Charlotte Deane, PhD, MBE, Exscientia Chief Scientist of Biologics AI and Professor of Structural Bioinformatics at University of Oxford.

After spending the past decade pioneering the use of artificial intelligence (AI) toward designing small molecule drugs, Exscientia has expanded its AI-based platform to begin designing precision engineered and optimized, fully human biologics, starting with novel antibodies.

The expanded platform will enable Oxford-based Exscientia to develop next-generation therapeutic antibodies through generative AI design—from computationally designing initial antibodies that bind to a specific epitope on a target protein, through virtual screening alongside generative design in order to identify candidates that bind to the specific target epitope with high specificity, while simultaneously optimizing multiple parameters.

The expansion of the platform will enable a near doubling of the universe of potential targets for new treatments, Exscientia says.

Of the approximately 20,000 genes coding for expressing proteins that are contained within the human genome, Exscientia estimates that about 10% of the genome expresses proteins—drug targets—that are druggable with small molecules, while an additional 10% of proteins are accessible to antibodies.

The quest to discover more druggable targets sparked the founding of Exscientia a decade ago by Andrew Hopkins, DPhil, who remains the company’s CEO. As a young scientist at Pfizer, Hopkins published in the early 2000s a groundbreaking review of the “druggable genome”, which estimated that some 3,000 genes within the human genome could potentially code for a protein with the ability to bind drug-like molecules. Together with Pfizer colleague Colin Groom, PhD, Hopkins identified almost 400 nonredundant molecular targets in 130 protein families that met Chris Lipinski’s famous “Rule of Five” oral bioavailability criteria.

“We think of ourselves as a tech company that develops drugs. AI has been the core to what we do,” Hopkins said earlier this year on GEN’s CEO interview series “Close to the Edge,” adding: “Some people have called us a pharma tech, how we combine technology and pharma together. But we are just starting the journey.”

Exscientia’s latest leg of that journey, the platform expansion, also opens the door to developing more complex biologics in the future. The first company to have AI-designed molecules entering clinical trials, Exscientia says its engineered antibodies will advance its goal of designing all of its biologics de novo for specific target epitopes without the need for physical screening.

“It’s a very natural next step in the evolution of Exscientia,” Charlotte Deane, PhD, MBE, the company’s Chief Scientist of Biologics AI, told GEN Edge. “The mission of Exscientia is to change the way the world invents drugs and make it more efficient and make more impactful medicines. And biologics are just part of that. A huge percentage of the top drugs now are antibodies, and we’re seeing other formats coming through from things like building antibody bispecifics.”

Antibodies accounted for five of the top 10 drugs and vaccines of 2021 as measured by revenue—AbbVie’s Humira® (adalimumab); Merck & Co.’s Keytruda® (pembrolizumab); Janssen Biotech (Johnson & Johnson)’s Stelara® (ustekinumab); Bristol Myers Squibb’s Opdivo®(nivolumab); and Sanofi/Regeneron Pharmaceuticals’ Dupixent® (dupilumab).

“Like a Set of Lego”

“In my head, it’s almost like a set of Lego building bricks. Here we’re picking, if you like, the easiest brick to start with, in the sense that it’s the one we understand best,” added Deane, who is also Professor of Structural Bioinformatics at University of Oxford. “But there is nothing that stops that from being a process where you can build other types of molecules and build more complex ones.”

Sequencing of antibodies has traditionally been limited to single chains, even though the binding site of the antibody consists of two chains, heavy and light. Exscientia has begun building a proprietary database of paired chain sequences and will use the data in machine learning algorithms designed to describe and model human antibody space with greater accuracy.

Earlier this year, Deane and collaborators from the University of Oxford and Roche Pharma Research and Early Development published a study detailing an initial version of the antibody design technology—namely a structure prediction tool for the six complementarity determining regions (CDRs) that serve as the key area for antigen binding and the main area of structural variation in antibodies.

According to the study, published in Bioinformatics, that initial version produced accurate protein modelling up to 35,000 times faster than AlphaFold2, the updated AI system of Alphabet-owned DeepMind that predicts the 3D structure of a protein from its amino acid sequence.

