Why it's getting harder (and more expensive) to make new drugs

70 years of Eroom's Law

Dec 10, 2023

Well, this LinkedIn post got a lot of attention last week. Today's Signal aims to share lessons, themes, and quotes from the threads that followed. Share your thoughts and subscribe to follow along.

On December 5th the FDA approved its 53rd drug of 2023.

You might be thinking: "More than one a week. That's pretty good, right?"

Do you know how much pharma companies will spend on R&D this year? Around $200 billion.

But here's the kicker: the number of new drugs approved per billion USD spent on R&D has fallen around 80-fold since 1950.

Simply put, it costs a lot more money to make drugs today than it did in the past.

Today we'll talk about why this is happening and what we can do buck the trend. But first, let's call this trend by its name...

Eroom's law

Eroom's law says the cost of developing a new drug roughly doubles every nine years.

Sound familiar? That's because "Eroom" is actually "Moore" spelled backwards. Moore’s law is a concept from the 1960s which observes that the number of transistors in a dense integrated circuit doubles every two years.

In their 2012 Nature paper, Dr. Jack Scannell and colleagues first coined this term when they pointed out the alarming trend in drug development. They said there are (4) main causes of Eroom's law:

The 'better than the Beatles' problem, which says there's a progressively higher bar for improvements over existing therapies. It turns out it's pretty tough to compete against our greatest hits from the past (e.g. Lipitor, Aspirin, Ibuprofen).
The 'cautious regulator' problem, which describes the progressive lowering of risk tolerance by regulator agencies that make R&D costlier and harder.
The 'throw money at it' tendency, or the tendency to add other resources to R&D that could lead to project overrun.
The 'basic research–brute force' bias, which is the tendency to overestimate the ability of advances in basic research and brute force screening methods.

So, how do we fix it?

Post-mortems and AI

Dr. Scannell's team suggested two possible solutions. First, that pharmaceutical companies should appoint a Chief Dead Drug Officer, who would be responsible for uncovering the reasons behind a drug failure at each phase of the R&D process, and publish the results in a scientific journal. Collaboration may be a prerequisite for solving Eroom's law.

More recently, biotech companies are now using AI to try to break Eroom's law— and they may be closer than we think. Here are a few companies using AI to tackle speed, cost, and accuracy across dozens of therapeutic areas:

Exscientia, Insilico Medicine, Schrödinger, Recursion, Genentech, Atomwise, Relay Therapeutics, Molecule.one, CellVoyant, CHARM Therapeutics, Isomorphic Labs, Cradle, AQEMIA, Iktos, Multiomic Health, April19 Discovery, Owkin, Deep LS, Delta4ai, BenevolentAI, Pfizer, AstraZeneca, Janssen Inc., TrialReach, Turbine, PrecisionLife, Deep Genomics, Bristol Myers Squibb, Verge Genomics, Immunai, Synthego Corporation, Gero, BioAge Labs, Ochre Bio, BenchSci, Kuano, Standigm, Genesis Therapeutics, Xbiome, xilis, Atavistik Bio, Diamyd Medical, Omniscope, Alto Predict, Vevo Therapeutics

Author's note: In their 2012 paper, Dr. Scannell and his colleagues actually anticipated that Eroom's law would not continue to hold— and they were right. We've seen an uptick in drug approvals over the past decade. This is a great Q&A with Dr. Scannell from March, 2020 that touches on what's changed.

Discussion

THEMES

Okay, at this point your curiosity may be piqued. But if you're thinking AI is the answer to Eroom's law, I'm afraid I have some awkward news. The truth is a bit more nuanced than that. In fact, what I learned from many of you this week is that Eroom's law is merely a symptom of much longer list of challenges in drug development.

For this section I did my best to go through all of your comments and questions to come up with 7 themes worth digging into further. They are:

Pharma Critiques: Insights into current pharmaceutical practices and their shortcomings.
Industry Hurdles: Discussing various systemic challenges in the pharmaceutical sector.
New Paths: Covering innovative alternatives and novel approaches in drug development.
AI Doubts: Skepticism and concerns regarding AI's role in pharma.
AI Advocates: Positive views on AI's potential in drug discovery and regulation.
AI Middle Ground: Balanced opinions acknowledging both the potential and limitations in pharma.
Models Matter: Focus on the need for improved drug development models, including in vivo, in vitro and in silico models, among others.

