Atom­wise inks Chi­na deal as list of AI col­lab­o­ra­tions length­ens

The long list of ma­jor AI bio­phar­ma col­lab­o­ra­tions has got­ten longer as one of the first ar­ti­fi­cial in­tel­li­gence star­tups has inked its first deal in Chi­na.

San Fran­cis­co-based start­up Atom­wise has signed an agree­ment to de­vel­op tar­get­ed drugs with Han­soh Phar­ma­ceu­ti­cals, a deal that could ul­ti­mate­ly be worth up to $1.5 bil­lion. Han­soh is flush with cash af­ter a $1 bil­lion IPO on the Hong Kong ex­change in June.

The promise of ma­chine learn­ing to speed up pre­clin­i­cal work and save de­vel­op­ers mil­lions of dol­lars has led to a string of new col­lab­o­ra­tions be­tween a hand­ful of soft­ware star­tups and some of the biggest drug de­vel­op­ers and re­search in­sti­tu­tions in­clud­ing Mer­ck, As­traZeneca, J&J, Bris­tol-My­ers Squibb, Pfiz­er and Duke Uni­ver­si­ty School of Med­i­cine. They’ve agreed to work on re­search rang­ing from on­col­o­gy to chron­ic dis­ease.

Part of the swarm like­ly comes from the hype that pe­ri­od­i­cal­ly sur­rounds a new tech­nol­o­gy — and few words are buzzi­er right now in both tech and pop­u­lar cul­ture than “ar­ti­fi­cial in­tell­gien­ce” and “ma­chine learn­ing” — and that has con­cerned some key fig­ures in phar­ma­ceu­ti­cal de­vel­op­ment. But al­though it’s too ear­ly for the AI plat­forms to have brought a drug to mar­ket, ear­ly stud­ies have in­di­cat­ed there could be some­thing be­neath the buzz. That in­cludes last week’s land­mark study from In­sil­i­co in Na­ture Biotech­nol­o­gy, in which over 21 days the com­pa­ny found six mol­e­cules that could be po­ten­tial treat­ments for fi­bro­sis.

At its most ba­sic, ar­ti­fi­cial in­tel­li­gence works like this: You feed an AI sys­tem a vast num­ber of, say, im­ages of a cow and im­ages not of a cow, and you tell it which is which. With each im­age of a cow and not-cow, the AI de­vel­ops a more and more re­fined set of cri­te­ria for what con­sti­tutes a cow (even if that cri­te­ria is far dif­fer­ent from what a hu­man might give). Pret­ty soon it can very ac­cu­rate­ly rec­og­nize whether a new pic­ture has a cow or not. You can al­so do this with, say, an im­age of your mom. It’s how your iPhone’s fa­cial recog­ni­tion works.

And you can do this with a mol­e­cule.

Atom­wise works by what’s called “vir­tu­al screen­ing,” mean­ing it us­es its AI sys­tem to rapid­ly search data­bas­es for mol­e­cules that re­sem­ble what its part­ners are look­ing for. Its June part­ner­ship with Ukraine-based Et­a­mine, the world’s largest chem­i­cal sup­pli­er, gives it ac­cess to a data­base of bil­lions of com­pounds to scan. Atom­wise can scan 10-20 mil­lion per day, up from con­ven­tion­al com­put­er meth­ods that cap out at about 100,000. This lat­est deal with Han­soh will see the com­pa­ny de­sign and dis­cov­er drugs for 11 undis­closed tar­get pro­teins.

How­ev­er, the In­sil­i­co study that grabbed head­lines was for a slight­ly dif­fer­ent form of AI.

This new­er AI, on­ly put forth in 2014, goes fur­ther. Rather than rec­og­niz­ing a face, it can imag­ine a face (or, say, art). The idea In­sil­i­co is bet­ting on and get­ting close to prov­ing is that if it can imag­ine a face, it can imag­ine a drug. Ac­cord­ing­ly, these are called “gen­er­a­tive” net­works, as op­posed to the “con­vo­lu­tion­al” ones Atom­wise us­es.

