*“We're going to look back on history and realize that that was one of the pivotal moments that changed our industry,” says Absci founder and CEO Sean McClain, “when we put a de novo antibody that was designed on a computer in humans.”*

The world was dazzled last month by the debut of an artificial intelligence program called ChatGPT, from the Microsoft-backed startup OpenAI. Posing questions in natural language via the keyboard, a person can prompt GhatGPT to give a full-paragraph answer to a factual question, such as, When did people first land on the moon?

But the same program can spit out endless reams of text, fulfilling much more ambitious queries, such as, Write a poem in the style of Walt Whitman about scuba-diving in Paris.

The wide-open nature of GPT is at the heart of its intrigue. The program seems to have such a broad nature that it prompts one to imagine all sorts of potential applications.

One afternoon this month, at the Manhattan satellite offices of Absci, an eleven-year-old, promising biotech firm, ChatGPT was being put to good use.

“We had a board meeting this week, and I needed something to open up with,” says Sean McClain, founder and CEO of Absci.

“We used ChatGPT to write a rap,” explains McClain, standing amidst the cubicles where the company’s AI team works.

“We asked it, Write a rap on AI drug discovery for biologics.”

“Here we go,” says McClain, nodding to a staffer who starts a beat track on his phone. McClain gives himself a three-count, then launches into the rap with gusto, swaying as he recites each line displayed on the screen in appropriate hip-hop swagger.

AI in the lab, making moves with biologics
Fast and precise, our algorithms are prolific
Discovering new drugs with efficiency and speed
Leaving our competitors in the dust, with no time to bleed

The full poem goes on for several stanzas, as you can see in the accompanying sidebar. The staff gathered around on this Tuesday afternoon are delighted with their boss and give him a round of applause when he’s done.

“ChatGPT wrote that, it’s incredible,” McClain marvels.

He will later inform me, in a follow-up email, that his board of directors, at dinner the next day, upon hearing his rap, “suggested I stick with my day job rather than moonlighting as a rap superstar.”

ABSI stock chart by TradingView But that’s alright. The playful diversion is merely meant as an illustration of what McClain believes is being heralded for life sciences by the increasing power of AI programs.

ChatGPT is just one example of a broad class of AI programs known as “generative AI,” applications that don’t just label pictures of cats and dogs, but that instead produce a new artifact in the world.

Today, it is rap lyrics, but down the road, says McClain, generative AI will generate formulas for novel biologics, or drug products, with highly specific powers tuned to the biology of illness.

“The exciting part is being able to be the first company to put a biologic, or an antibody, into humans that was designed on a computer,” says McClain.

“We're going to look back on history and realize that that was one of the pivotal moments that changed our industry, when we put a de novo antibody that was designed on a computer in humans.”

AI’S BROKEN PROMISES

That is a bold statement about drug discovery, and as a journalist, I’m wary. AI in life sciences has been a long time coming, and I’ve heard many claims over the years about its potential. Following McClain’s rap, we sit down in the conference room, and I have a chance to convey my skepticism.

> Write a rap on AI drug discovery for biologics.

OUTPUT:

AI in the lab, making moves with biologics
Fast and precise, our algorithms are prolific
Discovering new drugs with efficiency and speed
Leaving our competitors in the dust, with no time to bleed

(Chorus)
AI drug discovery, for biologics
Revolutionizing the way we treat our illnesses
With precision and power, we're making waves
In the world of science, we're the crafters of new days

Our machine learning models are top of the line
Finding cures for diseases, one at a time
We're changing the game, with our cutting-edge technology
Making strides in healthcare, for the betterment of humanity.

(Chorus)
AI drug discovery, for biologics
Revolutionizing the way we treat our illnesses
With precision and power, we're making waves
In the world of science, we're the crafters of new days

(Bridge)
Gone are the days of trial and error
We're using AI to find the right answer
Bringing hope to those who need it most
With our cutting--edge technology, we're taking the reins and we're on the rise

(Chorus)
AI drug discovery, for biologics
Revolutionizing the way we treat our illnesses
With precision and power, we're making waves
In the world of science, we're the crafters of new days

Journalists, including myself, have been like the boy who cried wolf, promising trusting readers about amazing breakthroughs with AI just around the corner.

And yet, to date, no drug has emerged from Phase III clinical trials that can uniquely trace its success to AI. None of the supposed efficiency benefits of AI have changed the economics of the industry.

Instead, the marketplace continues to see upwardly spiraling costs for drugs. Take the recently approved Hemgenix, from the U.S. pharma company CSL Behring of King of Prussia, Pennsylvania. Designed to restore precious clotting proteins to hemophiliacs, Hemgenix is the most expensive drug in the world, at three and a half million dollars per dose, according to Nature.

