Demis Hassabis, DeepMind CEO

No­bel Prize win­ner Demis Has­s­abis talks about his vi­sion for Iso­mor­phic and AI’s fu­ture in biotech

In Sep­tem­ber, End­points News in­ter­viewed Deep­Mind’s Demis Has­s­abis. We’re re­pub­lish­ing the in­ter­view to­day af­ter Has­s­abis won the No­bel Prize in chem­istry, shared with David Bak­er and John Jumper.

While Demis Has­s­abis is best known for steer­ing Al­pha­bet’s re­search ef­forts in ar­ti­fi­cial in­tel­li­gence, the 47-year-old is si­mul­ta­ne­ous­ly lead­ing a se­cre­tive biotech start­up look­ing to bring about an AI rev­o­lu­tion in drug dis­cov­ery.

Has­s­abis is the CEO of Google Deep­Mind as well as Iso­mor­phic Labs, an Al­pha­bet-backed com­pa­ny found­ed in 2021 to take Deep­Mind’s AI re­search in bi­ol­o­gy fur­ther in­to the drug in­dus­try. Deep­Mind has won a myr­i­ad of awards for Al­phaFold2, its AI sys­tem that pre­dicts pro­tein struc­tures and was pub­lished in 2021. Head­quar­tered in Lon­don, Iso­mor­phic has shared few de­tails about what it’s work­ing on in its rough­ly two years of ex­is­tence.

In a rare, ex­clu­sive in­ter­view, Has­s­abis spoke with End­points News about his vi­sion for Iso­mor­phic and the fu­ture of AI in the life sci­ences. This is his first in­ter­view where he’s dis­cussed Iso­mor­phic, an End­points 11 re­cip­i­ent this year, in-depth. This con­ver­sa­tion has been sub­stan­tial­ly edit­ed and con­densed for length and clar­i­ty.

An­drew Dunn: How’d Iso­mor­phic Labs first get start­ed?

Endpoint News

All of the news, delivered with full-text to your inbox. For professionals discovering, developing, and marketing biopharmaceutical drugs.

Demis Has­s­abis: I’ve had in mind the idea AI could be used in a fun­da­men­tal way for drug dis­cov­ery for a very long time. I’d say it’s the num­ber one thing I al­ways want­ed to ap­ply AI to once it was so­phis­ti­cat­ed enough and pow­er­ful enough.

That’s ob­vi­ous­ly the big ques­tion: When is that mo­ment the right mo­ment? Through­out my start­up and en­tre­pre­neur­ial ca­reer, I’ve learned tim­ing is very im­por­tant. Even if you have a great idea, if you’re 10 years too ear­ly, you’re in for a world of pain.

What I was wait­ing for was a sig­nal that the tim­ing was right. That, for me, was Al­phaFold. I want­ed that to work, not on­ly to use as a foun­da­tion­al build­ing block for what we’re do­ing at Iso­mor­phic, but as a proof point that the AI meth­ods were pow­er­ful enough, so­phis­ti­cat­ed enough that ac­tu­al­ly we could make a se­ri­ous dent on this prob­lem and reimag­in­ing drug dis­cov­ery from first prin­ci­ples us­ing com­pu­ta­tion­al meth­ods, ob­vi­ous­ly in­clud­ing AI. So Al­phaFold2 was the mo­ment for me to press go on this.

Dunn: Be­fore get­ting to the sci­ence, how is Iso­mor­phic fund­ed?

Has­s­abis: It’s an Al­pha­bet com­pa­ny, of­fi­cial­ly, so it’s a sis­ter com­pa­ny of Deep­Mind. The idea is a stand­alone com­pa­ny. It takes ad­van­tage of be­ing part of the Al­pha­bet fam­i­ly, most­ly the vast amounts of com­pute we need if we’re go­ing to achieve our mis­sion. There’s close col­lab­o­ra­tions with Deep­Mind.

A lot of the cul­ture things are sim­i­lar to Deep­Mind. But it’s op­er­at­ing at a start­up speed and en­er­gy and in­ten­si­ty. All our fund­ing comes from Al­pha­bet to be­gin with. That se­cu­ri­ty al­lows us to go for the most am­bi­tious ver­sion of the op­por­tu­ni­ty, which is to build this plat­form that will hope­ful­ly rev­o­lu­tion­ize the drug dis­cov­ery process, as op­posed to just go af­ter one drug can­di­date or what­ev­er.

