Minkyung Baek (University of Washington)

'They have shown that this is not some im­pos­si­ble thing': Aca­d­e­m­ic lab copies Google’s big bi­o­log­i­cal break­through

When Demis Has­s­abis, CEO of Google’s AI out­fit Deep­Mind, an­nounced last year that they had cracked one of the tough­est puz­zles in bi­ol­o­gy — suc­cess­ful­ly pre­dict­ing a pro­tein’s shape from its amino acid se­quence — Minkyung Baek watched with a cu­ri­ous mix­ture of dread and ex­cite­ment.

“It felt like I just lost my job,” said Baek, a post­doc at the Uni­ver­si­ty of Wash­ing­ton’s In­sti­tute for Pro­tein De­sign.

Baek had spent the last three years try­ing to do the ex­act same thing, work­ing with what for years had been the lead­ing pro­tein de­sign tech­nol­o­gy at the lead­ing pro­tein de­sign lab. Now Deep­Mind had ful­ly eclipsed her. She felt the sci­en­tif­ic com­mu­ni­ty’s ex­cite­ment, even as she pon­dered her pro­fes­sion­al fu­ture.

Deep­Mind’s break­through, though, had a caveat: The com­pa­ny hadn’t ac­tu­al­ly shown how they cracked the puz­zle, nor had they made their head­line-grab­bing soft­ware avail­able to any re­searchers out­side of Google’s dis­parate of­fices. Baek won­dered if she could, with the few bread­crumbs Google left, re­con­struct Deep­Mind’s soft­ware and dis­trib­ute it to the world.

On Thurs­day, she showed she could. Baek and her team at UW’s Bak­er lab de­tailed ma­chine learn­ing soft­ware in Sci­ence that was al­most — al­beit not quite — as pow­er­ful as Deep­Mind in pre­dict­ing a pro­tein’s struc­ture from its se­quence, and demon­strat­ed how it could be de­ployed to probe ques­tions in­tri­cate­ly linked to un­der­stand­ing dis­ease and de­sign­ing new drugs. They have made the tool, known as RoseTTaFold, avail­able on GitHub, where UW claims it has al­ready been down­loaded by over 140 re­search teams.

John Moult

“If you want­ed to be neg­a­tive about it, you could say they’re play­ing catch-up and got re­sults that were not quite as good,” said John Moult, a com­pu­ta­tion­al bi­ol­o­gist at the Uni­ver­si­ty of Mary­land who in 1994 launched the an­nu­al chal­lenge where Deep­Mind de­buted their re­sults. “I think the more pos­i­tive and prop­er way of look­ing at it is that they have done it near­ly as well and they have al­ready pro­vid­ed a serv­er, which works at least in a cou­ple of times. And they have done a full re­lease of their code.

“They have shown that this is not some im­pos­si­ble thing for oth­er peo­ple to achieve,” Moult added.

The pa­per came out the same time Na­ture pub­lished the de­tailed meth­ods be­hind Deep­Mind’s work — a co­in­ci­dence the folks at the Bak­er lab chalk up to their de­ci­sion to re­lease the preprint and open source soft­ware last month. Has­s­abis’ team sim­i­lar­ly promised to make their soft­ware pub­lic, al­though they have not pro­vid­ed a sim­i­lar serv­er to Baek’s.

The two ap­proach­es are broad­ly sim­i­lar on a high lev­el, said Baek, re­ly­ing on broad datasets of known pro­tein se­quences and struc­tures and sim­i­lar­i­ties be­tween co-evolved pro­teins. But they dif­fer vast­ly in how they tech­ni­cal­ly car­ry out their vi­sion.

Col­lec­tive­ly, they give re­searchers two so­lu­tions to a decades-old prob­lem and of­fer im­prove­ments over the orig­i­nal re­sults re­leased last year, in­clud­ing the abil­i­ty to pre­dict struc­ture in min­utes or hours, rather than days.

A pro­tein’s func­tion is de­ter­mined by its struc­ture, but for decades the on­ly way to de­ter­mine that struc­ture was through a va­ri­ety of imag­ing tech­niques, such as X-ray crys­tal­log­ra­phy, that could be lengthy or dif­fi­cult and didn’t pro­duce ac­cu­rate de­pic­tions for every pro­tein. Al­though those tech­niques have im­proved over time, the po­ten­tial to pre­dict a a struc­ture from its se­quence alone — known as the fold­ing prob­lem — con­tin­ued to be a holy grail.

David Bak­er

Baek’s lab, run by bio­chemist David Bak­er and fa­mous for its work in de­sign­ing pro­teins from scratch, had been lead­ing the race for years be­fore Deep­Mind leapfrogged them. See­ing the com­pa­ny’s progress, they adopt­ed more ma­chine learn­ing tech­niques to catch up.

