Tech VCs and Biotech VCs: Talking Past Each Other on AI Drug Discovery

[Editor’s Note: this is a new column called “Astounding HealthTech” by TR contributor David Shaywitz.]

David Shaywitz

If you want to know the difference between tech venture capitalists and biotech VCs, look at their respective views on AI applications for drug discovery.

Many of the most prominent AI drug discovery startups boast exceptionally rich valuations, driven by the enthusiastic participation of tech VCs. 

While many life science VCs view these opportunities as intriguing, they struggle to make sense of the often-stratospheric valuations.

For example, Salt Lake City-based Recursion Pharmaceuticals was valued this July at $699M, according to Bloomberg, while UK-based BenevolentAI was recently said to be valued at over $1B – a staggering figure that actually represents a haircut of about 50% from its April 2018 valuation, according to the Financial Times.

The thesis of tech VCs seems to be that these sorts of approaches will revolutionize biopharma, and anyone seeking to be competitive will need to acquire such a platform or risk becoming obsolete. 

Life science VCs, of course, recognize the limitations of contemporary biopharma drug discovery.  Successful drugs are so rare, and so hard to predict in early R&D, they are considered “a miracle,” according to the Novartis CEO and the head of research at Merck. Most life sciences VCs and research executives are wary that AI, however useful, may not represent the silver bullet for what they’re up against.

As David Weitz, the head of Takeda’s La Jolla site, who has thought deeply about this issue, recently told me in the context of this piece (disclosure: I work at Takeda’s corporate VC arm):

“The value of AI is demonstrated when its predictions are validated and the combined AI prediction / validation effort is catalytic. Do AI predictions provide insights not otherwise attainable by other methods?  Are the predictions actionable and have a reasonable probability of success to merit experimental follow-up? Taken as a whole, was a meaningful roadblock more efficiently overcome than traditional methods?

David Weitz

Even if AI supports a step in the drug discovery workstream, success only comes when the remainder of the drug discovery efforts are executed effectively. That may be outside an AI solution provider’s core expertise.  The ultimate measure is the efficient delivery of high-quality discovery candidates. The ‘how’ is secondary.

So build vs buy?  Neither.  Collaborate smartly where the partners synergize their respective capabilities.  Measure the program’s holistic success and secondarily the extent to which AI provided a breakthrough or an efficiency.”

This view seems to align with cautions raised by Derek Lowe in his recent critique of the breathless excitement around AI purportedly discovering a drug class via virtual screening. Lowe asserts this effort was not much better, if better at all, than traditional approaches. Even if the claimed AI breakthrough is found to be a legitimate reproducible advance over time, “the costs at this point are but tiny little roundoff errors in the total cost of a real drug development project,” he wrote.

This perspective seems generally echoed by noted life science investor and commentator David Grainger (as I recently quoted here).

“When you get under the hood, there isn’t really anything special [about the AI approach]… I think more data and better data science can revolutionize some human endeavors, but drug discovery is not one of them – it can help make better the parts of the process we do adequately now, but it doesn’t address the real rate-limiting step on innovation, which is understanding biology. Here, what is needed is not data and data science, but education about complex systems (a la the Santa Fe Institute). Current ‘AI’ approaches are so linear and reductionist, it makes me shake my head in disbelief at all the hype.“

As another well-regarded life science VC recently pointed out to me about one prominent AI-for-drug-discovery company, “What’s the business model?  I don’t think it’s there.”

What’s been demonstrated thus far hasn’t persuaded most of the “Where’s the Beef?” crowd that tends to populate life sciences VC firms. The tendency of life science investors is to especially prize assets that have generated some interesting data from experiments (i.e., they look like potential drugs). It isn’t unusual for even a so-called “platform” biopharma to be valued essentially based on the asset furthest along in clinical testing. Often, little, if any, value is placed on earlier-stage assets or the platform itself (see this relevant discussion of Nimbus model by Bruce Booth of Atlas, who describes a potential solution).

Tech VCs, by contrast, tend to break out the checkbook for exciting ideas that have seemingly boundless potential – especially ones that address large total addressable markets. AI for drug discovery can be put in that category.

Of course, extreme skepticism and extreme optimism aren’t the only ways to look at emerging technologies such as AI for drug discovery.

There’s always the hope – realized infrequently but not never – that a platform will truly prove to be the “gift that keeps on giving,” that a new approach will generate a series of attractive assets that will provide sustained value and ultimately revenue over time.

Unfortunately, this rarely materializes. A few disappointing examples I remember vividly, because they occurred at around the time I was there, was Merck’s acquisition of the GlycoFi platform (which was going to leverage a yeast-based platform to accelerate the development of biologics including “biobetters”) and Merck’s acquisition of Sirna Therapeutics, an RNAi company, which it acquired in 2006 for $1.1 billion and subsequently unloaded, eight years later, for $175 million to Alnylam Pharmaceuticals, with little if anything to show for either of these efforts.

