CEO, Insilico Medicine
While ChatGPT, released in 2022, brought advances in generative AI into the consumer domain, many advances in deep learning and generative AI for drug discovery were made in the past decade. My first peer-reviewed paper on generative AI for small molecule generation (DOI: 10.18632/oncotarget.14073) was published in 2016. In 2019, after a year of review, we published experimental validation (DOI: 10.1038/s41587-019-0224-x) up to mice studies of a molecule generated using Generative Tensorial Reinforcement Learning (GENTRL). Nevertheless, today the AI-powered drug discovery (AIDD) industry is going through major consolidation costing investors billions of dollars. During the past decade, I saw hundreds of overhyped AI startups aiming to disrupt pharma emerge. Several generations of these AIDD companies have failed, are failing, or have transformed into traditional biotechnology companies. In this article, I will opine on the hype and reality of AIDD. Please note that I am writing from my perspective as a founder of Insilico Medicine, so I am deeply biased and conflicted.
What is hype?
Recently, we have heard some AI drug discovery companies assert that generative AI is overhyped, and that they will need more high-quality data, additional resources, more computing power, or something else to deliver on their promises. In other words, they are trying to either lower expectations or bring attention to themselves as “realists,” “veterans,” or “wise gurus.” Unsurprisingly, these companies failed to set drug discovery and development records with AI after raising massive amounts of capital, despite many years of technological development.
What usually is pointed out as “hype” are the catchy headlines in the media promoting new companies or product launches or statements by AI scientists proclaiming that AI will soon solve biology and cure all diseases. Some may criticize the mainstream media popularity of AlphaFold, with no drugs developed using this groundbreaking technology nor any other popular AI technologies or companies. How dangerous is this type of hype, and should we fight it? In my opinion, media attention geared toward science and scientists is a very positive thing. When it comes to drug discovery, it provides basic education to nonscientists who need flashy headlines to pay attention.
However, there is another type of hype that is very dangerous: Financial hype. Every technological revolution creates a wave of media coverage and a wave of capital chasing the next Internet. The most overhyped companies are usually incubated by famous venture capitalists. They have a simple formula—identify the trend, speak with many startups to understand what works, craft a compelling story, bring together several big names, spice it up with academic research, and syndicate a substantial capital raise. Often, these companies make headlines because the media naturally assumes that the amount of capital raised, investor reputation, and the big names will produce the winning formula.
In the next round, the company makes a few high-profile partnerships with the pharma companies that provide upfronts “on faith,” attract money raises from pension funds, and make headlines. But then, years go by without a real drug getting into human clinical trials, and big names clash and leave to pursue new trendy ideas or retire.
What is surprising is that the media, the investors, and even the pharma companies that hailed this kind of company rarely look back and draw conclusions. In pharma, drug discovery and development cycles are simply too long, and people often change their jobs faster than their failures become apparent. Many of the industry players are committed to their personal interests and careers rather than therapeutic programs. Financial hype is very dangerous and results not only in failed expectations and lost fortunes, but also in patients not getting the drugs in time and in the loss of thousands or even millions of lives.
How to separate hype from real progress in AI?
In 2013 and 2014, when the deep learning revolution started, it was difficult to distinguish who is who. It took several years for the pioneering companies to learn, publish, and set some standards. Today, the barriers to entry are much lower, and there are examples of end-to-end programs published in peer-reviewed literature (DOI: 10.1038/s41587-024-02143-0). Entire data rooms with timelines and explanations are available for review and benchmarking, making it very easy for sophisticated analysts and journalists to do their research. Unfortunately, most still follow the big names and money raises instead of looking at measurable results and experimental validation.
The easiest way to avoid the dangerous financial hype is to ensure that AI-powered drug discovery companies receive funding gradually, based on the milestones they achieve and not on promises and big names. Substantial capital deployment is justified when the company is ready to start a certain number of Phase I or Phase II trials—not earlier. If the company needs to invest in computational resources, it can easily partner with a cloud provider to get the necessary resources instead of building its own infrastructure that will be obsolete by the time one drug reaches the clinic. It is also important to review their capabilities with experimental validation. Recently published work by École Polytechnique Fédérale de Lausanne, Columbia University, and Oxford University scientists (DOI: 10.1038/s42256-024-00843-5) provides a summary of most experimental validation studies in generative chemistry and a framework for evaluating the technology readiness.
