The Fruits of AI Drug Abundance: Indigestion In The Near Term
The Fruits of AI Drug Abundance: Indigestion In The Near Term
AI is transforming industries, and biopharma is no exception. AI will help make better drugs, hopefully more efficiently, and will aid in bringing those medicines to patients with expedience.
But here’s my prediction: AI will drive a wave of drug abundance and exacerbate the industry’s aggregate R&D productivity challenges over the next 3-5 years, rather than improve those metrics.
Great new technology waves often make things worse before they make them better. In 2001, at the peak of the Genomics Bubble, Lehman Brothers and McKinsey put out a report called “The Fruits of Genomics: Drug Pipelines Face Indigestion Until the New Biology Ripens” – which hypothesized that with the increased understanding of the human genome sequence, industry labs would be flooded with first-in-class unprecedented biologic targets – and new era of Target Abundance – all of which had little validation, which would lead to a spike in failure rates, elevated costs, and longer timelines. All very much counter to the “promise” that genomics would improve those things. The industry’s R&D Productivity declines for much of this century largely validate the Target Abundance thesis of that report. But no one would doubt that the Human Genome Project was of huge value to medicine today.
While there are many applications of AI across the biopharma R&D value chain, I’m talking here specifically about its role in “molecular design”. It feels like every day since the AlphaFold moment there’s yet another insight or case study for how AI is helping making drug candidates, be they small molecules, antibodies, antisense, or other modalities. All of our drug discovery companies are deploying some aspect of AI in their efforts, as there are definitely real use cases and valid traction in discovery. I’ve invested actively in advances in computational drug discovery (now often called AI) for two decades, so I’m a believer. That said, many of the current hyperbolic claims are over-hyped, in my opinion, especially as they relate to upending the industry’s R&D model in the near-term (or curing all disease).
AI is democratizing molecular design and bringing about Drug Abundance. You don’t need huge internal armies in drug discovery, or big HTS platforms and other technological infrastructure. You also don’t always need 30 years of medicinal chemistry or antibody engineering experience. Nimble small teams can do things only big R&D shops did previously. In addition, most “proprietary” AI models quickly disseminate in various incarnations across the drug discovery industry. Closed and open models alike are and will be used to drive new molecular designs. Competitive moats are harder in an era of AI proliferation, for both foundational labs and the drug-designer application layer.
But, critically, AI today is predominantly democratizing design around known targets, known drug scaffolds, and known biology. The training sets for most AI models involve the “known” drug discovery landscape, and predictive ability outside of those sets is poor (here, here). Cracking the undruggable – the novel and complex biology, the unique binding sites, the novel undiscovered pharmacophore, the orthogonal approach – remains elusive to most AI models, as Daphne Koller described in her recent blog that there’s no AI “magic wands” today. A similar AI realism theme was explored by Bender et al in August 2026 on AI in Drug Discovery and concluding that the “clinically relevant impact is… disappointingly limited.”
Because many AI-led players largely focus on the “known” drug discovery space (through the limits of their training sets), by definition it means they’ll be more work, more investment, and more drugs against known targets. I think AI will promote an unprecedented period of Drug Abundance: more candidates moving into late discovery and early development than ever before, especially since “AI native” startups are raising so much capital.
In light of this, I believe AI will undoubtedly exacerbate the crowding challenge in our sector – that intense competition that has brought over 100 global programs into R&D against each of the top eight targets (e.g., PD1, EGFR, CD19, VEGF, HER2, JAK, GLP, KRAS). There are at least 30 “hot” targets with over 30 programs against them in preclinical or clinical R&D around the globe.
Crowding around hot targets is as old as the industry itself. I’ve opined on the phenomenon here many times: March of the Lemmings in 2012 and the Strategic Supernova in I/O in 2016. It’s the nature of our business. And in 2025’s Wisdom of the Crowds, I highlighted that the culling of the crowded herd is typical in risk-taking ventures; while it’s bad for the unsuccessful me-too’s, and for industry R&D productivity in general, it may be good for patients (should be best drugs make it to them).
Since AI is democratizing work in these known spaces, we should be prepared for a tsunami of programs against validated biology, making the competitive crowding issue against hot targets more intense than before. As noted above, crowding around impactful targets is a time-honored tradition in pharma, and every era since the 1980s has seen complaints about the density of “me-too” programs (here, here). In fact, the term “me-too drugs” was coined in the 1950s! AI exacerbates this “me-too drug” issue by accelerating the pace.
Competitive intensity into known spaces has also been super productive for the industry historically: as Walter Sneader detailed in the seminal 1996 work on small molecule drug discovery, “Drug Prototypes and their Exploitation”, most drugs at that time fell into a relatively short set of core chemotypes. Famous drug hunter Sir James Black said it best: “The most fruitful basis for the discovery of a new drug is to start with an old drug.” This fact makes AI super powerful at making drug candidates against known targets: mine known drug/target data sets for training, and then predict new candidates with high probability.
As an aside, interestingly, some of the hottest programs today break these “known drug” rules and thus wouldn’t be in the “state of the art” training sets when they were discovered. Instead they were the product of hard-fought battles on the edge of novel drug discovery (e.g., RAS(on) molecular glues). First in class drug discovery remains hard, and lots of drug discovery will remain on the frontier (and outside of training sets). And while AI is a very helpful tool in the discovery toolbox, it’s not a golden ticket for these “undruggable” tough targets.
Back to the historic theme of crowding: in contrast to the past, today AI quickly enables working around and extending “known” science – to identify patentable new candidates – much much faster than ever before. Back in Sir James’ day, it still took lots of medicinal chemists years to find the new incrementally different drug. AI does that quickly, for both chemistries and for genetic sequence-driven drugs (like mAbs).