Two ways that modelling was sped up was by not having to accommodate proteins with lots of different sizes, and by not requiring a sequence alignment for antibodies, since they evolve differently from a standard protein.

“Antibodies are undergoing this very rapid somatic hypermutation, and very small changes in sequence lead to big changes in structure. The alignment doesn’t help you with that, because it’s not that same evolutionary process that they’re undergoing,” Deane said. “Removal of that is actually advantageous, both in terms of the accuracy of your answer, and in terms of increasing your speed.”

Since the Bioinformatics study was published, Exscientia has expanded the scope, speed, and accuracy of its antibody design algorithms, and integrated those capabilities into its drug discovery platform. This fall, the platform earned the company the Prix Galien USA 2022 Award for Best Digital Health Solution.

Exscientia says its virtual screening methodology for antibodies is now over three times more accurate than the initial version. Details of some enhancements are expected to be published by Deane and colleagues in coming weeks.

The most important CDR for binding is the CDR-H3 loop, as it allows for the most variability in structure due to H3 being made up of portions of three separate gene segments, V (variable), D (diversity), and J (joining). Deane estimated that the structures of about 6,000 antibodies—both therapeutics and natural antibodies—are available in public domain.

DeepMind and EMBL’s European Bioinformatics Institute oversee the AlphaFold DB open access database containing more than 200 million protein structure predictions.

Meta Genomics

In recent weeks, a larger source of AI-based protein structure data has emerged: On October 31, Alexander Rives of Meta AI’s Fundamental AI Research team (FAIR) and investigators from Meta AI and research partners published a preprint in bioRxiv presenting the ESM Metagenomic Atlas, a database of more than 617 million metagenomic protein structures, based on a new protein-folding approach that applies large-scale language models of protein sequences.

Developers of the Atlas say that it offers the first comprehensive view of the structures of proteins in a metagenomics database, and accelerates the speed of structure prediction by up to 60 times that of AlphaFold, while maintaining resolution and accuracy.

Exscientia reasons that its virtual modeling approach to antibody design will lead to antibodies better suited for a specific target than laboratory screening approaches that involve animal immunization to find antibodies that bind to a target of interest, or make use of a phage display library, or a humanized animal.

“You’re basically going fishing in an experiment to try and get something that hits,” Deane said. ”So, one, you don’t know if you’ll get any hits at all. Two, you have to be able to go fishing with your target in the first place. Membrane proteins are a good example where that starts to become quite difficult. And three, of course, even if you get something that binds, it may or may not bind where you want it to.

An example of the last challenge, Deane said, would be the antibodies developed to treat COVID-19, whose efficacy has been limited by the ability of newer variants to evade the drugs as the virus has mutated over the past three years.

“It may not be possible on the receptor binding the main [coronavirus] to come up with a site that actually blocks binding, but also is relatively not mutated,” Deane said. “But there are definitely examples where you can think of generating antibodies that you know will be specific across multiple variants much more easily If you can decide where you target rather than randomly hitting somewhere on the surface.”

That’s not to say that Exscientia has given up on small molecules. On November 14, the company announced a strategic collaboration with MD Anderson Cancer Center in Texas to develop cell-intrinsic small-molecule compounds against jointly identified cancer targets, by combining MD Anderson’s drug development expertise with Exscientia’s precision medicine platform. Successful target discovery programs may be advanced into proof-of-concept clinical trials at MD Anderson, the partners said.

Growing pipeline

Still uncertain is how the expansion to biologics will ultimately reshape Exscientia’s growing pipeline. More than half (17) of the company’s 31 wholly-owned and partnered pipeline candidates are in development for oncology indications—including its first candidate to reach the clinic, EXS-21546, which was co-invented and is being developed with Evotec.