Nor has AI solved the extraordinarily low success rate of drugs that get approved, about four percent of those that are attempted. Even the success rate for Phase I or II trials is stuck at about eighteen percent. For successful drugs, the time frame is still stuck at a decade or more from basic chemistry to Phase III trials.

It feels as if the promise of AI in life science, if not a broken promise, is one that has been extraordinarily over-hyped.

As I recite my chastened view, McClain nods. “I think that we are in the early innings,” he says. “We have shown fundamental advancements in the space, and actually showing, yes, this technology does what they say it can do.”

“That should translate into increased clinical success — but we're not there yet.”

The technology breakthrough to which McClain alludes is encapsulated in a research paper posted by he and his team on the free bioRxiv pre-print server in August. The paper, which has not yet been peer-reviewed, describes how the company used a neural network to predict whether a protein with a certain pattern of amino acids would be more or less likely to “bind” to an antigen — in lay terms, how likely it would be for the protein antibody to attack the pathogen in the body.

Like ChatGPT, the neural network in question is a generative AI program, in this case one introduced in 2019 by scientists at Facebook called RoBERTa, which anyone can grab off the shelf and play with.

The paper showed two remarkable results. One, RoBERTa’s predictions of which antibodies would bind were “highly accurate” compared to what could be measured in the lab by actually observing binding under a microscope. In other words, you could run the experiment on the computer instead of a lot of the lab work, potentially a huge time and materials saving.

More dramatically, McClain and team asked the neural network to invent novel combinations of amino acids by altering sections of known antibodies to create variants. Again, RoBERTa predicted how these new variants would bind and, again, the predictions of the machine were highly accurate compared to what a lab results showed.

The point of that second step is that the computer using AI can run many more explorations of possible amino-acid variants than can be run through a lab where the actual assays, even “high-throughput,” take tremendous time and care to prepare.

Similar to how ChatGPT spits out whole stanzas of poetry, the Absci program is able to spit out reams and reams of amino acid variants. In the paper’s results section, McClain and team declared, “Deep language models can expand the search space of an experimental dataset by orders of magnitude.”

As McClain explains it to me, “For drug discovery, it’s essentially saying, create me a drug that has these attributes — that is the future.” Pharma, he says, then goes “from drug discovery to drug creation, where you’re actually using AI to create novel drugs that don’t exist.”

FROM WET LAB TO AI

It has been a long time coming to this point. McClain founded Absci eleven years ago after graduating a year early from the University of Arizona, where he began as a mechanical engineer but then switched to molecular biology. His focus back then was not on AI, but on the art of protein expression.

McClain’s accomplishment at the start of Absci was a refinement of the mechanism by which E. coli cell lines in the lab can be made to produce enormous volumes of proteins. The cells become like little factories for producing custom proteins that a drug maker would want, such as monoclonal antibodies that can fight viruses.

That was a breakthrough in scale. A typical test tube of animal cells would produce merely thousands of antibodies. With McClain’s approach, “You could basically take a single test-tube of our engineered E. coli, take a billion different antibody sequences, and in that single test-tube you have a billion different drug candidates.”

Absci has patents and patent applications on that protein synthesis with McClain as the lead author. Because it’s rather like a production line for proteins, in an article about Absci for ZDNet earlier this year, I dubbed McClain the Elon Musk of protein manufacturing, a moniker he told me gave him some amusement.

With the ability to mass-produce proteins, McClain’s company was waiting for the proper vehicle to exploit that laboratory capability. The arrival of “deep learning” forms of artificial intelligence appeared on the scene as a perfect complement.

Deep learning is generally data hungry. “There wasn’t enough data” with the mammalian cells of traditional protein research, observes McClain. “We solved that problem,” he says. A billion proteins in a test-tube, rather than thousands, meant there was suddenly enough training data to fit the power of those neural networks like GPT-3.

“It was basically eleven years at Absci developing the wet lab technology that would allow us to leverage generative AI,” he says.

“The craziest part is, I had no idea that E. coli was going to be the key to unlocking data for biologics with genitive AI.”

Shortly before going public in July of last year, Absci bought another startup, Denovium, a three-year-old firm pioneering deep learning AI in medicine. The company was using AI to tease out novel proteins from DNA sequences.