Dunn: How much fund­ing has Iso­mor­phic re­ceived?

Has­s­abis: I don’t think it’s pub­lic in­for­ma­tion, so I can’t give you ex­act num­bers. But it’s very well-fund­ed for the stage that it’s at.

Dunn: You’ve de­scribed the goal as build­ing more Al­phaFolds, par­tic­u­lar­ly in pre­dict­ing how pro­teins in­ter­act with oth­er pro­teins or small mol­e­cules, and in de­sign­ing small mol­e­cules. What have you been able to do since get­ting start­ed in No­vem­ber 2021?

Has­s­abis: I guess you’ve been fol­low­ing what I’ve said pub­licly quite care­ful­ly. It’s ex­act­ly that.

Al­phaFold is on­ly one small part of the whole drug dis­cov­ery process. There’s all these ad­ja­cen­cies. We are fo­cus­ing on small mol­e­cules first. We’re in­ter­est­ed in bi­o­log­ics too, but we think small mol­e­cules is where we can make the most dif­fer­ence most quick­ly.

We’re very in­ter­est­ed in the in­ter­ac­tion space: pro­tein-lig­and in­ter­ac­tions, pro­tein-pro­tein in­ter­ac­tions, and the dy­nam­ic na­ture of bi­ol­o­gy. Al­so go­ing in­to chem­istry space, un­der­stand­ing chem­i­cal com­pound struc­tures, how they might bind, bind­ing affini­ties, these kinds of prob­lems.

Just in that short sen­tence, I’ve out­lined years worth of re­search work and many ad­ja­cent prob­lems us­ing ad­vanced ver­sions of Al­phaFold as part of that mix. But we need many oth­er things, things like pre­dict­ing ADME prop­er­ties, these types of things.

We need to make an­oth­er half-dozen big break­throughs.

Dunn: That re­minds me of writ­ing about the jump from the first ver­sion of Al­phaFold to Al­phaFold2. A lot of re­work­ing from the ground up. How much of Iso­mor­phic’s vi­sion is based on ex­tend­ing Al­phaFold ver­sus build­ing new mod­els?

Has­s­abis: We’re al­ways do­ing both ac­tu­al­ly. We’re al­ways push­ing ex­ist­ing meth­ods to the lim­it. There’s usu­al­ly more juice that can be squeezed. But we’re al­so in­ves­ti­gat­ing new ideas, new ar­chi­tec­tures.

We’re ap­ply­ing dif­fer­ent tech­niques to these new do­mains. Al­phaFold wouldn’t do things like ADME prop­er­ties. That will be com­plete­ly new tech­nolo­gies.

But with the in­ter­ac­tion space, we’re work­ing hard on that in fun­da­men­tal ways with new ver­sions of Al­phaFold as well. There’s a ton of very in­ter­est­ing work go­ing on, both at the Deep­Mind sci­ence team and Iso­mor­phic. It’s com­ple­men­tary work. Iso­mor­phic is more ap­plied and more spe­cif­ic to drug dis­cov­ery, where­as Deep­Mind is more fun­da­men­tal re­search.

Dunn: How do you see Iso­mor­phic pro­vok­ing the drug in­dus­try’s sta­tus quo?

Has­s­abis: We’re pos­ing some very in­ter­est­ing ques­tions. One is how would drug dis­cov­ery look if you did it from an AI-first per­spec­tive. Not as an add-on. My im­pres­sion is, in tra­di­tion­al phar­ma, these are nice-to-haves. The chemists re­al­ly do the work and then you dou­ble-check things in ADME sim­u­la­tions or what­ev­er. We want to re­think that from the ground up.

That’s hard for Big Phar­ma to do be­cause they’re en­trenched in the ways of work­ing they’ve been do­ing for decades. It re­quires a new ap­proach, and that’s what we have.

If that’s true, can the drug dis­cov­ery process do most of the re­search in sil­i­co and leave the wet lab stuff to val­i­da­tion? That’s the the­sis, as op­posed to do­ing the search for the com­pound ex­per­i­men­tal­ly, which goes much slow­er. My dream is we could re­duce the time it takes by an or­der of mag­ni­tude, maybe, cost and time, and have a high­er per­cent­age suc­cess rate at the next stage if we’re cor­rect and can build these half a dozen more Al­phaFold-lev­el break­throughs.

