Curing Disease With Genetics And AI

Aug 26, 2023

Curing Disease With Genetics And AI

Manolis Kellis, an accomplished Computer Science Professor at MIT and member of the Broad Institute, is a trailblazer in computational biology. Renowned for leading the MIT Computational Biology Group, his impactful research spans disease genetics, epigenomics, and gene circuitry. With numerous cited publications and leadership in transformative genomics projects, Kellis has garnered prestigious accolades, including the PECASE and Sloan Fellowship, shaping the field with his international perspective from Greece and France to the US.

Subscribe to FORBES: https://www.youtube.com/user/Forbes?s…

Fuel your success with Forbes. Gain unlimited access to premium journalism, including breaking news, groundbreaking in-depth reported stories, daily digests and more. Plus, members get a front-row seat at members-only events with leading thinkers and doers, access to premium video that can help you get ahead, an ad-light experience, early access to select products including NFT drops and more:

https://account.forbes.com/membership…

Stay Connected
Forbes newsletters: https://newsletters.editorial.forbes.com
Forbes on Facebook: http://fb.com/forbes
Forbes Video on Twitter: http://www.twitter.com/forbes
Forbes Video on Instagram: http://instagram.com/forbes
More From Forbes: http://forbes.com

Forbes covers the intersection of entrepreneurship, wealth, technology, business and lifestyle with a focus on people and success.

Content

0 ->

0.94 -> How do we use artificial intelligence

2.95 -> to truly transform human health, to make aging optional,

6.76 -> to make disease a thing of the past?

10.04 -> How do we do that?

11.18 -> The way to do that is to start with causality.

14.74 -> And causality is very hard in human epidemiology.

17.8 -> The vast majority of data is about correlations.

21.1 -> But the one place where causality is there is genetics.

25.07 -> So we start with genetic information, which basically

29.23 -> tells us across the 6 million common variants that all of us

33.79 -> have inherited from our mom and our dad, which ones

36.7 -> are associated with disease?

38.68 -> That allows us to know that there's something causal going

41.62 -> on there.

42.25 -> But that doesn't tell us how it works.

44.38 -> The beauty and the power of genetics

46.69 -> is that it doesn't matter what genes are made of.

48.79 -> It doesn't matter if they're made of green cheese.

50.873 -> Genetics will still work.

52.94 -> But the challenge is that after you have the genetic hit,

55.84 -> you have no idea how it works.

57.52 -> You have no idea what cell type, what gene

59.65 -> might be the target, what tissue of action.

63.02 -> The mechanism is simply unknown.

65.51 -> And how do you develop therapeutics without mechanism?

67.98 -> So the way that we do that is by systematically starting

71.09 -> from genetics and then profiling the molecular impact

75.74 -> of those genetic differences at the cellular level

79.37 -> to basically understand, for example, for patients that

81.92 -> show Alzheimer's, how is APOE4 impacting

85.61 -> the brain, and not just the brain as a whole,

89.7 -> but every single cell type of the brain?

92.02 -> So what we need to do is understand the molecular impact

95.07 -> of every one of those genetic variants

96.93 -> at single-cell resolution in every tissue

100.44 -> and every cell type of the body.

102.48 -> That's where AI begins.

104.64 -> AI begins with data.

106.15 -> So we have the genetics.

107.19 -> We have the molecular phenotypes.

108.63 -> We can now start integrating the data sets together.

111.81 -> And we can basically ask, what are

113.76 -> the driver genes, regulatory control regions, and cell types

118.95 -> that are mediating the effect of the genetic variants?

121.71 -> It doesn't matter if you carry the genetic variant.

124.41 -> You carry the circuit, whether or not you have the variant.

127.683 -> And that basically means that we can intervene

129.6 -> for everyone in this room, if only two--

131.7 -> even if only two people contain the specific genetic variant

135.18 -> associated with obesity, Alzheimer's,

137.25 -> cardiovascular disease, and so on and so forth.

