AI background, pill
Credit: Just_Super / iStock / Getty Images Plus

When Google DeepMind’s AlphaFold was unveiled in 2020, the new protein structure prediction program was hailed as a breakthrough that could serve as a magic bullet for the high costs and dismal failure rates in drug discovery and development. AlphaFold uses an end-to-end deep neural network trained to produce protein structures from amino acid sequences, multiple sequence alignments, and homologous proteins.

Leveraging the success of this new program, just about two years from its launch DeepMind’s AI spinout Isomorphic announced two drug discovery deals, worth $3 billion each, with Eli Lilly and Novartis.

Earlier this year, microprocessor giant NVIDIA also dove head first into AI for drug discovery, making big investments and deals with leaders like Recursion Pharmaceuticals and Genentech.

AI in drug discovery seems to be having a moment.

All of these AI advances are dovetailing with a key trend: “The entire industry is pivoting to precision medicine,” Alban de La Sablière told Inside Precision Medicine. He is the chief operating officer of Owkin, which uses patient-data-based AI to speed clinical trials. “A lot of drugs have failed because companies were reaching for the 6th millionth mutation and it didn’t have that big an effect. Pushing an asset forward, but in a smaller subgroup using AI with the right data makes more sense.”

The latest version of AlphaFold is even better than the first and there is good reason to expect this advance to be transformational. “The application of AlphaFold to drug discovery is very direct, since structure prediction of soluble globular proteins, or regions of proteins, is useful at many points along the pharma and biotech R&D pipeline,” said Sarah Teichmann, FMedSci, FRS, group leader in cellular genetics at the Wellcome Sanger Institute.

Artificial Intelligence concept brain with CPU
Credit: MF3d / Getty Images

A Boston Consulting Group (BCG) February 2022 report showed that from 20 AI-based drug discovery companies, about 15 candidates were in clinical trials. Eight of the candidates had reached the clinical trial stage within 10 years of discovery. Another BCG report, published in June 2022 and drafted in collaboration with Wellcome, showed that AI could produce time and cost savings of at least 25–50% in the drug discovery phase.

TechCrunch said in November 2021 that AI “supercharged” the field of drug discovery. Eric Topol, MD, wrote in a recent Substack post (October 1, 2023), “Not long ago, scientists might spend 2 or 3 years to define the 3-dimensional structure of a protein, … Now that can be done for nearly all proteins in a matter of minutes, thanks to advances in AI. Even new proteins, not existing in nature, never previously conceived, can now be created.” 

AlphaFold was taught by being fed the sequences and structures of around 100,000 known proteins. The latest version can now generate predictions for nearly all molecules in the Protein Data Bank, the world’s largest open access database of biological molecules. The algorithm has been reported to frequently reach atomic accuracy.  

But big challenges still face AI. “Companies such as Exscientia and Insilico have demonstrated you can accelerate the drug cycle,” de La Sablière said. He believes that using AI drug developers can cut at least 1–2 years off every phase of drug discovery and development. Picking the right patients to study, he pointed out, is key. In heart disease, for example, it’s very hard to anticipate people at highest risk to properly power a study. The goal of big pharma, he said, is to cut 40% off the drug discovery cycle.

He posed a looming question: “Are the molecules better than what a chemist will find?”

Researchers also ask how many data points are needed for a powerful AI algorithm. “We have a good idea about how many data points we need to build a good model. Something interesting is that you can pretrain an algorithm on lots of data and use it to solve problems with smaller data sets,” said de La Sablière’s colleague, Jean-Philippe Vert, PhD, chief research and development officer of Owkin.

Lots of optimism

A wealth of other tools and strategies are fueling the field’s growth (see sidebar, More and New Types of Data). Dozens of AI drug discovery companies have been around for at least a decade, bolstered by a healthy batch of fresh financings and deals. 

For example, Insilico signed a deal last year with Sanofi of up to $1.2 billion to advance drug development candidates for up to six new targets.

“Although there was plenty of skepticism by pharma companies toward AI’s potential during Insilico’s early days, most have now added AI capabilities into their R&D operations. Some companies like Sanofi have now even made AI a core part of their strategy,” said Alex Zhavoronkov, PhD, chairman of the board, executive director, and CEO at Insilico. The company has several drug candidates in clinical trials and over 30 programs.

