Aviv Regev leads Genen­tech's next rev­o­lu­tion with AI

SOUTH SAN FRAN­CIS­CO  At Genen­tech, many of the com­pa­ny’s drug de­sign­ers are mak­ing mol­e­cules for a ma­chine that are, for lack of a bet­ter word, weird.

The new mol­e­cules don’t look like they’ll be bet­ter drugs. They don’t al­ways make sense. But they make the com­put­er hap­pi­er. And as Amer­i­ca’s old­est biotech rein­vents it­self to be at the fore­front of a ma­chine-learn­ing race, this odd mol­e­c­u­lar work is train­ing the al­go­rithms up­on which Genen­tech’s new lead­ers are bet­ting its fu­ture.

“Those might not look great for our mol­e­cule team,” said Aviv Regev, the com­pu­ta­tion­al bi­ol­o­gist who joined Genen­tech in 2020 to lead its re­search and ear­ly de­vel­op­ment unit. “They’re like, ‘This is nev­er go­ing to be a drug.’ Maybe it won’t. But it will make an al­go­rithm that will be good for all drugs.”

Regev is lead­ing the re­search re­design, which fo­cus­es as much on chips and com­pute pow­er as on chemists and com­pounds. Genen­tech rose up near­ly half a cen­tu­ry ago as a pi­o­neer in ge­net­ic en­gi­neer­ing, and Regev be­lieves the com­pa­ny can lead the next great R&D rev­o­lu­tion by har­ness­ing the grow­ing pow­er of ar­ti­fi­cial in­tel­li­gence.

Vir­tu­al­ly every phar­ma gi­ant has dipped their toes in­to AI. But Genen­tech has plunged in since hir­ing Regev, poach­ing aca­d­e­m­ic lu­mi­nar­ies, cre­at­ing labs, and ear­li­er this year es­tab­lish­ing a 400-em­ploy­ee com­pu­ta­tion­al sci­ences unit. The 52-year-old sci­en­tist left a dis­tin­guished ca­reer at the Broad In­sti­tute to move across the coun­try with the vi­sion of de­vel­op­ing a new way of do­ing R&D.

Thomas Schi­neck­er

“I plan to be here for the very long term,” Regev told End­points News. “There is no Plan B.”

Genen­tech, too, has tied its fu­ture to Regev, who over­sees a team of about 2,400. Suc­cess is cru­cial not just to Genen­tech’s fu­ture, but to its par­ent com­pa­ny, Roche. The Swiss phar­ma gi­ant has re­cent­ly bat­tled in­vestor con­cerns over a thin pipeline, as its stock has fall­en by about a third from an April 2022 peak. Roche CEO Thomas Schi­neck­er has pitched in­vestors on “evolv­ing the R&D en­gine” with a unique set­up of four re­search units op­er­at­ing in rel­a­tive au­ton­o­my. Genen­tech is one of those, along­side groups based in Switzer­land, Japan, and Chi­na.

Regev’s hir­ing “is a clear ex­am­ple of Roche’s firm be­lief in the pow­er and im­por­tance of AI and ML to dra­mat­i­cal­ly ac­cel­er­ate drug dis­cov­ery,” Schi­neck­er said in a state­ment.

***

Three years in, the biotech has bought and built its way to de­vel­op­ing dozens of ma­chine-learn­ing mod­els.

In 2021, the biotech ac­quired a New York-based AI start­up called Pre­scient De­sign, when it had just a few em­ploy­ees and one pro­tein-mod­el­ing al­go­rithm. Genen­tech has built up Pre­scient’s team to 80, which has de­vel­oped over 20 ma­chine-learn­ing mod­els tack­ling a range of R&D prob­lems.

A few months lat­er, Roche and Genen­tech paid $150 mil­lion up­front for a wide-rang­ing col­lab­o­ra­tion with Re­cur­sion Phar­ma­ceu­ti­cals. The com­pa­nies didn’t start from a list of drug tar­gets but in­stead are cre­at­ing bi­o­log­i­cal maps to bet­ter un­der­stand neu­ro­science and can­cer, which could pro­duce as many as 40 projects and over $12 bil­lion in mile­stone pay­ments.