139.33 -> So once we've integrated all these data,

141 -> we then make predictions about what is driving the disease.

144.46 -> And then we go and change these elements.

147.49 -> We perturb them.

148.76 -> The beauty of circuits is that you can then intervene.

152.17 -> And we, for example, did this for the strongest

155.02 -> genetic association with obesity.

157.07 -> So I carry two copies.

158.29 -> Thank you, both mom and dad.

160.06 -> They basically gave me a risk allele for obesity.

162.91 -> And that basically means that if you look at my circuit,

168.4 -> there's something going on there.

170.2 -> And what changes is not the genes directly.

172.96 -> What changes instead is individual nucleotides

176.11 -> that are sometimes affecting genes that are far, far away.

179.74 -> And that's what we've uncovered here.

181.42 -> We've basically found that underlying the strongest

184.39 -> association with obesity lies a circuit that ultimately

187.39 -> controls whether your fat cells will burn calories

191.795 -> that you don't consume through the day,

193.42 -> that you don't use through the day,

195.55 -> or whether they will store them for a rainy day.

198.76 -> And the storing allele is great if you're living

202.36 -> in food-deprived societies.

204.52 -> But in today's world, where we have office jobs and McDonald's

208.45 -> and cookies outside in the break,

210.37 -> you basically don't want to store every calorie.

213.01 -> So by understanding these circuits,

215.23 -> we can now start intervening.

217.61 -> So we've basically developed methods

219.49 -> for changing the upstream or the downstream target genes

223.57 -> from these regulators.

224.77 -> We can also go with genome editing

227.35 -> with CRISPR-Cas9 genome editing and change

229.84 -> that single nucleotide letter that

232.03 -> makes me at risk for obesity.

234.64 -> And what we find in every one of those cases

236.74 -> is that when we understand the circuits,

238.78 -> we can intervene like a switch.

241.54 -> We can flip our cells back and forth

243.58 -> between thermogenic, burning calories,

246.37 -> and lipogenic, storing fat.

248.92 -> And what we saw is that with a single letter alteration out

252.19 -> of 3 billion letters of human genome,

254.05 -> we can basically flip the switch and restore

256.839 -> thermogenesis in fat cells from individuals like myself.

260.529 -> We can also go into mice and change that circuit

263.59 -> to downregulate the genes that get derepressed

267.25 -> when humans have this allele.

270.53 -> And what we find is a dramatic shift,

272.5 -> where mice simply cannot gain weight.

275.315 -> You put them on a high-fat diet, normal mice gain weight.

277.69 -> These mice are unable to gain weight.

279.74 -> So they can exercise as little as they want.

282.41 -> They can sleep as much as they want.

284.03 -> They just simply don't put on calories.

285.938 -> So if you guys want some pills for that, I have--

287.98 -> [LAUGHTER]

289.48 -> I have them in the back.

291.74 -> So that's one example for obesity.

294.13 -> What's another great killer of our time?

298.07 -> Alzheimer's disease.

299.15 -> So, again, with an aging population,

300.65 -> Alzheimer's has become extremely important.

302.65 -> And the previous variant that I showed you

304.4 -> increases your obesity risk by one standard deviation.

307.19 -> Here, this increases your risk by a factor of 10.

310.7 -> Homozygous risk for APOE4 have 10-fold higher risk.

314.57 -> Again, doing the single-cell dissection,

316.58 -> understanding the circuitry, we found

318.32 -> that APOE4 controls the transport of lipids,

323.27 -> and specifically myelinating cholesterol,

326.45 -> through oligodendrocytes.

327.89 -> These are the cells of your brain that protect the neurons.

331.01 -> And the cholesterol gets stuck in the endoplasmic reticulum.

333.543 -> It doesn't make it all the way to myelin

335.21 -> to protect the neurons.

336.51 -> So by understanding the circuitry, we said, OK, great.

338.93 -> Let's restore transport.

340.22 -> And we found that it restores myelination.

