A scientist designing a new protein using computer software

Af­ter years of hype, the first AI-de­signed drugs fall short in the clin­ic

The first AI-de­signed drugs have end­ed with dis­ap­point­ment.

Over the last year-plus, the first hand­ful of mol­e­cules cre­at­ed by ar­ti­fi­cial in­tel­li­gence have failed tri­als or been de­pri­or­i­tized. The AI com­pa­nies be­hind these drugs brought them in­to the clin­ic full of fan­fare about a new age of drug dis­cov­ery — and have qui­et­ly shelved them af­ter learn­ing old lessons about how hard phar­ma­ceu­ti­cal R&D can be.

Ear­li­er this month, UK-based Ex­sci­en­tia slipped in­to a pipeline up­date that a Phase I/II study of its can­cer drug can­di­date EXS-21546 was wind­ing down. That cut fol­lowed a de­ci­sion last year by its part­ner Sum­it­o­mo Phar­ma to aban­don an­oth­er of its AI-de­signed drugs. In April, a test of Benev­o­len­tAI’s der­mati­tis drug fell short as well. And Re­cur­sion Phar­ma­ceu­ti­cals — the third of AI’s ear­ly gen­er­a­tion — hasn’t record­ed a tri­al fail­ure but has had a hand­ful of clin­i­cal set­backs that don’t nec­es­sar­i­ly bode well.

Patrick Mal­one

There’s no short­age of AI naysay­ers, and the 0-for-3 start sug­gests that AI hype has set un­re­al­is­tic ex­pec­ta­tions. Clin­i­cal wins are rar­i­ties in biotech, where an es­ti­mat­ed 5% or 10% of drugs that head in­to hu­man test­ing ac­tu­al­ly get ap­proved.

“If you take the hype and PR at face val­ue over the last 10 years, you would think it goes from 5% to 90%,” Patrick Mal­one, a prin­ci­pal at KdT Ven­tures, said of AI. “But if you know how these mod­els work, it goes from 5% to maybe 6% or 7%.”

These three com­pa­nies have been at this for rough­ly a decade, com­bin­ing to rack up an ac­cu­mu­lat­ed deficit of over $1.5 bil­lion.

Ivan Grif­fin

The re­al­i­ty check of the clin­ic, paired with a dour biotech mar­ket, has beat­en up these first-gen­er­a­tion biotechs that went pub­lic in 2021 or 2022. Their stock prices are all down at least 75%, un­der­per­form­ing the biotech mar­ket, even as new AI star­tups have con­tin­ued to raise sub­stan­tial sums of mon­ey. Gen­er­ate:Bio­med­i­cines, In­cep­tive, Iambic and Gen­e­sis, for in­stance, have com­bined to raise $673 mil­lion over the past few months.

Ex­ec­u­tives at these first-gen­er­a­tion com­pa­nies say it’s too ear­ly for a ver­dict on whether or not AI boosts the odds of suc­cess, par­tic­u­lar­ly giv­en the vast like­li­hood that any drug can­di­date — AI-de­vel­oped or not — will fail.

“The things that are eas­i­est to show — speed and cost, par­tic­u­lar­ly on the pre­clin­i­cal — have been done,” Ivan Grif­fin, Benev­o­len­tAI’s co-founder and chief op­er­at­ing of­fi­cer, said in an in­ter­view with End­points News. In­creas­ing the prob­a­bil­i­ty of suc­cess “will in­evitably take the longest to prove out.”

Mile­stone mo­ments to qui­et can­cel­la­tions

Back in the fi­nal days of Jan­u­ary 2020, as Covid-19 was emerg­ing as a pan­dem­ic threat, Ex­sci­en­tia CEO An­drew Hop­kins hailed a “key mile­stone in drug dis­cov­ery.”

His biotech an­nounced the first AI-de­signed drug had en­tered the clin­ic. Its part­ner Sum­it­o­mo led the Phase I study in ob­ses­sive-com­pul­sive dis­or­der, with the Fi­nan­cial Times call­ing the tri­al’s start a “crit­i­cal mile­stone for the role of ma­chine learn­ing in med­i­cine.”

About two years lat­er, in Jan­u­ary 2022, Sum­it­o­mo dis­closed they had aban­doned the drug, which failed to meet the study’s cri­te­ria. In an in­ter­view, Hop­kins said Ex­sci­en­tia’s job was to just de­sign the mol­e­cule, with Sum­it­o­mo mak­ing the clin­i­cal de­ci­sions.

Ex­sci­en­tia had more con­trol over the next drug, a can­cer treat­ment called EXS-21546, which it brought in­to the clin­ic in De­cem­ber 2020. Ear­li­er this month, the biotech said it was dis­con­tin­u­ing an on­go­ing Phase I/II study with mod­el­ing sug­gest­ing “it will be chal­leng­ing for ‘546 to reach a suit­able ther­a­peu­tic in­dex.”