And last month, Genen­tech signed a mul­ti­year deal with chip­mak­ing gi­ant Nvidia, tap­ping its su­per­com­put­er pow­ers to op­ti­mize its al­go­rithms faster.

While it will take time for the vi­sion to play out, Regev be­lieves these new ideas will ul­ti­mate­ly de­liv­er more and more trans­for­ma­tive drugs.

“We have put the frame­work in place, and we are tru­ly us­ing it for our ac­tu­al work,” she said. “We’re test­ing it on the re­al thing. It’s not a pa­per. It’s not just fun sci­ence, which I love and do. But it has to pass the re­al test.”

That frame­work is called the “lab in a loop.” The idea leaves be­hind tra­di­tion­al R&D ways, which Regev de­scribed as sim­ply too slow.

“If we pro­ceed with every­thing in this step­wise, be­spoke, al­most hand­craft­ed way, we can’t pro­ceed fast enough and on enough prob­lems,” Regev said. “We have to find ways to work at a dif­fer­ent scale.”

The loop puts ma­chine learn­ing at the heart of R&D. Al­go­rithms are trained on moun­tains of ex­per­i­men­tal da­ta, mak­ing pre­dic­tions that are then test­ed in labs. Re­sults are fed back in­to the mod­els to fur­ther im­prove them. The process re­peats over and over.

John Mar­i­oni

Com­pu­ta­tion­al and ex­per­i­men­tal sci­en­tists are work­ing to­geth­er, Regev said, with both learn­ing to think dif­fer­ent­ly about do­ing re­search. That leads to projects like de­sign­ing mol­e­cules that may seem use­less to vet­er­an drug hunters, but can train an al­go­rithm’s un­cer­tain eyes.

John Mar­i­oni, the for­mer re­search head of the Eu­ro­pean Bioin­for­mat­ics In­sti­tute, joined last year to lead the new­ly-formed com­pu­ta­tion­al sci­ences unit. The group has now reached “crit­i­cal mass” to start ap­ply­ing da­ta and mod­els to the most in­ter­est­ing sci­en­tif­ic ques­tions, he said.

“We are at this cusp of do­ing some­thing re­al­ly in­cred­i­ble with da­ta in the phar­ma­ceu­ti­cal sci­ences,” Mar­i­oni said. “It’s been re­al­ly ex­haust­ing on oc­ca­sion, but al­ways fun.”

***

Genen­tech’s pipeline isn’t yet full of new mol­e­cules from the ef­fort, which will be the true test of the bet’s suc­cess.

“You can­not judge suc­cess three years in, in this par­tic­u­lar field,” Regev said. “Un­for­tu­nate­ly, still, it’s too slow.”

But Regev said the strat­e­gy is im­pact­ing more and more pro­grams. A foun­da­tion­al mod­el, for in­stance, sug­gest­ed a new dis­ease to test a clin­i­cal-stage drug in, prompt­ing Genen­tech to launch a new tri­al, Regev said, with­out dis­clos­ing the ther­a­py. Pre­scient De­sign’s mod­els are “part and par­cel” of mul­ti­ple pre­clin­i­cal pro­grams as well, she added.

And some re­search part­ner­ships, like one with BioN­Tech that start­ed in 2016, are lit­er­al­ly let­ting al­go­rithms de­sign the drug. The two are de­vel­op­ing can­cer vac­cines, now in hu­man test­ing, where ma­chine learn­ing an­a­lyzes a sam­ple of a pa­tient’s tu­mor and de­signs an in­di­vid­u­al­ized shot for that per­son.

But re­search com­ing out of the com­pa­ny hints at its ear­ly progress: The team has de­vel­oped mod­els and writ­ten pa­pers on AI sys­tems that pre­dict­ed pro­tein struc­tures faster than Google Deep­Mind’s Al­phaFold, stud­ied tens of mil­lions of im­mune cells, and made huge CRISPR screens faster and cheap­er.

“Peo­ple think it will work im­me­di­ate­ly,” Regev said. “No, no, no. It’s a set of mul­ti­ple dif­fi­cult prob­lems, and you have to tack­le all of them, and you’re on­ly as good as your weak­est link. There are many is­sues like that com­bined to­geth­er, but if we don’t do it, what are we here for?”

AUTHOR

Andrew Dunn

Biopharma Correspondent