AI Drug Discovery
AI and drug discovery, a match made in heaven? Read how this industry is being transformed by new, powerful technology.
EDGE100 Report, 2023
The number that drives the major worldwide industry of pharmaceuticals: 1060. Street name: Novemdecillion. This formidable figure is cheminformatics’ ballpark estimate for all drug-like molecules. The chemicals useful to modern medicine are mixed among 1060 possibilities, and sifting through this colossal chemical space is the early work of drug discovery. It’s the basis of some of pharmatech’s cutting-edge sectors including molecular mining and high-throughput screening (HTS), supported by biochemical compound libraries and drug discovery hardware and software. For all its numerous competing and complementary tech, efficiency means everything, and these companies constantly nudge at the boundaries of speed and cost.
Since emerging in the late ‘80s, HTS has kept pace with technology as one of the best answers to pharma’s needle-in-haystack compute problem. Not without substitute, HTS is a popular choice in commercially analysing in vitro, on an industrial scale, what happens when drug targets (such as cells or proteins and other biomolecules) are exposed to matter from the drug-like chem space. Chemicals that impact a target may be chosen as drug candidates and progress from “hits” to “leads”, with about one in ten drug candidates succeeding to clinical trials. Hit-to-lead services and lead optimization are some of the major sectors catering to drug development and clinical application.
With ever more miniaturized assays, tools borrowed from mass spectrometry, genomics, etc., and state-of-the-art automation, discovery has been hammering at the confines of conventional analytical biochemistry and computational science. However, R&D efficiency has been dulling. For the most part, breakthroughs only shorten drug discovery time and cost at most by months and millions. Each novel drug on average still takes more than a decade and a billion dollars before commercial launch. For some critical categories, like antibiotics, that are truly at an impasse, speed to market is now irrelevant—chronic infectious disease care badly needs a paradigm shift. For pharmatech, humbled by universe-sized data and long thirsting for a thinking machine, it seems like AI is just what the doctor ordered.
How do AI and Generative AI fit into drug discovery?
In the early stages of the drug discovery process, companies spend billions of dollars on many decade-long projects that are far more likely to fail than to succeed. Some sources give the current probability of success of drug discovery projects at a daunting 0.004%. Based on this statistic, adopting the best technology for data processing is a no-brainer strategy to save a portion of the trillion dollars the pharmaceutical industry would commit to wastage, and to shed years off pre-clinical discovery
In this light, AI has two broad roles in drug discovery. Naturally, one of those roles would be the discovery of drugs. In these early days of AI adoption, however, the second role might be as important as the first: the role of making failures cheaper, in both cost and time. This explains why pharma investments into AI keep pouring in even though the world is yet to see a single true AI drug on the market (less than a handful of AI drugs were successful in phase III trials in 2023, according to one optimistic estimate). In the industry’s best example of AI saving time and money, one startup project shortened drug candidate generation to 46 days, down from 2-5 years in traditional research, and spent just a fraction of the traditional costs, USD 2.6 million versus an estimated USD 430 million in out-of-pocket expenses.
Even if that drug candidate fails clinical stages, it has already demonstrated value, making the AI technology behind it monetizable. Other AI tools, too, have shown various capabilities, such as being able to analyze medical data at 100x the traditional speed or generating 80% cost savings in drug discovery. The promise of such large savings is magnetizing the industry.
Applications of AI and Generative AI in drug discovery
AI in its most basic form mimics human thinking and behaviour for industrial application but at exponentially higher processing powers. AI can analyse information and differentiate and respond to patterns in massive amounts of data that stump conventional analysis (aka ‘big data’). Machine learning involves programming AI systems beyond pattern recognition, to learn tasks and decision-making. The impact of AI is alluring, but how specifically are organizations leveraging its power to generate results? We've identified a number of AI drug discovery use cases below:
- In cell line development, generative AI has already reduced antibody discovery work by four months and provided greater B-cell diversity with the ability to screen from multiple organs
- In drug candidate identification, it has been used to identify and develop compounds for oncology, neurology, and immunology, significantly reducing the number of cycles required in the drug discovery process
- In lead optimization, it has helped develop potential novel anti-cancer therapies, increasing the success rate of lead optimization and reducing time and cost associated with drug development.
- In breast cancer diagnosis, it has reduced total read times by over 55%, including a ~65% reduction for slides suspicious of cancer, and improved diagnostic accuracy with 70% pathologist efficiency improvements
- In identification of drug targets, generative AI is being used to identify novel targets to treat systemic lupus erythematosus and heart failure.
Tech and medicine come together in a diverse competitive landscape
AI drug discovery gathers companies in various forms and sizes. Commercial AI drug discovery models are typically run on specially developed AI platforms, although technically nothing should stop AI from being run on a laptop. Drug discovery startup Superluminal Medicines, for example, began its AI program using cloud computing in a small office in its home base of Maine, USA. At the other end of the spectrum, to supercharge AI to find a covid-19 vaccine in record time, Pfizer went to Spain seeking the supercomputer called MareNostrum 4.