THE FEEDBACK LOOP

Hitched to McClain’s protein factory, the Denovium AI becomes a way engineer a kind of feedback loop. One first manufactures tons of proteins, then sends the proteins into the AI software, as symbols of amino acid chains, and out come predictions from the program about binding. Those predictions are then sent to the wet lab to be validated in the test tube. The lab validation then becomes further data for the AI programs, and the process starts all over again.

Absci as a company functions a little like a feedback loop. The main headquarters where McClain built his wet lab is in Vancouver, but he flies out once a week or so to New York, where he maintains an apartment, to visit the AI hub on the 43rd floor of 152 West 57th, an imposing modern high-rise next to Carnegie Hall where we are having our chat.

What is discovered in the AI hub in New York becomes the input to the wet lab’s test tubes in Vancouver.

“The fact that we can go from wet lab data, to training the [AI] models, in a six-week time period, that’s what’s allowed us to make these huge advancements in a short amount of time,” says McClain. Other companies have wet labs, such as Recursion Pharmaceuticals, but that is for small-molecule drugs, not large, complex molecules such as antibodies, McClain points out.

The feedback loop has attracted some of the top talent in the field, including AI lead Joshua Meier, who was previously with Facebook’s AI team. Another top AI scientist recently joined from Tesla’s self-driving team.

“They joined because we spent ten years building that feedback loop, and that’s an advantage now no one else has.”

That tight coupling, says McClain, puts the company ahead of other firms that don’t have a lab, that just use software and data.

“A lot of these papers that are coming out, they say, Hey, we developed this computational metric that got improved,” he says of competing AI efforts. “But they never showed that it worked in the lab.”

The lab is important for ruling out false positives, he says.

"Because we have that six-week cycle time that we have and no one else has, that's what's really allowed us to figure out what directions are important to go in.”

“Sometimes, we see the industry going in a certain direction on a metric that they think is important,” he says, “and we find out that that metric doesn't actually correlate to the wet lab and what we actually want it to do.”

Eventually, says McClain, it will be possible to do most of the work on the computer, in silico, as it’s called. He expects to get there, but it won’t magically happen overnight. It will happen as a progression.

“Ultimately, you're still going to want to validate everything you do” in the lab, he says, "but you're not going to have to validate it to the extent we do now.”

The AI software program developed by Google’s DeepMind program, AlphaFold, is, of course, an important advance in what can be done on the computer. AlphaFold has essentially solved the problem of how proteins fold, the problem of structure, in other words.

Structure is important, but it won’t solve the problem of an antibody binding to an antigen, says McClain.

“Structure is a part of the solution, it’s not the solution,” says McClain. “At the end of the day, you don’t actually care about the structure, you care about, Does the antibody bind to where I want?

“AlphaFold won’t tell me this structure will bind to this target at the affinity I want,” he says. “AlphaFold isn’t telling you how to design the right drug.”

TURNING THE TIDE FOR DRUG DISCOVERY

To my point about the unmet promise of AI, McClain is convinced the feedback loop will dramatically change the success rate in drug development.

“There's never been a technology out there that's been able to, let's say, take a brand new target and be able to design antibodies that hit every single epitope,” meaning, the part on the antigen to which the antibody has to attach. “And, then, to be able to go instantaneously into the lab and say which of these gives me the biology I want to achieve.”

“This is the huge game changer — boom! You can instantaneously know what’s going to give you the biology that you want.”

What’s more, not only binding but also other qualities can be predicted at the same time. In the August paper, the research showed that RoBERTa could predict what Absci has christened “naturalness.” Naturalness means how close is an antibody to naturally occurring antibodies. Greater naturalness can make an antibody easier to produce, and make it more effective against a target in practice.

*“You’re not going to have this iterative traditional drug discovery process that takes years and, ultimately, gets sub-optimal hits,” says McClain. “We can get everything right the first time and dramatically reduce the time it takes to get into the clinic.”*

By divining both the binding ability of a protein, and its naturalness, “You're no longer having to sacrifice different attributes for each other, and, kind-of, taking suboptimal hits,” says McClain. “You're able to take the optimal hit the first time.”

In AI, the ability of something like ChatGPT to spew out rap lyrics the first time you ask it, without practice, is called “zero shot.” Effectively, McClain is saying that his company’s AI models, in conjunction with the wet lab, will get so good, they’ll be zero shot at generating a good antibody.

“Again, it’s just like how I told ChatGPT to write a rap on drug discovery; we’re going to be able to do that same thing for biologics, feed in the target sequence and have the AI then give us an antibody with all the attributes we want.

“You’re not going to have this iterative traditional drug discovery process that takes years and, ultimately, gets sub-optimal hits,” he says. “We can get everything right the first time and dramatically reduce the time it takes to get into the clinic.”