Alex Snyder
Alex Snyder, MD
Vice President of Research and Development

Generate:Biomedicines recently closed $273 million in series C financing to advance its generative AI pipeline of preclinical and clinical protein candidates. Soon after, the company inked a deal with the Roswell Park Comprehensive Cancer Center to create optimized chimeric antigen receptor (CAR) T-cell therapies.

“Most CAR Ts have binders [molecules that make them stick to cancer cells] that have been used for years and mainly work in liquid tumors,” said Alex Snyder, MD, executive vice president of research and development at Generate:Biomedicines. One big question, she said, is “how do we get into refractory liquid tumors and solids?”

She added that Generate “can test thousands” of constructs, using computational and wet lab feedback, “whereas most groups can only test hundreds.”

There are practical applications of the technology in play already. Researchers at Nagoya University in Japan used AI to synthesize a candidate compound for a new gastric acid inhibitor with a better binding affinity than existing drugs. Their findings were published in Communications Biology in September 2023.

“We ‘knew’ where Drug A is bound to, and where Drug B is bound to. But binding site A and B are different. The idea is that if we could combine all the binding sites with one compound,” said associate professor Kazuhiro Abe of Nagoya University.  

The group focused on the steric structure of the gastric proton pump, a protein in the stomach lining that transports the protons that make up hydrochloric, or gastric, acid. Using their “Deep Quartet” AI-driven drug discovery platform, the group generated more than 100 candidate compounds with unique chemical structures. They then selected the most promising candidates for synthesis and tested how strongly they bound and inhibited the gastric proton pump. The sixth compound synthesized (DQ-06) exhibited stronger binding than existing reference compounds. 

Other encouraging recent developments are:

  • Iambic Therapeutics’ oversubscribed $100 million series B financing to advance AI-discovered therapeutics.
  • Venture firm Andreessen Horowitz co-led a $200 million investment in Genesis Therapeutics, which says it will use the funding to get its first AI-based drug candidates into the clinic.
  • PostEra announced a multi-target collaboration with Amgen to leverage PostEra’s AI platform, Proton, to advance up to five small molecule programs. Terms were not announced, but PostEra said the agreement includes upfront funding and milestone payments in addition to royalties on any approved products arising from the collaboration.

The biotech rollercoaster routine

With skepticism typical for a relatively new technology, biotech analysts have been counting the number of early, high-profile, AI-based drugs going down the tubes. So far, these seem to be balanced out with new deals.

In April 2023, a test of BenevolentAI’s dermatitis drug (BEN-2293) failed to meet secondary efficacy measures in a Phase IIa atopic dermatitis study. The drug inhibits multiple tropomyosin-related kinases. But in September 2023, Benevolent signed a new drug discovery deal with Merck, initially for three targets in oncology, neurology, and immunology. The agreement reportedly includes a low upfront payment but milestones of up to $594 million.

Exscientia announced that it wound down a Phase I/II study of its AI-based cancer drug candidate EXS-21546 in October 2023. However, the company also announced that it had delivered its third AI-generated candidate to partner Sumitomo.

Also in October 2023, Bayer and Recursion Pharmaceuticals announced a shift in their collaboration away from fibrotic diseases to oncology, with the goal of advancing up to seven AI-based programs. The new deal includes potential total payments of up to $1.5 billion to Recursion based on milestones and royalties on net sales.

What’s next?

None of this can happen without a lot of computing power.

Emphasizing this need, Recursion recently announced that it was increasing the compute capacity of its BioHive-1 on-premise supercomputer four-fold by buying more than 500 NVIDIA H100 Tensor Core GPUs. The company already had 300 A100 Tensor Core GPUs. In a press release, Recursion announced that this expansion would make BioHive-1 one of the top 50 most powerful supercomputers in the world.

Such a marriage of keen algorithms, expansive data, and heavy-powered computer power is required to turn around AI’s record in drug discovery and development.

Malorye Branca is a contributing editor at Inside Precision Medicine and a freelance medical science journalist. She has written hundreds of articles, as well as managed and launched health and science magazines, newsletters, and market research report businesses. She has also co-authored two books: “Moneyball Medicine” and “Walmart’s Second Opinion.”

Also of Interest