An­drew Hop­kins

Hop­kins said that the tri­al didn’t fail, as the biotech doesn’t have full re­sults back.

“It wasn’t a clin­i­cal da­ta de­ci­sion,” he said. “It was a strate­gic de­ci­sion” to pri­or­i­tize two oth­er can­cer drugs that the com­pa­ny be­lieves have bet­ter chances.

“We don’t want to be one of those com­pa­nies that keeps push­ing a pro­gram for­ward be­cause it’s the on­ly thing they have,” Hop­kins said.

Clin­i­cal fail­ure is less con­testable for Benev­o­len­tAI, a fel­low UK-based AI en­thu­si­ast whose lead drug was un­able to beat a place­bo in a Phase IIa atopic der­mati­tis study, the biotech an­nounced ear­li­er this year. That read­out led to the drug’s end, a stock plunge, and siz­able lay­offs.

Re­cur­sion stands apart as the on­ly one of the three to main­tain a val­u­a­tion above $1 bil­lion to­day. (Ex­sci­en­tia is worth about $650 mil­lion, while Benev­o­len­tAI is val­ued at $117 mil­lion.)  The biotech has had sev­er­al pos­i­tive Phase I read­outs cen­tered on safe­ty and tol­er­a­bil­i­ty, such as a C. diff drug re­cent­ly clear­ing a healthy vol­un­teer study of 42 peo­ple. The com­pa­ny used AI to iden­ti­fy ex­ist­ing com­pounds rather than de­sign new drugs for its ear­ly pipeline.

Dy­lan Reid

“The mar­ket has nev­er known what to make of these com­pa­nies,” said Dy­lan Reid, a part­ner at Zetta Ven­ture Part­ners. “They’ve been way too ex­cit­ed and way too down. At one point, they val­ue the plat­form at X bil­lions of dol­lars, and to­day, it’s prob­a­bly a drag on val­u­a­tion.”

While Re­cur­sion hasn’t had a clin­i­cal fail­ure, its de­vel­op­ment plans have had hic­cups. It qui­et­ly dropped a rare dis­ease drug last year, cit­ing “noise in the po­ten­cy” and de­lays in get­ting a tri­al go­ing. Ear­li­er this month, the Salt Lake City-based biotech slimmed down an on­go­ing Phase II study for an­oth­er of its drugs, drop­ping a place­bo arm and cut­ting ex­pect­ed en­roll­ment to 37 peo­ple. A spokesper­son said the changes will help get to da­ta and Phase III faster.

If a decade sounds too soon to judge these biotechs, con­sid­er Re­cur­sion’s lead­ers set that time­line them­selves, pub­licly de­clar­ing the goal of dis­cov­er­ing 100 clin­i­cal-stage drug can­di­dates in its first 10 years. Rough­ly 10 years on, Re­cur­sion’s pipeline has four clin­i­cal-stage mol­e­cules.

Re­cur­sion ex­pects two Phase II read­outs in the sec­ond half of 2024. All three biotechs have on­go­ing part­ner­ships with drug­mak­ing gi­ants like Mer­ck, Bris­tol My­ers Squibb, and Roche’s Genen­tech.

The AI clin­i­cal pipeline is full of oth­er play­ers as well, such as Verge Ge­nomics’ ALS drug, BPG­bio’s brain can­cer treat­ment, In­sil­i­co Med­i­cine’s id­io­path­ic pul­monary fi­bro­sis drug, Gen­er­ate:Bio­med­i­cines’ Covid-19 an­ti­body, and Au­ransa’s liv­er can­cer ther­a­py.

Strate­gies evolve as more play­ers emerge

AI back­ers say suc­cess­ful pro­grams like Mod­er­na’s Covid-19 vac­cine or Nim­bus Ther­a­peu­tics’ TYK2 in­hibitor used AI to a de­gree. But those drug­mak­ers don’t brand those med­i­cines as AI-de­signed, while Ex­sci­en­tia, Benev­o­len­tAI, and Re­cur­sion mar­ket their ap­proach as AI-dri­ven or AI-en­abled.

A decade in, lead­ers of the first-gen­er­a­tion com­pa­nies say they are still learn­ing.

Benev­o­len­tAI, for in­stance, says its failed atopic der­mati­tis drug can­di­date didn’t use the com­pa­ny’s tar­get iden­ti­fi­ca­tion ap­proach, which is be­hind its ul­cer­a­tive col­i­tis drug that en­tered the clin­ic ear­li­er this year.

Ex­sci­en­tia has in­cor­po­rat­ed more hu­man tis­sue sam­ples in its re­search process and hired ex­pe­ri­enced clin­i­cal hands like Michael Krams, Hop­kins said.

“We’ve al­so now re­al­ized if we want to change the prob­a­bil­i­ty of suc­cess in the clin­ic, it’s not just bet­ter mol­e­cules,” Hop­kins said. “We al­so need bet­ter trans­la­tion­al mod­els.”


Andrew Dunn

Senior Biopharma Correspondent