Indeed, patent-busting is one of AI’s strong suits: mine a patent application, identify what is likely the “lead” drug structure or sequence, and figure out where the soft spots in the patent are to make new molecules, incrementally different and potentially better than the patented version. These tools are widespread today. Not only for patents, but scientific literature in general. Many biotech companies are using them, and not just in China.
Just look at the pipelines of most of the major “AI Native” discovery shops who claim to be using transformative AI models for drug discovery: they are mostly populated with well-credentialled targets and lots of pre-existing chemotypes or protein sequences: known chemical scaffold-target pairs (e.g., MEK, HDAC, KRAS, TYK2, 5HT, etc…), validated antibodies (e.g., TSLP, IL13), precedented T-cell engagers (e.g., CD3-CD19), and common ADCs (e.g., HER2, TROP2), among others. Two biotech IPOs this year by “AI native” players, Generate and Iambic, are working on TSLP and HER2; these may very well have credible shots of “best in class” but they certainly aren’t “novel” programs. And, for the record, there are some novel AI-driven programs (e,g., Insitro’s TNIK inhibitor, Anthropic’s recent Mithril case study), but they are the exception rather than the norm.
Importantly, some of these new AI-inspired drugs against known targets could end up being clinically differentiated, and offer great new profiles to patients; as we all know, later entrants have often improved on the profiles of earlier medicines, like Lipitor in the statin class. It will take years and head to head clinical studies to know a winner, though. But its fair to acknowledge that not all of this crowding is bad for patients in the end, as I noted in Wisdom of the Crowds. Unfortunately, a plethora of programs all trying to claim differentiation, when only a few will deliver it, is certainly a recipe for lower aggregate return on investment for the industry.
Some specific examples for how crowding will have profoundly negative implications for the industry’s return on investment (“R&D Productivity”), even if it brings good medicines to patients:
- Wasteful spending in discovery in pursuit of elusive “Best In Class” profiles, with heightened competition for limited slots at preclinical toxicology vendors and CMC partners
- Clinical competition for patients as waves of similar programs enter the clinic, extending enrollment timelines for everyone and degrading the inclusion/exclusion criteria in studies, not to mention competition for the best trial sites
- Greater failure rates as so many AI-enabled “me-too drugs” fail to beat the “standard of care” precedent drugs in clinical trials
- Larger sums of capital invested (and lost) to get to a “no-go,” where the failure mode is often likely late-stage misses vs. standard-of-care or, even worse, commercial failures
Will AI models help us sort through this wave of Drug Abundance – helping us pick only the best programs to fund and preventing the waste? Maybe, if it coordinated across the entire sector and synced up decision-making across the thousands of private and public biopharma companies. That’s obviously not happening. So while the tools of AI could help any individual player prioritize the best programs and efficiently get them through development, this is unlikely to help the broader industry’s R&D Productivity metrics in a substantial way.
Will AI help us drive more investments in the clinic, via more throughput or faster timelines? Maybe, but I think on the margin. Players like Formation Bio are making this their AI development thesis, and I hope it plays forward favorably. I’m somewhat skeptical that the embedded inefficiencies and latencies of the industry’s Development machine (like “in-life” treatment periods) offer more than incremental 10-20% improvement opportunities for AI impact. And congestion around access to patients is a real concern today that AI-led throughput may not be able to solve; certainly more crowded clinical landscapes make other operational aspects harder.
In this brave new world, what should startup biotechs do? For those pursuing true novelty (like first-in-class targets, novel modalities or unique drug scaffolds, etc), it seems clear that one needs to fundamentally change the disclosure norms in this new AI (and China) reality: file patents later, use more decoy filings, keep PR to a minimum, fewer conference posters for early programs, limit the number and depth of scientific publications, etc…. To keep the AI wolves at bay, at least temporarily, novel biotechs need to adopt much more protective posture to their own efforts and “trade secrets”. Interesting that the biggest Pharma in the world, Lilly, has decided recently not to share it’s Phase 1 portfolio. Further, at a policy level, we should also figure out what changes we need to make to the patent law to enable better intellectual property protection (maybe considering going back to first-to-invent vs first-to-file). Lots of other strategic changes at the company and investment level, but we’ll keep those secret…
AI has enormous positive potential to change the drug industry and help patients – and it certainly will in the long term. But my prediction is we’ll face some serious “indigestion” as an industry before the impact of AI begins to “cure all disease” – hopefully I’m wrong.
Fundamentally, if today’s AI were excelling at novel medicine creation, we’d see tons of new programs addressing targets we’ve barely heard of with uniquely-AI-discovered drug scaffolds – but this isn’t really happening. While there have been one-offs, I’ve yet to see repeatable, scalable examples of generative AI finding totally new scaffolds for novel drugs against conventionally undruggable targets: using AI to tackle grievous diseases with first-in-class novel drug matter. This is the drug abundance we’d love to see, but not the drug abundance we’re likely getting today.
Maybe that’s too high of a bar right now for a transformational technology in its earliest innings, like it was in the years following the Human Genome Project when the Fruits of Genomics drove a wave of risky bets and hurt the ROI on R&D spending. At least in the near term, AI-led Drug Abundance isn’t likely to be the R&D productivity panacea the industry is hoping for…
The post The Fruits of AI Drug Abundance: Indigestion In The Near Term appeared first on LifeSciVC.
Email subscriptions powered by FeedBlitz® • 650 E. Palisade Avenue Ste 2, PO Box 173, Englewood Cliffs • NJ 07632, USA