The normal time to get from chemistry to clinic is four years. “We believe we can get that to about eighteen months,” says McClain.

Following on the success of the paper posted in August, McClain expects that “soon, very soon here, we're going to be releasing where we sit on on the de novo design,” he says, meaning, tailoring a drug from scratch. That may come at an investor conference, he says.

“Things are accelerating faster than we had anticipated,” says McClain.

WORKING WITH BIG PHARMA

Startups don’t generally do their own drug development, and Absci is partnering with multiple drug giants to take its AI and wet lab into the clinic, the most prominent partnership being with Merck.

“Our goal is to be in the clinic in 2024,” says McClain.

To do so, McClain has lured star talent.

A third hub for the company is in Zug, Switzerland, south of Zurich, where pharmaceutical luminary Andreas Busch runs the company’s “innovation center.”

Busch had been on the board for four months when McClain said, “We need you full time” to oversee the company’s work with the drug makers.

Busch has the important duty of bringing traditional Swiss cookie samplings with him on his trips to the New York office, which he puts out in the common area for all to enjoy. He finds New York fascinating, he tells me, but confesses the pace and complexity make him happy to return to the sleepy terrain of Zug.

He also has the important duty of being a steady hand who’s seen the full cycle of drug development. The sleepy Zug canton is, in fact, the locale of many Big Pharma companies, and Busch has helped to run R&D at many of them, including Sanofi, Bayer, and Shire.

“It’s incredible that we even landed Andreas,” says McClain. “He is one of the most prolific R&D, large pharma executives in the industry, I mean, he’s gotten over ten drugs approved, all the way from the bench, which, I think, is more than any other large pharma exec.”

*McClain with chief innovation officer Andreas Busch, center, and chief AI officer Joseph Meier, in the company’s New York satellite office. The unique assets such as the wet lab have been a big factor attracting “the best of the best” in talent, says McClain.*

EXPLAINING THE ODDS TO INVESTORS

All of this research has to come to market, and what expectations to set with investors is a complex matter. Absci is growing very fast off of a very small base of revenue. The Street models sales doubling to a little over nine million dollars this year, and almost doubling again next year to eighteen million.

Absci releases quarterly press releases, but it has not held the traditional conference call with Street analysts since coming public. The stock is covered by a handful of analysts including Credit Suisse, Cowen & Co., and Stifel Nicolaus.

“We didn’t want to set a precedent of doing it [conference calls] because we’re not an earnings story,” says McClain.

All investors, including Fidelity Investments, the second-largest holder, says McClain, know that “It’s going to take time for revenue to ramp up.” In the breach, the right metric to watch is the company’s programs with Big Pharma. Those programs pay out in multiple ways, starting with up-front payments, followed by payments for milestones achieved, followed by, someday, royalties, assuming a drug succeeds.

So far this year, Absci is ahead of its intended goal of signing eight programs with pharma companies, having achieved ten, including three with Merck. The Merck deals, which carry the option of collaborating on three different drug targets, are valued at $610 million in milestone payments and eventual royalties. Meaning, up-front payments and milestone payments to Absci would be about $200 million per drug.

“Building up that portfolio [of programs], things get more advanced, and that's when you start getting that cascade of large milestone payments being hit, and ultimately, ramp up to royalties,” explains McClain.

The company doesn’t say how much the up-front payments are that are baked into each deal. “I will say it is a significant payment that definitely covers the cost of the work that will be done.” Revenue of $2.4 million in the most recent quarter was mostly from milestone payments by Merck, the company has said.

Given that it can take time for the total $610 million value of something like the Merck deals to be realized — if ever — Absci’s CFO, Greg Schiffman, will point out at investor meetings that the net present value, the discounted cash flows, of future deals is in the neighborhood of $15 million to $20 million per program.

“The way to think about it is, if we were today to go to a royalty farm that buys royalty streams, they would pay you today $15 million to $20 million for that particular program,” says McClain. “And so, if you did ten programs, that’s $150 million to $200 million of lifetime value that you’ve created; you’re not recognizing it today, but if you wanted to, you could go off and sell those.”

McClain is quick to add “But that’s not the business model,” meaning, selling the royalties.

As far as the ramp, while it’s highly dependent on what happens in the lab, and the AI hub, and the clinical path, it’s also tied to clinical success and commercial marketing if a drug happens to make it that far.

“You could think of it as, a good chunk of it is in-clinical development, and then we get royalties on top of that,” says McClain. “I would expect in the next three to five years, we will see revenues ramping up significantly, and really starting to create that hockey stick from a revenue perspective.”