In their pa­per, Baek showed a few of the ways wide­spread ac­cess to such tech­nol­o­gy could be used to build new drugs. They pre­dict­ed struc­tures for three class­es of pro­teins key to a range of dis­eases, in­clud­ing can­cer and de­men­tia, show­ing how rare mu­ta­tions could warp the pro­tein’s shape and re­veal­ing pos­si­ble open­ings to tar­get drugs.

“It’s not as good as Deep­Mind,” said Jin­bo Xu, a com­pu­ta­tion­al bi­ol­o­gist at the Uni­ver­si­ty of Chica­go. “But some of the struc­tures are very ac­cu­rate. It will be use­ful.”

Un­til Deep­Mind re­leas­es their open-source work, RoseTTaFold, Xu said, will be more help­ful to the field. They may not have long to wait, though.

Baek post­ed their pa­per as a preprint on June 15. Three days lat­er, Has­s­abis tweet­ed out a “brief up­date” on Deep­Mind’s pro­tein fold­ing soft­ware, say­ing that a full pa­per out­lin­ing their meth­ods was un­der re­view and that they would be pro­vid­ing the source code and “broad ac­cess” to the soft­ware to re­searchers. Xu said a serv­er for re­searchers to ac­cess Deep­Mind’s tech would be the ide­al tool.

Moult not­ed the Bak­er pa­per took their soft­ware to places that Deep­Mind didn’t. In ad­di­tion to pre­dict­ing the struc­ture of in­di­vid­ual pro­teins, they al­so pre­dict­ed how dif­fer­ent pro­teins come to­geth­er to form larg­er struc­tures, as they do to per­form man­i­fold bi­o­log­i­cal func­tions. Baek pre­dict­ed the dif­fer­ences be­tween two im­mune com­plex­es: one formed by IL-23 and one by IL-12. The re­sults could help drug de­vel­op­ers iden­ti­fy mol­e­cules that block one but not the oth­er, cre­at­ing more pre­cise drugs for au­toim­mune dis­eases.

Baek said her next big project is im­prov­ing the soft­ware’s abil­i­ty to pre­dict these in­ter­ac­tions.

Adap­tive De­sign Meth­ods Of­fer Rapid, Seam­less Tran­si­tion Be­tween Study Phas­es in Rare Can­cer Tri­als

Rare cancers account for 22 percent of cancer diagnoses worldwide, yet there is no universally accepted definition for a “rare” cancer. Moreover, with the evolution of genomics and associated changes in categorizing tumors, some common cancers are now characterized into groups of rare cancers, each with a unique implication for patient management and therapy.

Adaptive designs, which allow for prospectively planned modifications to study design based on accumulating data from subjects in the trial, can be used to optimize rare oncology trials (see Figure 1). Adaptive design studies may include multiple cohorts and multiple tumor types. In addition, numerous adaptation methods may be used in a single trial and may facilitate a more rapid, seamless transition between study phases.

Marianne De Backer (L) and Jeff Hatfield

Bay­er nabs star biotech Vi­vid­ion with a $2B buy­out and an ‘arms-length’ pact, pulling a part­ner out of the IPO con­ga line

Vividion is canceling that IPO it filed. Instead of following the industry-wide migration to Nasdaq, the biotech that has captured considerable attention for its still-preclinical work finding cryptic pockets to bind to on proteins is going to work for Bayer now.

The pharma giant is putting out word today that it has bought out Vividion for $1.5 billion in cash and another half-billion dollars in milestones.

Endpoints News

Keep reading Endpoints with a free subscription

Unlock this story instantly and join 113,500+ biopharma pros reading Endpoints daily — and it's free.

Tadataka Yamada (Photographer: Kiyoshi Ota/Bloomberg via Getty Images)

Sci­ence pi­o­neer, phar­ma re­search chief, glob­al health ad­vo­cate and biotech en­tre­pre­neur Tadata­ka ‘Tachi’ Ya­ma­da has died

Tadataka Yamada, a towering physician-scientist who made his name in academia before transforming drug development at GlaxoSmithKline and developing vaccines for malaria and meningitis at the Gates Foundation, died unexpectedly of natural causes at his home in Seattle Wednesday morning.

He was 76. Frazier Healthcare Partners’ David Socks confirmed his death.

Known widely by the mononym “Tachi,” Yamada had a globetrotting career and arrived in industry relatively late in life. A 2004 Independent article noted GSK had asked Yamada to stay on beyond his approaching 60th birthday, the company’s usual retirement age. Yamada would continue working for the next 17 years, steering the Gates Foundation’s global health division for 6 years, funding Jim Wilson’s gene therapy work when few would touch it, launching Takeda Vaccines and co-founding a series of high-profile biotechs.