Like pharma seeking out tech, computer companies are not on strange terrain exploring healthcare. Commercial AI drug development models benefit from extreme processing efficiency not just from industry-standard GPUs, but also from AI-specific hardware, like NVIDIA’s deep-learning accelerators. In AI drug discovery, Silicon Valley partnerships are a natural evolution of the foundational link between medicine and computing, between wet labs and dry labs. Many industry players specialise in software-as-a-service (AI SaaS) for specific stages of drug discovery. Established pharma firms are regularly seen collaborating with AI SaaS companies as well as with tech incumbents, like NVIDIA, that show interest in drug discovery. AI drug discovery made up the biggest share of pharma automation activity at the start of 2024.
Today, AI drug discovery companies span the entire drug discovery value chain. At the start of the value chain, AI companies are ramping up studies of proteins and other biomolecules by leveraging AI. Next, AI platforms offer an alternative screening technology to HTS or to remarkably speed up parts of HTS, such as iterative screening. In the drug candidate stage of discovery, AI can predict clinical outcomes for AI drug candidates. Similarly, AI is playing a growing role in target discovery, lead optimization, and drug design. The fields of pre-clinical development, digital pathology and diagnosis, and clinical trial management and patient recruitment also benefit from AI.
We’ve identified a few top trends in this quickly-changing landscape:
The AI drug discovery industry comprises AI SaaS firms and AI-integrated biotech players
Industry participants are either pharma incumbents that integrate AI in various ways, either through acquisitions or partnerships and collaborations, or startups of various origin, whether coming from the tech industry, the biotech industry, academia, or collaborations between these.
Pharma-tech partnerships are common
Pharma giants have been quick to scoop up the capabilities of top AI drug discovery companies. French giant Sanofi is one example of big pharma with an extensive AI program, signing up early with AI providers Atomwise, Exscientia (known for its oncology focus), Insilico Medicine, and many others.
Tech giants in AI drug discovery
Alphabet, Amazon, and Microsoft, among other large tech firms, are actively scoping out AI drug discovery, and partnerships are an important strategy for innovation for these. Google’s DeepMind and Isomorphic Labs’ AlphaFold 3, for example, is one iteration of an AI that predicts protein structures in 3D and, as of its May 2024 launch, interactions between a wider range of biomolecules. Alongside other software, AlphaFold won its team the 2024 Nobel Prize in chemistry.
Startups in early discovery – some highlights
That 2024 Nobel Prize in chemistry was shared with the key scientist behind Xaira Therapeutics, an AI drug discovery startup partnership also worth watching. Xaira uses an AI model called RFdiffusion for protein design and raised a historic USD 1 billion in its first round. Like Xaira, Cradle and EvolutionaryScale are also among many players focusing AI power on proteomics services. Novel proteins, other synthetic molecules, and interactions that work in vitro, though some distance from being put into the human body, have an important place in R&D.
Molecular studies complement research that’s even closer to actual medicine, starting with AI-powered hit discovery. Atomwise, for example, says its AtomNet AI platform could be a cost-effective alternative to HTS. Further along the drug discovery path, Iambic Therapeutics claims its AI called Enchant launched in 2024 can predict clinical outcomes for drug candidates, aiming for more efficacious medicine and for pharma clients to shave off significant costs by dropping any leads that might lead nowhere.
Challenges and risks to growth in AI drug discovery
AI in drug discovery is young and still undergoing curious experimentation, with some projects going smoothly, others quietly dying, and some causing happy accidents. A team of MIT researchers might set out using AI looking for a drug for diabetes and instead stumble on a resistance-resisting antibiotic, which is what happened with the AI-discovered drug-like molecule Halicin. This surprise discovery was named after Hal, one of the most-loved AI characters in sci-fi and memorable for the problems it eventually posed—what better segue into looking into the challenges and risks of AI drug discovery. The main sources of risk for AI drug discovery are problems of data volume, quality, and bias; regulatory and ethical considerations; and a talent shortage.
Inherent limitations in data
AI output can only be as good as the quality of its input data. Biomedical data is fragmented among various stakeholders. The data that’s accessible for AI drug discovery can be skewed or poorly validated, slowing growth in AI drug discovery.
Threats to data privacy
The US National Institute of Health talks about data privacy and security as a possible ethical issue in biotech R&D. The attraction that large amounts of personal medical information hold for the pharmaceutical industry is very real. The temptation to unethically augment such data, this time to feed voracious demand for AI analysis, makes the issue even more sensitive. On the other hand, until governments find a middle ground where medical data can be sanitized and ethically repurposed for R&D, stricter regulatory scrutiny on data sharing and blanket bans could seal off valuable information.