Those milestones can come quicker if Absci can boost those dismal success rates of four percent overall and eighteen percent for Phase I and II trials. The bet McClain and team are making is that they improve those success rates, which would improve the payoff represented by net present value.

“All of our investors and analysts know that that's highly conservative,” he says of the $15 million to $20 million estimates. “Because with our technology, it should — and it will — in future increase success rates,” he says. “Even if you go from four percent to eight percent” success rate for drugs, he says, “that’s huge, even just on a net present value basis, that’s going from $15 million or $20 million to $30 million to $40 million — you double success, you double the NPV.”

PROOF OF CONCEPT

At the same time that it partners with Merck, says McClain, Absci will pursue some programs of its own. “We’re actually going to be developing our own pipeline” of drugs he says. “We don’t plan on taking these to Phase III, because that’s extremely expensive, but if you can take it to a proof of concept, let’s say, get an efficacy readout in Phase I, that’s a huge value to the asset.”

He predicts “we’re going to be able to get there faster than anybody else can,” meaning, getting to a Phase I. As for what those drugs might be, it won’t be a cure for cancer, he says. Rather, the orphan drug market, where disease cohorts are small enough that they usually don’t attract much investment, “is a really interesting area,” he says. He declines to say which indications those might be, but says the company intends to disclose that next year.

Getting to a Phase I trials with its own drug, and proving the Absci feedback loop, “gives you huge credibility, huge validation that the platform works, and that’s going to be driving even more partnerships our way,” observes McClain.

While oncology is off the table for the moment, McClain does allow as the topic is an intriguing one.

“The issue with oncology is that everyone’s going after the same known targets,” he says. “What we need to do is actually find new targets.”

PATENTS PENDING

If the company finds new targets for anything, it opens up a whole other aspect of the business: patents. The more that drug discovery turns to drug design, the more that Absci may be able to establish patents on both antibodies and drug targets, says McClain.

“We have very broad IP here,” says McClain. That includes both the wet lab technology of protein expression and whatever is developed with AI.

“If you can take a novel target, use the platform to develop antibodies against all the epitopes [locations on the target], then that enables you to make patent claims that you couldn’t have enabled in any other way.”

The dynamic of designing antibodies and linking them to the antigen becomes a kind of circular dynamic that establishes exclusivity, says McClain.

“You give yourself a big runway to go after that target where no one else could come in after you because you’ve defined the target as a function of the sequence variety [of the antibody] that goes after it.”

The company’s general counsel, Sarah Korman, who had been the head of IP and licensing at Amgen, was lured to Absci in part because of the patent prospects, McClain tells me.

“She saw the diversity that our AI models could create, and how that could actually enable very broad patent claims that previously were unattainable,” says McClain.

“Broad claims means you can block other people, and ultimately, kind-of, dictate who can come into a target.”

A power, no doubt, one must wield carefully, I offer.

“No, absolutely,” replies McClain. “You have to always remain, What is best for patients? How do we do what’s best for patients, and make money, and create shareholder value?”

THE BEGINNING OF THE ROAD?

If, as he says, Absci gets into the clinic in 2024, and if its AI models are really “zero shot” drug development machines, how soon will it be clear that this whole approach is going to make good on the promise of AI? Is it ten years down the road? Is it more than that?

“No, I think you’re going to see clinical proof of concept way sooner than ten years,” says McClain. “Even being able to get a Phase I efficacy readout, so you can actually show something that gets people excited, and then you go to Phase II and show that proof of concept.”

“This goes back to having somebody like Andreas on board that really knows the clinical development side, knows where to look for targets that could give us early efficacy signals in a phase one.”

It is still early innings, McClain reminds me again, as we wrap up. “I think the public needs to know that,” he says. “But, early innings are exciting.”

I’m reminded of cancer biologist Robert A. Weinberg’s great book, Racing to the Beginning of the Road. Weinberg wrote that the 1970s and 1980s were the decades that taught scientists the mechanism of cancer, why cells go rogue, why programmed cell death fails to rein in chaos.

Looking back from the 1990s on those decades of fitful, meandering research, Weinberg declared, hopefully, “after so long, we finally know where to look” for a cure.

Perhaps after years of work in wet labs and in AI, McClain and others have the tools they need to begin to make serious breakthroughs.

As we walk to the elevators, McClain, Absci’s biggest shareholder, with a ten percent position, tells me, “I think this will show you my bullishness: in eleven years, I haven’t sold a single share of stock.”

Absci shares, at a recent $2.38, are down seventy-one percent this year, and down eighty percent since IPO.