UK re-in­ves­ti­gates Pfiz­er's eye-pop­ping price goug­ing on an epilep­sy drug

When a drugmaker raises the price of a drug in the US by more than 2,000% overnight, and without any particular reason for that increase, nothing typically happens to the company. No fines, no court orders, just business as usual.

Martin Shkreli’s decades-old anti-parasitic drug Daraprim was the perfect example — massive price spike on an old drug, lots of media attention, public outcry, Congressional committees dragging his former company through multiple hearings, and at the end of it? Nothing happened to the price or the company (until generic competition came).

Thomas Lingelbach, Valneva CEO

A small vac­cine de­vel­op­er fa­vored by the UK gov­ern­ment in Covid-19 touts a PhI­II first in chikun­gun­ya

Before Valneva garnered the favor of the UK government as a potential supplier of Covid-19 vaccines, the French biotech prided itself on being the first company to bring a chikungunya vaccine into Phase III.

It now has positive pivotal results to back up the breakthrough therapy designation the FDA granted just weeks ago.

There are currently no approved jabs to prevent chikungunya virus infection despite decades of R&D efforts, a fact that underscores just how arduous traditional vaccine development can be, particularly for neglected tropical disease. In a absence of a major commercial market, the US government and NGOs such as CEPI have deployed various grants and incentives to spur on a small crew of academics and industry players, with Merck, via its acquisition of Themis, claiming a spot in that race.

Josh Hoffman, outgoing Zymergen CEO (Zymergen)

UP­DAT­ED: Syn­bio uni­corn Zymer­gen jet­ti­sons found­ing CEO, cuts guid­ance as cus­tomers re­port lead prod­uct does­n't work

Zymergen, just months off a $500 million IPO that put the synthetic bio firm in rarified air, has now ejected its founding CEO and downgraded its revenue forecasts after customers reported its lead film product doesn’t work as advertised, the company said Tuesday afternoon.

CEO Josh Hoffman will leave his role and sacrifice his board seat immediately in favor of Jay Flatley, the former CEO of Illumina who will take the lead role on an interim basis as the company conducts a search for its next leader.

Endpoints News

Keep reading Endpoints with a free subscription

Unlock this story instantly and join 113,500+ biopharma pros reading Endpoints daily — and it's free.

Zymergen co-founders Zach Serber, Josh Hoffman, and Jed Dean (Zymergen via website)

Zymer­gen's sud­den im­plo­sion shocked biotech. A lin­ger­ing loan could make things even worse

As former synbio unicorn Zymergen picks up the pieces from its spectacular implosion Tuesday, an outstanding loan from Perceptive Advisors — the only blue-chip biotech crossover investor to touch Zymergen’s fundraising efforts — could make the situation worse, according to public documents.

In December 2019, more than a year before Zymergen filed for what would eventually become a $500 million IPO, the “biofacturing” firm signed a $100 million credit facility with Perceptive to help supplement the nearly $700 million the company had raised across four VC rounds.

Bio­gen, Ei­sai are push­ing for an­oth­er ac­cel­er­at­ed Alzheimer's OK — this time for BAN2401

Now that the door at the FDA has been opened wide for Alzheimer’s drugs that can demonstrate a reduction in amyloid, Biogen and its partners at Eisai are pushing for a quick OK on the next drug to follow in the controversial path of aducanumab.

In a presentation to analysts, Eisai neurology chief Ivan Cheung outlined some bullish expectations for their newly-approved treatment and set the stage for what he believes will be a fast follow for BAN2401 (lecanemab) — after a dry spell in new drug development that’s lasted close to 20 years.

Endpoints News

Keep reading Endpoints with a free subscription

Unlock this story instantly and join 113,500+ biopharma pros reading Endpoints daily — and it's free.

Paul Hudson, Sanofi CEO (Eric Piermont/AFP via Getty Images)

UP­DAT­ED: Sanofi buys mR­NA play­er Trans­late Bio for $3.2B. And the price fits a pop­u­lar range for biotech M&A

Sanofi CEO Paul Hudson is dead serious about his intention to vault directly into contention for the future of mRNA vaccines.

A year after paying Translate Bio $TBIO a whopping $425 million in an upfront and equity payment to help guide the pharma giant to the promised land of mRNA vaccines, Sanofi closed the deal with a buyout early Tuesday, spending $38 a share in a $3.2 billion buyout.

Translate’s stock $TBIO soared after the market closed Monday when Reuters reported the first word of the acquisition just hours ahead of the formal announcement. The wire service, though, didn’t have a price to report in its scoop, and investors chased the stock up 78% in the wild ride that followed. Once the price was announced, gains shriveled to 29% ahead of the bell.

Endpoints News

Keep reading Endpoints with a free subscription

Unlock this story instantly and join 113,500+ biopharma pros reading Endpoints daily — and it's free.