SELECT QUOTES

Here is a short sample of comments from some of you that give us more context and insight into these themes. Some comments have been shortened or lightly edited for brevity:

On New Paths:

"I think there is a third solution: understanding patient reaponse to drugs. Despite the use of technologies such as NGS by the research community for over a decade (e.g. immunotherapy), this is still not used in routine clinical practice. This is especially important to leverage AI - we are still leaving out cheap and high value data. Guardant Health is a great example of the power of NGS in the clinic. We need to move to cellular medicine and intercept disease." ( Vijay Vaswani, CEO at Omniscope)
"There are mainly 2 reasons why drugs fail in development: compound's intrinsic properties (eg low efficacy, unacceptable safety) and Type 2 error (drug is OK, but the experiment, ie the clinical trials, were badly designed). If any post-mortem is due, then it should be on Type 2 errors committed in development -- however, I doubt that any company will openly come forward with such post-mortems, either internally, because no one wants to be in such a negative spotlight (there is not Biggest Flops award category in traditional R&D awards) or more so publicly given the confidential nature of the data being examined." (Armen Asatryan, MD, MPH, 25+ years in pharma & biotech)
"Disease insights, data and accepting that there will always be responders and non-responders to any treatment is key. Find those responders, through AI, large-scale meta analyses like we did in Diamyd Medical with our antigen-specific immunotherapy for type 1 diabetes, or any other means, and the pathway to approval looks much brighter." (Ulf Hannelius, CEO at Diamyd)

On AI Advocacy:

"Agreed - 52 drugs in a year is dismal. And only a handful of them are truly innovative. Some of the top scientists researching R&D productivity are Alexander Schuhmacher and Oliver Gassmann. On a positive note, we nominated 9 preclinical candidates last year and got 5 human clinical trials with AI. Some were sold." (Alex Zhavoronkov, CEO at Insilico Medicine)
"How about think about the force of AI in another way? Fifty-three drugs approved by FDA in a year. Can AI make regulation authorities more effective and face less risk in evaluating whether a drug is qualified for marketing. Perhaps AI also can help or is helping (I hope) FDA, EMA, NMPA, PMDA, HSA, etc, in regulating?" (Chengyu LIN, Radiologist, CRO Exec)

On AI Doubts:

"AI will not help if big pharma continues to waste time and money on "processes" and endless "decision making" the way they have been doing for the last 25y. AI will be just another area in which to waste more money. A fundamental redo is needed, not adding another buzzword to the mix." (Tomasz Sablinski, 30+ years in biologics & drug development)
"AI will not fix any of the problems of pharma. The main challenge is the very low reproducibility rate of results published in scientific articles from life sciences (eg. most articles are based on super small number of samples N<10, paper mills, etc.). The most optimistic rate is somewhere around 50%. This means that the scientific articles from life sciences (and corresponding databases) are littered mostly with noise and red herrings." (Daniel N.)

On Critiques & Hurdles:

"Bad scales, insensitive measures, lack of appreciation of inflection points of disease trajectory, heterogeneity underlying disease labels for inclusion criteria. Many missed opportunities to identify therapy efficacy in a data driven manner. So much to be done but few really listen or take action." (Mark Gudesblatt, Neurologist)
"One aspect that I would mention is that we are running out of low hanging fruits. For diseases that are easy to treat we have found some solution. What we are left of are the hard ones. The ones that mutate a lot, the ones with targets very similar to healthy peptides, the rare ones, the ones we still don't understand, etc." (Karol Bubała)

On Better Models:

"I'd also add the reliance on poorly predictive animal models. We've run through all the diseases where there was reasonably good translation. Time for better human in vitro & in silico models." (Ellen Berg , CSO at Alto Predict)
"Testing drugs on humans is expensive and has profound ethical implications. Experimental models fall short of faithfully modelling humans as whole organisms. Modelling the entirety of human physiology (and relative drug interactions) is beyond the event horizon even using state-of-the-art-AI - so far AI has been successful or has shown promise tackling specific problems in drug discovery, where there are more data available, and/or model pre-training is more meaningful." (Daniele Merico, VP at Vevo Tx)

Suggested Media

Thanks for reading this week’s Signal. Subscribe, share this post with your network, and let us know what else you’re reading, following and excited about in the comments below.