342.71 -> And it also restores cognition in the mice.

346.04 -> We have some of that behind the other door.

349.41 -> So here's a third example by Jackie,

351.433 -> who's actually sitting at the back of the room over there.

353.85 -> So, basically, by looking at a cohort

355.76 -> of immunotherapy patients, some of whom responded, some of whom

360.59 -> do not respond to immunotherapy for metastatic melanoma,

363.92 -> we basically were able to predict a regulator that sits

367.22 -> upstream of the nonresponders.

369.8 -> And we're able to overexpress that regulator

373.25 -> and suddenly use a combination therapy that

375.77 -> restores the effects of immunotherapy for all patients.

379.92 -> So that's what we want to do systematically

381.74 -> for every disease.

382.77 -> So how do we do that?

383.72 -> It starts with big data.

385.28 -> We basically start with understanding the genome

387.553 -> systematically, understanding the regulatory control

389.72 -> circuitry, being able to map, where

392.06 -> are the genetic variants across thousands

394.04 -> of diseases and thousands of tissues,

396.17 -> and being able to build these maps that tell us what tissue

399.32 -> is every disease acting in based on the overlap

402.53 -> between the control regions of that tissue

404.87 -> and the genetic variants associated with that disease.

409.23 -> We can then understand the causality

411.92 -> through mediation analysis, Mendelian randomization,

414.74 -> and other approaches to understand

416.45 -> how genetic variants are acting through these intermediate

419.49 -> molecular phenotypes to be able to predict

421.89 -> the impact in the brain based on--

425.55 -> at 93 years of age based on the genetic variants

428.16 -> that you have when you are first born.

430.09 -> So we're able to, for 50,000 loci,

432.54 -> start predicting the epigenome, predicting the regulatory state

436.23 -> of your cells.

437.25 -> And we can use this to discover new disease

440.4 -> loci that are causal in the disease circuitry.

443.85 -> We can then go at single-cell resolution.

447.21 -> And we have written a series of grants in my group

449.61 -> to understand cardiac disease, obesity,

453.48 -> 12 different neurodegenerative disorders, ALS, FTD,

456.84 -> schizophrenia, bipolar disorder, autism.

459.81 -> And for all of those, we're basically

461.76 -> looking at cohorts of individuals

463.83 -> working with clinics, working with doctors,

465.81 -> working with patients to get samples,

467.79 -> to then understand, how are these changing

470.07 -> at single-cell resolution?

471.72 -> What that allows us to do is now start

473.52 -> asking about the progression of disease.

475.68 -> We can treat every cell as being in a different stage

479.49 -> on the progression from non-AD to AD

481.502 -> and start understanding, what are the genetic variants that

483.96 -> are acting early or late in the disease progression?

487.02 -> We can go at single-cell resolution

489.27 -> and start mapping even in subcellular resolution

492.18 -> the spatial organization of where

494.82 -> are these genes expressed, and where are they

497.1 -> differentially expressed in the context of complex tissues

500.49 -> to try to understand the networks that are guiding,

503.34 -> in this particular case, cardiac disorders and coronary artery

506.79 -> disease and heart failure, which we are finding

509.16 -> are, in fact, altering the expression of foam cells

512.58 -> macrophages that engulf the lipids that your fat cells are

516.833 -> not able to store and that are now

518.25 -> going into your veins, which are the driver of the number one

522.15 -> cause of mortality, which is, of course, cardiac arrest.

526.71 -> We can then start asking about the impact

529.26 -> not just in one tissue, but in multiple tissue simultaneously.

532.56 -> We can basically take cohorts of humans and mice,

535.14 -> subject them to diet and exercise interventions,

537.99 -> and then see how those are, in fact, changing multiple tissues

541.41 -> in the body simultaneously.

542.802 -> And what are the cells and circuits

544.26 -> that are mediating the beneficiary effects of exercise

548.07 -> so that we can then go and alter those circuits

550.86 -> and recapitulate the effects of exercise

553.62 -> for people who might be bedridden

555.24 -> or simply not have the time?