We’ll use cows again for the mod­el. These new AIs ac­tu­al­ly con­sist of two sys­tems. Loaded with da­ta, the “gen­er­a­tive” one at­tempts to come up with an im­age of a cow. Then a sec­ond one, which is called the “dis­crim­i­na­tor” and works likes the tech de­scribed above, tells the gen­er­a­tive one if it got a cow or not. The gen­er­a­tor learns from the dis­crim­i­na­tor, which learns from the vast store of up­loaded in­for­ma­tion. You have a learn­ing feed­back loop that should even­tu­al­ly gets you a brand new pret­ty pic­ture of a cow.

In the In­sil­i­co study, they were search­ing for a new ty­ro­sine ki­nase in­hibitor for dis­coidin do­main re­cep­tor 1 (DDR1). The sys­tem was taught all DDRI lit­er­a­ture, a larg­er set of ki­nase in­hibitors, data­bas­es of med­i­c­i­nal­ly ac­tive struc­tures and a data­base of struc­tures that have al­ready been patent­ed. The re­sult? 30,000 can­di­date struc­tures, which the com­pa­ny then whit­tled down to 40. They pro­duced 6 of them in the lab, test­ed 2 of them on cells and one on mice.

Promi­nent sci­ence writer and not­ed skep­tic of biotech AI hype Derek Lowe damp­ened the ex­u­ber­ant head­lines, not­ing the DDRI is al­ready well re­searched (cre­at­ing an ide­al sam­ple size to train the neur­al net­works), the dis­cov­er­ies weren’t drugs but pos­si­ble drug tar­gets, and gen­er­al­iz­ing these tech­niques to oth­er drug ar­eas will take years and lots of cash. This ac­cords with a con­sen­sus view of the tri­al as a proof-of-con­cept. Still, he found it one of the most in­ter­est­ing pa­pers he had read on vir­tu­al screen­ing.

“The good news, though, is that there is no rea­son that vir­tu­al screen­ing can’t do great things, even­tu­al­ly,” he wrote in his blog, In the Pipeline. “We just have to get a lot bet­ter at it than we are now, and that’s as true as it was when I first heard about it in the mid-1980s.”

Da­ta Lit­er­a­cy: The Foun­da­tion for Mod­ern Tri­al Ex­e­cu­tion

In 2016, the International Council for Harmonisation (ICH) updated their “Guidelines for Good Clinical Practice.” One key shift was a mandate to implement a risk-based quality management system throughout all stages of a clinical trial, and to take a systematic, prioritized, risk-based approach to clinical trial monitoring—on-site monitoring, remote monitoring, or any combination thereof.

Pfiz­er's big block­buster Xel­janz flunks its post-mar­ket­ing safe­ty study, re­new­ing harsh ques­tions for JAK class

When the FDA approved Pfizer’s JAK inhibitor Xeljanz for rheumatoid arthritis in 2012, they slapped on a black box warning for a laundry list of adverse events and required the New York drugmaker to run a long-term safety study.

That study has since become a consistent headache for Pfizer and their blockbuster molecule. Last year, Pfizer dropped the entire high dose cohort after an independent monitoring board found more patients died in that group than in the low dose arm or a control arm of patients who received one of two TNF inhibitors, Enbrel or Humira.

Endpoints News

Keep reading Endpoints with a free subscription

Unlock this story instantly and join 98,900+ biopharma pros reading Endpoints daily — and it's free.

Covid-19 roundup: EU and As­traZeneca trade blows over slow­downs; Un­usu­al unions pop up to test an­ti­bod­ies, vac­cines

After coming under fire for manufacturing delays last week, AstraZeneca’s feud with the European Union has spilled into the open.

The bloc accused the pharma giant on Wednesday of pulling out of a meeting to discuss cuts to its vaccine supplies, the AP reported. AstraZeneca denied the reports, saying it still planned on attending the discussion.

Early Wednesday, an EU Commission spokeswoman said that “the representative of AstraZeneca had announced this morning, had informed us this morning that their participation is not confirmed, is not happening.” But an AstraZeneca spokesperson later called the reports “not accurate.”

Endpoints News

Keep reading Endpoints with a free subscription

Unlock this story instantly and join 98,900+ biopharma pros reading Endpoints daily — and it's free.

