Demis Hassabis, Isomorphic Labs CEO

Ex­clu­sive: Al­pha­bet's Iso­mor­phic inks first phar­ma deals with Eli Lil­ly and No­var­tis for $83M up­front, $2.9B in mile­stones

Iso­mor­phic Labs, Al­pha­bet’s close­ly watched AI start­up, has signed its first phar­ma part­ner­ships, and will work with Eli Lil­ly and No­var­tis in deals that could be worth $3 bil­lion, CEO Demis Has­s­abis ex­clu­sive­ly told End­points News.

Since its found­ing in No­vem­ber 2021, Iso­mor­phic has quick­ly be­come a lead­ing fig­ure in the fast-grow­ing AI biotech space. The Al­pha­bet sub­sidiary is ap­ply­ing Google Deep­Mind’s bi­ol­o­gy re­search — par­tic­u­lar­ly its pro­tein-struc­ture pre­dict­ing mod­el Al­phaFold — to drug dis­cov­ery. Has­s­abis is CEO of Google Deep­Mind, where he over­sees Al­pha­bet’s AI ef­forts, as well as CEO of Iso­mor­phic, which now em­ploys about 90 peo­ple across its Lon­don head­quar­ters and an of­fice in Lau­sanne, Switzer­land.

These first phar­ma deals, an­nounced the day be­fore the JP Mor­gan Health­care Con­fer­ence gets un­der­way in San Fran­cis­co, are Iso­mor­phic’s “light­house part­ner­ships,” Has­s­abis said. Lil­ly will pay $45 mil­lion up­front, along with $1.7 bil­lion in po­ten­tial mile­stone pay­ments, to work on sev­er­al tar­gets. No­var­tis paid $37.5 mil­lion up­front with $1.2 bil­lion in mile­stones to work on three tar­gets. While the com­pa­nies are not dis­clos­ing tar­gets, both col­lab­o­ra­tions aim to dis­cov­er small mol­e­cules against hard-to-drug tar­gets.

“They’re not pi­lot projects. They’re re­al drug tar­gets,” Has­s­abis said in an in­ter­view. “We want to do things that ul­ti­mate­ly, if suc­cess­ful, are re­al­ly mean­ing­ful, ac­tu­al­ly cure dis­eases and lead to block­buster drugs, rather than just pi­lots and sort of aca­d­e­m­ic work.”

Iso­mor­phic will try vir­tu­al screens, mol­e­cule de­sign on chal­leng­ing tar­gets

Iso­mor­phic has had pre­lim­i­nary talks with a range of large drug­mak­ers, Has­s­abis said, but Lil­ly and No­var­tis were the most en­thu­si­as­tic about get­ting in­volved at “ground ze­ro.” Daniel Skovron­sky, Eli Lil­ly’s chief sci­en­tif­ic and med­ical of­fi­cer, and Fiona Mar­shall, No­var­tis’ pres­i­dent of bio­med­ical re­search, were in­volved in the deals.

Mar­shall first heard about Al­phaFold when Deep­Mind’s team won a 2018 pro­tein-fold­ing com­pe­ti­tion. As Iso­mor­phic got start­ed, Mar­shall met with Has­s­abis and Iso­mor­phic’s chief sci­en­tif­ic of­fi­cer Miles Con­greve, Mar­shall’s for­mer col­league at So­sei Hep­tares. She called the part­ner­ship a “nat­ur­al fit,” as No­var­tis is al­so work­ing with oth­er tech­nol­o­gy gi­ants like Palan­tir on da­ta stor­age and Mi­crosoft on gen­er­a­tive chem­istry.

Mar­shall asked the dis­ease area teams at No­var­tis what tar­gets they were strug­gling with to try out Iso­mor­phic’s ideas.

“These are not easy tar­gets that we’ve picked at all,” she said in an in­ter­view with End­points. She said they are look­ing at pro­teins where the struc­ture isn’t yet solved or chemists can’t find the right type of chem­i­cal mat­ter.

Iso­mor­phic will work to find mol­e­cules that hit those tar­gets by vir­tu­al­ly screen­ing through com­pounds us­ing high-per­for­mance com­put­ing. Mar­shall said a con­ven­tion­al screen typ­i­cal­ly can cov­er 1 mil­lion com­pounds, while a vir­tu­al screen can con­sid­er 20 bil­lion. A sec­ond ap­proach that Iso­mor­phic plans to ex­plore is ab ini­tio de­sign, or de­sign­ing an ide­al mol­e­cule from scratch to fit in­to a pro­tein’s pock­et.

Has­s­abis called the two meth­ods — vir­tu­al screens and ab ini­tio de­sign — the “ob­vi­ous low-hang­ing fruit” for its tech­nol­o­gy, with promis­ing ini­tial re­sults.