Dunn: Are you sole­ly fo­cused on these fun­da­men­tal break­throughs? Are you build­ing a pipeline, part­ner­ing with in­dus­try?

Has­s­abis: We’ll have more to say on this I think lat­er in the year, but for now, what I can say is we’re open to both. We’re think­ing about our own in­ter­nal pro­grams. We would pick tar­gets and pro­grams that we think are unique­ly suit­ed to our tech­nol­o­gy roadmap.

We’re open to part­ner­ships with Big Phar­ma on in­ter­est­ing tar­gets. I can’t tell you any­thing pub­licly right now.

Dunn: How do you sum­ma­rize Big Phar­ma’s in­ter­est in work­ing with Iso­mor­phic? I see Big Phar­ma CEOs as gen­er­al­ly smart peo­ple, who know AI is part of the fu­ture even if they don’t un­der­stand it. What have your in­ter­ac­tions been like?

Has­s­abis: We’ve had tons of in­ter­est. Pret­ty much every Big Phar­ma com­pa­ny CSO has reached out to us at some point. We’ve had in­ter­est­ing dis­cus­sions.

I think it’s bi­modal: those that think it’s still far away and tra­di­tion­al meth­ods will con­tin­ue to be the way for­ward. Maybe Al­phaFold is a one-off. They all use Al­phaFold, by the way, and I think they found it use­ful, but they re­gard it more as a one-off tool ver­sus those who think that maybe this is the fu­ture and would be in­ter­est­ed to part­ner in some way. I’ve met both types.

Dunn: What’s your con­fi­dence lev­el that Al­phaFold is not a one-off, that this can be repli­cat­ed and re­pro­duced?

Has­s­abis: Over the years at Deep­Mind, I’ve be­come very good at try­ing to as­sess the dif­fi­cul­ty of a prob­lem ver­sus the ideas we have, the tal­ent we have, and I think it’s per­fect tim­ing. We’ve got the sweet spot just right. You want to be five years ahead of the pack, but not 50 years ahead.

Dunn: In the next three to five years, what does suc­cess look like for Iso­mor­phic?

Has­s­abis: I would love us to have drug can­di­dates in clin­i­cal tri­als. Al­so hav­ing an in­cred­i­ble plat­form and an en­gine that is pro­duc­ing dozens of these can­di­dates on a very reg­u­lar, short time­line ba­sis. Our ad­vanced can­di­date should be in clin­i­cal tri­als by then, so there’s ac­tu­al, phys­i­cal proof points that this is all work­ing. And the plat­form has be­come pret­ty ma­ture and very ca­pa­ble, maybe we’ve re­duced the dis­cov­ery time from years to months.

Dunn: Your back­ground con­verges be­tween AI and bi­ol­o­gy. Should more of the AI field fo­cus on bi­ol­o­gy, or is there enough at­ten­tion al­ready?

Has­s­abis: I think more at­ten­tion should be paid on these. And not just life sci­ences, but al­so AI for sci­ence in gen­er­al.

But life sci­ences has al­ways been one of my pas­sions and the num­ber one thing I al­ways want­ed to ap­ply AI to, to help with cur­ing dis­eases and un­der­stand­ing bi­ol­o­gy. In many ways, it’s the per­fect regime. It’s very hard to imag­ine how you could math­e­mat­i­cal­ly de­scribe a cell with a few equa­tions.

Any­one can hand-wave and just go, “Ap­ply AI to bi­ol­o­gy, it must work.” You prob­a­bly hear that all the time with com­pa­nies pitch­ing you. But when you drill in, most of the time it’s not re­al. It al­so may just be, for that ques­tion, you should just use nor­mal stats. There’s no need for AI, or AI can’t help with it.

That’s where the hype part comes in. Every­one just says they’re do­ing AI with bio, but in my opin­ion, very few peo­ple are, if you ac­tu­al­ly think about which teams are ca­pa­ble of build­ing an Al­phaFold from scratch. Ob­vi­ous­ly we gave it away, pub­lished it, but imag­ine we hadn’t done that and were do­ing it be­hind closed doors. What oth­er groups could have built such a mod­el? Very few, I think. More of that is need­ed.

AUTHOR

Andrew Dunn

Senior Biopharma Correspondent

Director of IT, Security

Viridian Therapeutics

Waltham, MA, USA