556.95 -> We have some pills of that up there as well.

560.41 -> We can then start asking, can we understand

563.07 -> the diversity of disease?

565.05 -> Can we understand now how we can use these hallmarks

568.95 -> that we're finding?

569.97 -> Across hundreds of individuals, we

571.53 -> can basically see the subsets of cells

573.99 -> that are turning on and off repeatedly

575.94 -> in different patients and how they

577.89 -> turn on and off in combinations across those patients.

580.44 -> And we can use that to start understanding

583.14 -> the subtypes of patients.

584.61 -> And what we're finding is that at the transcriptional level,

587.17 -> when you look at gene expression alone,

589.15 -> we can start predicting differences

591.16 -> between patient subgroups that also manifest,

594.1 -> that are visible at the phenotypic level.

596.92 -> So individuals that show differences in their cells

600.64 -> are also showing differences in their phenotype.

603.1 -> This is extremely important because for the cells,

605.53 -> we can actually start using molecular information

607.84 -> to predict the subtype of that individual.

609.86 -> We don't want to say, oh, please give us a sample of your brain,

612.07 -> and then we'll tell you which drug to take,

613.93 -> because people kind of like to hold on to their brain.

616.18 -> But instead, we can genetically start

618.22 -> predicting the impact of these genetic variants

620.71 -> and then start predicting which patient subgroup

624.52 -> everyone should belong in.

625.9 -> So we're using all this to now develop target genes.

629.11 -> And we're using next-generation technologies

631.99 -> for modulating these in programmable and modular ways

636.28 -> so that we can go in and activate or repress

638.8 -> individual regulatory regions or individual genes

641.41 -> or individual variants.

642.82 -> And we can carry this out systematically.

644.85 -> And we now have this amazing collaboration

646.6 -> that's using large language models to predict the protein

650.5 -> language, the language of protein folding,

655.09 -> the language protein function, the language of gene

657.94 -> expression through separate towers,

659.98 -> these foundational models, and then glue them together

662.77 -> in the same multidimensional embedding so that we can

665.8 -> translate seamlessly between the type of therapeutic

667.99 -> that you need for every type of function.

671.53 -> And we can use this to now transform medicine.

675.4 -> We can use this to enable personalized medicine,

677.98 -> to understand the drivers, to develop drugs specifically

681.61 -> for these drivers.

682.51 -> And this requires a call to action.

684.41 -> This requires a coalition across every single discipline

687.55 -> represented in this room, across computer science, biology,

690.98 -> chemistry, engineering, finance, as we heard earlier.

694.85 -> So the last place where AI is making a dramatic difference

698.53 -> is the discovery process itself.

701.09 -> So a small team in my lab has now

703.18 -> started working on this Idea Navigator that allows us to now

706.33 -> project every idea through every meeting

709 -> that we have across an entire organization, across years

712.45 -> of work, across multiple different teams

714.76 -> working on different aspects, and connect people together

717.37 -> when they're working on the same ideas.

719.26 -> At the top right, you can see an example

721 -> of embedding 150,000 papers that have cited our work.

724.09 -> At the bottom, 10,000 meetings that we've had that

727.21 -> have been automatically transcribed,

729.01 -> connecting them, dragging them together, and seeing

731.44 -> how individual ideas traverse through the space

734.62 -> of the work in our own lab.

735.997 -> And I'm hoping that you can use that for your own corporation.

738.58 -> And this one, we actually do have it

740.68 -> at the back of the room.

742.52 -> So I just want to thank this extraordinary team

745.45 -> of individuals, many of whom are here in the room,

747.7 -> for the opportunity to work on all these different topics.

750.26 -> And I want to thank John for putting together

752.135 -> just such an awesome meeting.

753.35 -> And thank you all for coming.

755.4 ->

Source: https://www.youtube.com/watch?v=JWrZgiKTxms