Pascal Soriot, AP

As­traZeneca CEO Pas­cal So­ri­ot sev­ers an un­usu­al board con­nec­tion, steer­ing clear of con­flicts while re­tain­ing im­por­tant al­liances

CSL Behring chief Paul Perreault scored an unusual coup last summer when he added AstraZeneca CEO Pascal Soriot to the board, via Zoom. It’s rare, to say the least, to see a Big Pharma CEO take any board post in an industry where interests can simultaneously connect and collide on multiple levels of operations.

The tie set the stage for an important manufacturing connection. The Australian pharma giant agreed to supply the country with 10s of millions of AstraZeneca’s Covid-19 vaccine, once it passes regulatory muster.

Endpoints Premium

Premium subscription required

Unlock this article along with other benefits by subscribing to one of our paid plans.

Jim Tananbaum, Foresite

Fore­site re­turns to the SPAC well, as in­vestors won­der how long the run can last

Six months after launching his first biotech SPAC, Foresite’s Jim Tananbaum has started a second. On Tuesday, the longtime life science investor filed to raise $100 million by selling 10 million shares of the blank check company FS development II.

It’s a quick return to Wall Street for Foresite, although other firms have moved quicker. Perceptive Advisors raised a $130 million SPAC in June and were back before the end of July to raise another $125 million. By that point, the firm was evidently nearing a deal for the June SPAC, which would announce a half-billion-dollar merger with Cerevel Therapeutics on July 30.

As­traZeneca scores new goal on the pipeline front, adding its first AI-gen­er­at­ed tar­get to the port­fo­lio

As more and more biopharmas develop artificial intelligence platforms, the drug discovery process is being reshaped to fit new goals on cutting down the prodigious amount of time, energy and money that go into a drug program. Now one of the most ambitious players in the drive to improve on ROI, AstraZeneca, is marking a milestone on that front by adding the first target generated by AI to its portfolio.

Adeno-associated virus-1 illustration; the use of AAVs resurrected the gene therapy field, but companies are now testing the limits of a 20-year-old technology (File photo, Shutterstock)

Af­ter 3 deaths rock the field, gene ther­a­py re­searchers con­tem­plate AAV's fu­ture

Nicole Paulk was scrolling through her phone in bed early one morning in June when an email from a colleague jolted her awake. It was an article: Two patients in an Audentes gene therapy trial had died, grinding the study to a halt.

Paulk, who runs a gene therapy lab at the University of California, San Francisco, had planned to spend the day listening to talks at the American Association for Cancer Research annual meeting, which was taking place that week. Instead, she skipped the conference, canceled every work call on her calendar and began phoning colleagues across academia and industry, trying to figure out what happened and why. All the while, a single name hung in the back of her head.

Endpoints Premium

Premium subscription required

Unlock this article along with other benefits by subscribing to one of our paid plans.

Mer­ck scraps Covid-19 vac­cine pro­grams af­ter they fail to mea­sure up on ef­fi­ca­cy in an­oth­er ma­jor set­back in the glob­al fight

After turning up late to the vaccine development game in the global fight against Covid-19, Merck is now making a quick exit.

The pharma giant is reporting this morning that it’s decided to drop development of 2 vaccines — V590 and V591 — after taking a look at Phase I data that simply don’t measure up to either the natural immune response seen in people exposed to the virus or the vaccines already on or near the market.

Endpoints News

Keep reading Endpoints with a free subscription

Unlock this story instantly and join 98,900+ biopharma pros reading Endpoints daily — and it's free.

Anthony Fauci, NIAID director (AP Images)

As new Covid-19 task force gets un­der­way, threat looms of vac­cine, mon­o­clon­al an­ti­body-re­sis­tant vari­ants

Hours before President Biden’s Covid-19 team gave their first virtual press conference, the famed AIDS researcher David Ho delivered concerning news in a new pre-print: SARS-CoV-2 B.1.351, the variant that emerged in South Africa, is “markedly more resistant” to antibodies from convalescent plasma and vaccinated individuals.

The news for several monoclonal antibodies, including Eli Lilly’s bamlanivimab, was even worse: Their ability to neutralize was “completely or markedly abolished,” Ho wrote. Lilly’s antibody cocktail, which was just shown to dramatically reduce the risk of hospitalizations or death, also became far less potent.