I came up with a list of simple questions to help evaluate the companies in this reasonably mature field of AI-powered drug discovery and estimate their hype-to-results ratios:
- How many of your internally discovered drugs are in Phase I, Phase II, and Phase III?
- Did any of your pharma partnerships progress into Phase I, II, and III, and how long did it take?
- How many preclinical candidates (PCCs) did you nominate in a single year and in total?
- How long, on average, does it take you to nominate a PCC, and how much does it cost?
- What was the level of novelty of the targets and what is the level of novelty of the molecules?
- How long, on average, does it take you to go into clinical trials with an AI-discovered target?
- Did you in-license any of your drugs from other companies, repurpose them, or discover them internally?
- How many pharma companies use your AI software to discover drugs, and how many users do you have? How many renew every year?
- What was your revenue last year, and what was the revenue growth year over year?
- How much money have you raised, in how many rounds, and what is the valuation of your last round?
Unfortunately, most journalists, analysts, and even investors I know rarely ask these questions, even though the answers can be summarized in one small table or figure.
Are there benchmarks for AI-powered drug discovery?
I can share some internal benchmarks from Insilico Medicine with the disclaimer that these are approximations and cannot be used as financial advice or any form of solicitation. In 2020, we released the AI software suite, Pharma.AI, which is now comprised of Biology42, for target discovery and design of biologics; Chemistry42, for generative chemistry; and Medicine42, for clinical trial analysis. Today, 10 of the top 20 pharma use elements of this software, generating millions in recurrent software revenue and constantly validating the models. In 2019, we started using Pharma.AI internally to discover and develop novel therapeutics that we sell to other pharma companies to showcase the capabilities of the software.
To sell an asset, you need the highest level of quality and unique properties—the standards are often higher for external assets than for internal programs. From January 2021 to May 2024, we nominated 18 PCCs, with seven programs proceeding to human clinical trials and two progressing into Phase II. In 2022 alone, we nominated nine preclinical candidates—eight internally and one for a pharma partner. We aim only for first-in-class or best-in-class novel molecules for high- to moderate-novelty targets. The QPCTL program (now in Phase I) was partnered for co-development before the PCC stage, and USP1 and KAT6 were out-licensed at Phase I and PCC stages, respectively. The $80 million upfront hints at the quality and novelty. On average, it takes us 11–12 months to reach the PCC stage and 20–24 months to start human clinical trials for a moderately novel target.
Did generative AI improve drug discovery?
Our results suggest that we can reduce preclinical development time in half and increase the probability of success, based on 18 internal real and novel small-molecule programs. But in my opinion, AI-discovered drugs in clinical development should go through clinical trials without cutting corners and “move with the speed of traffic,” complying with all of the FDA regulations to provide convincing evidence for safety and efficacy. To provide convincing evidence for the impact of generative AI on drug discovery, a company needs to demonstrate the clean, well-documented pathway from target discovery and drug design to delivery of a blockbuster therapeutic benefiting a large patient population, not just a treatment for a rare disease. Given the state of the industry today, I expect this to happen between 2026 and 2029, in the absence of war or economic collapse. Although it is too early to tell if developing a new therapeutic can be accelerated, 18–24 months from program initiation to the start of human clinical trials should be standard for AI-powered biotech.
Unfortunately, there are only a few books written about drug discovery. My favorite is The Billion Dollar Molecule: The Quest for the Perfect Drug, published in 1994. I hope that a few years from today, someone will tell the story of deep learning in drug discovery, after the first drugs genuinely discovered and developed using AI receive regulatory approval.
Alex Zhavoronkov, PhD, is the founder and CEO of Insilico Medicine.
