ever proved that the table you're

sitting at exists?

Is there such a thing as a table made out of wood? In fact,

is there such a thing as wood? The answer is no.

Use wood to build houses. I use engineering

principles to calculate the stress and load on a beam.

How the hell do I do that if I don't know wood exists?

I describe wood, I assume it exists, I

characterize it in terms of a bunch of properties,

and then I can, based on that, make small correction

calculations again to see how the wood behaves

when I stand on it. But I have to start from the

assumption it exists and that there are properties

I can measure for it and make prediction based on that.

But the first principles thing that would exists, no way.

Nobody solved that problem. Okay? So I was

very interested in that because that's sort of a first principles problem,

right? It's very philosophical, isn't it? It's where the, a

hard science like physics kind of meets up against.

Oh, we meet up against soft stuff all the time and

we fail to solve the problem. But that's okay.

It, it's. At any rate,

I was always interested, always after many years in

phenomenology, I and papers

published in phenomenology and things like that,

getting into field theory and, and

trying to understand from first principles how to solve hard problems

that, like quantum chromodynamics.

That intrigued me because we're kind of using up this

perturbation theory paradigm, okay? It's very

useful, it's very good. But we're already running into lots of

problems where it doesn't work. We don't know a problem that's

approximately like the problem we want to solve.

So how do you solve it? So I got involved in that. I got involved

in what's called lattice field theory. And then I said, but how am I going

to know I'm right? Because

I could be Wrong in pushing my answer in the one

known direction. There got to be other problems,

but there's only one quantum chromodynamics. It's the one we

live with, it's the one we're made of.

So I don't know if I'm cheating or not, but there's

lots of condensed matter problems and they all have different

answers and many of them are strong coupling problems and you

can't treat them perturbated. So take the same methods

and change your field and go look at condensed matter and see if you can

develop techniques to do that. Then did that for

a long time and then developed some methods and decided,

oh, David Horn came into my office and I said,

oh, this looks interesting. So I can't stay

in one area now, to me it makes sense why I'm changing to other

people. It looks like I have no attention span. So that's

okay because I do this for me. And

so as long as I see the thread, I'm happy. But that's how I'm here.

I'm now in biology, quote. But we're

glad. We're glad that you're here. Glad that we got to learn

a bunch of stuff today. I think it's going to be really

exciting to unpack it and to

have you back because you are just a. Few

guys, but I'm going to bore you. So. No, I don't feel bored. I

mean, I'm more fascinated. I'm confused. It's about some things, but,

like, I'm also fascinated, too, and we want to be respectful of your

time and. But we'd love to have you back on the show.

I'm sitting in my office. I have

Nothing on until 5:00 clock this evening. Awesome. We'll

definitely have you come back then because again,

it's just really great information. It's important, it's exciting. I think it's very

exciting. So, unfortunately, we have a little limitation,

so. Yeah, but definitely. And so folks can

reach out to you on LinkedIn and engage with you directly, if you're cool with

that and let your AI. I don't promise to

answer everybody, and if they're a crackpot,

I don't promise to be polite. There you go. That's fair.

I'm liking that. I like that. Let our

AI finish the show. And that wraps this quantum

odyssey on impact. Quantum. A massive thank you to Dr.

Marvin Weinstein for taking us deep into the fractal jungle of

biology, data, science and quantum mechanics with

only his brain, DQC and a suspiciously

underutilized basement server farm. From classifying

glioblastomas with 99% accuracy to uncovering

biocordinates that could revolutionize precision

medicine. Marvin reminded us that sometimes the biggest

scientific breakthroughs don't require a billion dollar

lab, just a stubborn physicist, open source data,

and the audacity to ask what if? If you enjoyed

this episode, and really, how could you not? Be sure to

subscribe, share and let your fellow Quantum Curious friends

know. And as always, check the show notes for links to

Marvin's work, ways to connect, and possibly a

diagram that will make your head spin just a little less.

Until next time, stay curious, stay entangled,

and remember, just because you can't observe the Quantum doesn't mean it's

not observing you. Cheers.