Pharma executives have long memories for episodes like this. Battle scars from those disappointing acquisitions tend to remain sensitive for many years after the fact. They reinforce a sense of skepticism that the next shiny object will blow open biology or radically change drug discovery and development.

Tech investors, perhaps embolded by their experiences seeing tech deliver results in industries where this was said to be impossible, see in biopharma a high value problem that technology could help solve. 

My sense is that many tech investors view AI-enabled drug discovery as a “missing link” in drug development, and, tactically, believe pharma companies will be so smitten by the results of the approach that someone will acquire the underlying platform at an attractive premium. Meanwhile, at least right now, biopharmas seem to believe (as apparent in the comments earlier) that AI can perhaps slightly improve small aspects of drug development, but are unlikely to really move the needle.  Hence, as I discussed in a recent HBS podcast (listen here, read my summary here), pharma companies would much rather pay market value for individual assets (early-stage drugs), which the pharma feels comfortable assessing, versus paying for a platform of uncertain significance.

This creates both an opportunity and a challenge for AI-driven drug development companies.  The opportunity is that if the AI startup actually can create 10 (or even two) promising early stage drugs in the time it would typically take to create one, they could in theory sell each molecule and pocket a huge amount of revenue.  But the challenge is that many of these companies, while adept at AI, often have far less experience in making drugs, which requires an enormous amount of tacit knowledge and on-the-job experience. 

Once you’ve created a molecule, putting it through rigorous preclinical and clinical trials is also expensive. It’s non-trivial for an AI company to legitimately advance and adequately validate a number of candidate molecules. The pot of gold at the end of the rainbow, of course, is that if an AI-driven company could actually do this, and generate several promising candidates, then I imagine (as I said on the HBS podcast) either they’d make a lot of money supplying compounds to biopharmas, or they’d be in a position to acquire a biopharma for its late-phase clinical development capabilities, plus sales and marketing teams. I imagine this is the dream towards which many AI-driven drug developers are driving.

Presently, it’s difficult to know how the story will end; I am a huge fan of the phenotypic screening which is a key aspect of the efforts for a number of AI drug discovery companies – it was actually what led me to pursue my PhD in yeast genetics, as I was captivated by the power of screens and the opportunity they presented to discovery new biology.  (I suspect a similar excitement is behind the massive CRISPR screens that are driving both Maze and the GSK collaboration with UCSF and UC Berkeley that GSK’s research chief Hal Barron has set up; notably, both efforts involve pioneering UCSF biologist Jonathan Weissman, a former yeast biologist [disclosure: we attended high school together, he was a year ahead of me]); UC Berkeley’s Jennifer Doudna is co-leading the GSK partnership from the Berkeley side.

While it’s easy to see how the slew of AI-for-drug discovery companies might flounder and then fade, I can also envision – very easily – an upside scenario where one biopharma company takes the plunge, and acquires one of the high-profile AI-for-drug discovery startups. It’s exactly the sort of dramatic investment in innovation that CEOs love to do. 

If and when this happens, then other pharmas are likely to follow; as reluctant as so many are to be first movers in an uncertain new area, there’s a strong fast-follower reflex. Executives want to demonstrate involvement in this area before their boards start to press them on why they’re behind. 

If this occurs, the tech VCs investing in these AI-driven drug discovery companies are going to make out like bandits, even as many life science VCs continue to shake their heads, wondering what the fuss is all about.

[Author’s Note: I am thrilled to join the contributor roster of Timmerman Report with my Astounding HealthTech column. The title is a tribute to the iconic, mind-expanding, occasionally prescient (as I discuss here) science fiction magazine of the 1930s and 1940s. I expect to focus on the transformational opportunities and implementation challenges presented by emerging technologies in healthcare, biomedical discovery, and biopharmaceutical development, from next-gen sequencing and cell therapies to digital biomarkers and the meaningful application of AI. As part of this transition, I have now stepped away from Forbes, though my legacy posts will of course remain available on that platform. I have admired Luke for years, and have always appreciated his insight, instincts, and integrity, and it’s especially exciting and meaningful for me to have the opportunity to contribute to what he’s creating. I hope over time to deliver the insight and occasional delight that subscribers have come to expect from Timmerman Report.—David Shaywitz]

You may also like

Hot Topics in Biopharma: Initial Impact of Digital, Data Dilemmas in Clinical Studies, and the Search for ‘New Normal’
ICYMI – Recommended Reading and Listening for Biotech Innovators
You Have Chosen … Poorly: Why Drug Developers Make Bad Decisions
A Glimpse Into the Adjacent Possible: Incorporating AI Into Medical Science