EXS-21546 is set to enter the Phase I/II IGNITE-AI trial assessing the A2A receptor antagonist plus anti-PD-1 therapy in up to 110 patients with immunotherapy relapsed or refractory renal cell carcinoma (RCC) and non-small cell lung cancer (NSCLC). On Monday, Exscientia said it received clinical trial application (CTA) approval for IGNITE-AI, enabling site activation in Europe for the trial, which is set to start before the end of the year.

Another eight candidates are for inflammation and immunity, and three for psychiatry, including a Phase I clinical candidate co-developed with Sumitomo Dainippon Pharma, DSP-0038 for Alzheimer’s disease psychosis, a “design-as-service” candidate also developed with Sumitomo Dainippon, and a third candidate being partnered with an undisclosed company. (Exscientia and Sumitomo Dainippon also co-developed DSP-1181, a clinical candidate until Sumitomo Dainippon quietly ended its development after disappointing Phase I results).

Three of Excientia’s candidates are in IND-enabling phases—an oncology candidate and an inflammation and immunity candidate, both partnered with BMS; and GTAEXS-617, an oncology candidate partnered with China-based GT Apeiron. At the recent 34th EORTC-NCI-AACR (ENA) Annual Symposium, held in Barcelona, Spain, Exscientia presented biomarker data and novel patient stratification methods that it said support development of GTAEXS-617, a cyclin dependent kinase 7 (CDK7) inhibitor. A Phase I/II trial is set to launch in the first half of 2023 in ovarian cancer among multiple solid tumor indications.

The remainder of Exscientia’s pipeline consists of one candidate each for COVID-19 (partnered with the Bill and Melinda Gates Foundation), anti-infective (Gates Foundation), and unspecified rare disease (Rallybio). Exscientia focuses its internal pipeline on applying its platform in oncology, and its partnered pipeline in other therapeutic areas.

Looking forward

As 2022 winds down, so does Exscientia’s tenth year in business, and its first full year as a public company, following a $350.4 million initial public offering (IPO) and concurrent $160 million private placement in September 2021, which generated a combined $470.6 million in net proceeds.

Exscientia finished the first nine months of this year with a net loss of £79.64 million ($95.66 million), nearly triple the £29.14 million ($35 million) net loss of January-September 2021. The net loss heavily reflected a more than tripling of R&D expenses, which soared to £93.23 million ($111.98 million) from £25.31 million ($30.40 million) in Q1-Q3 of last year.

“We expect our research and development expenses to increase substantially for the foreseeable future as we continue to expand and advance our internal and partnered drug pipeline, invest in our technology platform and hire additional personnel directly involved in such efforts,” Exscientia stated in its latest quarterly regulatory filing on November 15.

Exscientia reported 414 employees as of mid-September. The company also appointed Caroline Rowland as Chief People Officer, a newly created role designed to lead Exscientia’s “people” (human resources) and “places” functions, as well as oversee the company’s talent, training and employee experience strategies.

Despite the growing loss, Exscientia still enjoyed a 74% year-over-year jump in revenue from collaborations, to £85.6 million ($102.81 million) in January-September. This year’s total included £74.2 million ($89.11 million) from Sanofi related to an up-to-$5.2 billion partnership launched in January to develop up to 15 novel small molecule candidates across oncology and immunology indications, and £11.4 million ($13.69 million) from two ongoing collaborations with BMS.

Also, Exscientia’s cash, cash equivalents, and short-term bank deposits have only dipped 0.2% during the year, to £561.08 million ($673.78 million) as of September 30 from £562.2 million ($675.12 million) as of December 31, 2021.

To facilitate the platform expansion, Exscientia plans to add 8,000 square feet to its headquarters at the Oxford (U.K.) Science Park for an automated biologics laboratory designed to enable the company to generate and profile novel antibodies internally. At the lab, the company plans to automate the production of proprietary data for each antibody, as well as measure essential qualities such as affinity, immunogenicity, aggregation, and stability.

The new lab is envisioned for small-scale, rapid expression and characterization of a large number of antibodies, with the resulting data to be fed back into Exscientia’s machine learning algorithms.

As for full-scale manufacturing, Deane said: “That’s something where we would partner with others downstream who have experience in doing those kinds of steps.”