Regulatory volatility
Regulations governing AI drug discovery companies are only just being created. The USA’s Center for Drug Evaluation and Research, which formed an AI Council in 2024, notes that the FDA “plans to issue draft guidance” on regulating the creation of drugs and biological products using AI. This creates some uncertainty for the industry, as new developments could bump into brand new rules written along the way.
Talent vacuum
AI drug discovery startups are hampered by the scarcity and high cost of AI engineering talent, a factor that might slow this industry’s growth engine.
Future trends
The future of AI drug discovery will be guided partly by AI’s ability to overcome these risks and challenges and turn around R&D efficiency and ROI for the pharmaceutical industry. AI drug discovery is an opportunity for pharma to turn the page on a decade of increasing R&D costs and to adjust to major revenue risks in the face of a slew of patent expirations.
AI drug discovery is also set to diversify drug discovery into areas in medicine that faced neglect because of either insufficient data, analytical power, or ROI due to market size. One such area is rare diseases: only 5% of the total number of rare diseases have FDA-approved drugs. In the future, AI also promises to advance precision medicine: AI helping fit drugs to patients’ genetic profiles would be game-changing for the medical field and for the human health span in the age of biohacking.
Commercial innovation could be accelerated by companies that piggyback on AI tools and technology from non-profits, academia, and technology companies that periodically develop open-source AI tools and technology. Artificial intelligence has some roots in open source, with big names like Google, Meta, and OpenAI each having dabbled in code sharing from time to time. Following this tradition, Chai Discovery, for example, has shared out its drug discovery AI, Chai-1, for purposes of innovation (under a license that restricts commercial use and service offerings).
As AI drug discovery gains traction, we might see more databases and compound libraries reanalysed by AI. The decades to come should birth novel drugs from old catalogues through AI’s new eyes. Efforts are also likely to be made to improve the quality of existing data and even to adapt data sets to suit specific domains of R&D. Meanwhile, computer processor manufacturers are likely to keep upgrading their products over time, supporting AI drug discovery.
Stay up to date
AI drug discovery is at an exciting and dynamic stage. The next three years could grow its global market just beyond 4 billion US dollars or to 14 billion, either way transforming biotech automation and breathing life into pharma. Sign up for our weekly AI newsletter at our blog homepage to stay up to date on developments, or try our platform for free and deep dive into anything and everything related to AI Drug discovery.
References
(1) “McKinsey research indicates that gen AI applications stand to add up to $4.4 trillion to the global economy—annually.” Source: https://www.mckinsey.com/featured-insights/mckinsey-explainers/what-is-generative-ai
(2) ‘Principles of early drug discovery’ published by the national library of medicine
https://pmc.ncbi.nlm.nih.gov/articles/PMC3058157/#:~:text=The%20output%20of%20a%20compound,the%20primary%20drug%20target%20protein.
(3) HTS was first introduced in the 1990s (more Googling says 1980s)
(4) Vaccines: Baidu and mRNA vats
(5) Atomise, Cradle, etc.
https://www.labiotech.eu/best-biotech/ai-drug-discovery-companies/
(6) A-Z glossary of AI in Healthcare
https://www.owkin.com/a-z-of-ai-in-healthcare
(7) Graphable’s Sherlock (healthcare LLM)
https://www.graphable.ai/blog/ai-in-drug-discovery-and-development/
(8) NVIDIA’s MegaMolBART (“A BART transformer language model trained on molecular SMILES strings”)
https://catalog.ngc.nvidia.com/orgs/nvidia/teams/clara/models/megamolbart
(9) Meta’s ESM metagenomics atlas
https://ai.meta.com/blog/protein-folding-esmfold-metagenomics/
(10) McKinsey
(11) List of AI tools:
https://roboticsbiz.com/ai-based-tools-and-databases-for-drug-discovery/
(12) The size of chemical space
https://www.mn.uio.no/hylleraas/english/news-and-events/news/2024/Vast_Chem_Spaces_Gen_AI.html
https://pubs.acs.org/doi/10.1021/ar500432k#:~:text=The%20so%2Dcalled%20%E2%80%9Cdrug%2D,too%20large%20for%20practical%20application.
(13) Superluminal’s history
https://cloud.google.com/transform/superluminal-medicine-interview-ai-drug-discovery-q-and-a-how-it-works
(14) NIH on risks
https://pmc.ncbi.nlm.nih.gov/articles/PMC10302890/#:~:text=Challenges%20and%20Limitations%20of%20Using%20AI%20in%20Drug%20Discovery&text=In%20many%20cases%2C%20the%20amount,of%20the%20results%20%5B10%5D.
(15) Merck’s QSAR program:
https://pmc.ncbi.nlm.nih.gov/articles/PMC7010403/
(16) Low success probability in drug discovery
https://www.mdpi.com/2673-7426/2/4/39
(17) McKinsey: USD 15-28 billion in drug research and early discovery
(18) https://insilico.com/phase1
(19) Zero shot: https://www.ibm.com/think/topics/zero-shot-learning#How++zero-shot+learning+works