“The proof will be in the pud­ding, which is why I think we’re ready to take on re­al drug projects,” he said.

Iso­mor­phic has al­so pro­gressed Al­phaFold to do more than pre­dict sta­t­ic pro­tein struc­tures. Pro­tein struc­tures, by them­selves, are “not that help­ful for drug de­sign,” Mar­shall said, and there’s much more val­ue in pre­dict­ing how a small mol­e­cule binds to a par­tic­u­lar pro­tein. The next gen­er­a­tion of Al­phaFold can pre­dict these com­plex struc­tures. Has­s­abis called it the “best co-fold­ing sys­tem in the world, both for place­ment of where lig­ands bind and pre­dict­ing bind­ing affin­i­ty.”

A holy grail for the field is link­ing the struc­ture of these mol­e­cules to func­tion. Mar­shall called that “the next lev­el to go to” for the tech­nol­o­gy.

“I don’t think we’re ac­tu­al­ly there yet,” Mar­shall said. “All you’re mea­sur­ing at the mo­ment is just the bind­ing affin­i­ty. That doesn’t re­al­ly tell you what it’s do­ing to the func­tion of the pro­tein.”

Iso­mor­phic is al­so work­ing on oth­er mod­els, still un­der wraps, that could help pre­dict cer­tain prop­er­ties in com­pounds like phar­ma­co­ki­net­ics and po­ten­tial tox­i­c­i­ty.

Fiona Mar­shall

“The low-hang­ing fruit, which we def­i­nite­ly think Iso­mor­phic is go­ing to be able to do, is re­duce the time from when you start with a tar­get to get your can­di­date mol­e­cule,” Mar­shall said in an in­ter­view. “On av­er­age, that’s three or four years us­ing con­ven­tion­al ap­proach­es. I think it’s rea­son­able to think you could halve that amount of time.”

Over­all, Mar­shall said she be­lieves AI will help with sev­er­al, but not all, of the steps in the lengthy process of drug dis­cov­ery.

“It isn’t go­ing to work mir­a­cles,” she said. “Maybe it can make 10 years in­to six years. I don’t think it’s go­ing to make 10 years in­to two years.”

Gen­er­al vs. spe­cif­ic mod­els for drug de­sign

The top goal for these col­lab­o­ra­tions is ob­vi­ous: mak­ing drugs. But Iso­mor­phic has a sec­ondary aim in mind.

Has­s­abis said he’s in­ter­est­ed in learn­ing just how much fine-tun­ing its mod­els will re­quire to work against spe­cif­ic drug tar­gets. He sees a slid­ing scale for these mod­els, rang­ing from the most gen­er­al­iz­able sys­tems to the most spe­cif­ic. The No­var­tis and Lil­ly part­ner­ships will test just how much a gen­er­al mod­el can achieve, and where it re­quires feed­ing in and train­ing mod­els on spe­cial­ized da­ta for a cer­tain drug tar­get.

“I don’t know what the op­ti­mal bal­ance is be­tween build­ing the gen­er­al sys­tem and do­ing spe­cial­ized sys­tems around the spe­cif­ic da­ta or the spe­cif­ic pro­gram,” he said. “At the mo­ment, I’m imag­in­ing it’s 50-50, but we’ll soon find out. Ide­al­ly, I’d like it to be 80-20 to­wards the gen­er­al, be­cause that’s more scal­able for us as a com­pa­ny.”

Has­s­abis said he hopes to an­swer those ques­tions with these first part­ner­ships, be­fore strik­ing more deals.

“For now we’re good, but we’d al­ways be open to dis­cus­sions and new part­ner­ships, es­pe­cial­ly if they pushed things in an in­ter­est­ing new way we didn’t al­ready have cov­ered,” he said.

Be­yond these col­lab­o­ra­tions, Has­s­abis said his aims in 2024 are for Iso­mor­phic’s ma­chine-learn­ing team to keep im­prov­ing the ac­cu­ra­cy of its pre­dic­tive sys­tems, while the com­pa­ny al­so pro­gress­es its in­ter­nal pipeline. De­tails about its own drug pro­grams re­main un­der wraps, but Has­s­abis said the biotech has short­ened a list of about 20 tar­gets down to a hand­ful it’s now go­ing af­ter. The com­pa­ny is us­ing con­tract re­search or­ga­ni­za­tions to make and test po­ten­tial com­pounds.

“We’re still at the ear­ly stages of test­ing out the pre­dic­tions of the plat­form,” Has­s­abis said. “How re­li­able are they? Once we’re feel­ing good about that, then we can re­al­ly go to town in push­ing an in­ter­nal project re­al­ly hard.”


Andrew Dunn

Senior Biopharma Correspondent