Eroom's Law and the FDA's Asymmetric Risk API

Silicon Valley looks at the pharmaceutical industry’s staggering 91% clinical failure rate and reflexively categorizes it as a legacy inefficiency waiting for a brute-force computational patch. You can scroll through recent Hacker News debates on breakthroughs in computational biology, or read the noted medicinal chemist Derek Lowe observing that pharma's attrition rate looks utterly alien compared to the auto manufacturing or restaurant industries, and find the exact same pervasive tech-utopian assumption. The operating theory goes that biology is ultimately just a slow, poorly indexed database, and applying sufficient artificial intelligence will finally debug the pipeline.

This fundamentally misapprehends the transaction occurring between pharmaceutical capital and the administrative state. If a modern automaker shipped cars with a 91% defect rate, the market would liquidate them by Tuesday, and a restaurant that poisoned nine out of ten patrons would face literal pitchforks. Pharmaceutical companies operate under entirely different constraints. They optimize for navigating regulatory approval against an asymmetric risk threshold, rather than streamlining a straightforward assembly line.

The dance here is governed by the FDA’s institutional incentive gradient, which structurally penalizes visible tragedies while heavily discounting invisible missed cures. This regulatory interface between scientific progress and public safety was largely codified by the 1962 Kefauver-Harris Amendments, a massive overhaul of federal law passed in the harrowing shadow of the thalidomide crisis. The administrative apparatus was deliberately re-engineered to optimize almost entirely against Type 1 errors. If a regulator approves a novel compound that subsequently causes severe birth defects or acute liver failure in three hundred patients, they will be hauled before a hostile congressional committee and their career will end in highly publicized disgrace.

Conversely, consider a regulator who delays a life-saving Alzheimer’s medication by five years to demand an extra Phase III trial. Thousands of patients will quietly expire in nursing homes during that window. Those deaths are statistically illegible. There are no investigative journalism exposés about the victims of Type 2 errors, because society does not naturally attribute the baseline mortality rate of terrible diseases to bureaucratic friction. Regulators are perfectly rational actors responding to their environment; they heavily prioritize mitigating the legible catastrophes that end careers.

This brings us to Eroom’s Law, the industry’s dark joke that drug discovery becomes slower and more expensive over time despite exponential improvements in technology. The tech industry assumes that if you inject massive computational surplus into a system—say, radically cheaper genomic sequencing or wildly accurate protein folding models—the baseline cost and failure rate of the end product will naturally drop. (You can forgive the naive assumption here. If a software engineer ships a ten-times faster compiler, the team ships more code; if you make drug screening ten times cheaper, surely humanity gets ten times as many cures, right?)

The assumption fails because the state’s appetite for risk mitigation is functionally infinite. In a heavily regulated market governed by an asymmetric risk threshold, any newly generated technological surplus is instantaneously consumed by increasingly esoteric compliance checkpoints. When technology enables a new way to measure a tiny, fractional risk, the regulator does not congratulate you on optimizing the workflow and invite you to enjoy your improved margins. The administrative apparatus instead dictates that you must now run this new test on every compound in your pipeline, and any compound that fails it is dead. The bar simply moves up to absorb the new capability.

To understand why artificial intelligence cannot simply handwave this dynamic away, we have to zoom all the way down from macroeconomic policy to a very specific, physical artifact: the human ether-a-go-go-related gene, or hERG, potassium channel. You will note that biological systems are deeply messy, and the hERG channel is notorious in medicinal chemistry for possessing what researchers politely term a biologically promiscuous inner cavity.

Because of its weirdly accommodating shape, completely unrelated drug molecules have a nasty habit of binding to it by accident. When a drug blocks the hERG channel, it delays the electrical repolarization of the heart muscle. The result is a specific, potentially fatal cardiac arrhythmia known as torsades de pointes. Take the pill. Block the channel. Delay the repolarization. Cardiac arrest. You have successfully engineered a sudden, catastrophic coronary event out of a promising antihistamine.

The hERG channel aggressively frustrates clean computational prediction. Because it is so uniquely promiscuous, simply yeeting structural models through a neural network yields terribly noisy results. You cannot solve it entirely in userspace. You are forced to synthesize the physical molecule, put it in a literal petri dish, and run manual empirical tests on mammalian cells. It is the ultimate meatspace bottleneck, effectively demanding that any purely digital workflow eventually halt to pay a toll to empirical reality.

For decades, pharmaceutical companies occasionally found out their novel drugs caused fatal arrhythmias only after millions of dollars of late-stage clinical trials (or, disastrously, post-market deployment). Eventually, the industry developed the operational capability to screen for this specific obscure toxicity much earlier in the pre-clinical pipeline. Did this lower the overall clinical failure rate of drug development? Absolutely not.

Once the screening capability existed, the FDA formalized it. In 2005, regulators implemented ICH S7B and E14 guidelines, effectively demanding rigorous pre-clinical hERG screening and massive, expensive Thorough QT clinical trials for practically every new drug molecule seeking approval. The moment the industry generated a technological surplus that could identify cardiac repolarization risks, the administrative state consumed that surplus, transforming it into a mandatory, incredibly expensive checklist item.

(You might assume that regulators would weigh this systemic cost against the benefits of faster drug discovery. But the internal logic of the administrative state dictates that if a risk can be legitimately quantified, it must be mitigated, regardless of the financial drag it imposes. A bureaucrat simply cannot afford to ignore a newly measurable vector of failure.)

This is the precise mechanism by which the regulatory state eats innovation. You build a better machine learning model to predict toxicity, and the state requires you to run that model on ten thousand edge cases to prove a negative. You develop a faster way to screen for liver damage, and the state raises the statistical confidence interval required to demonstrate liver safety. You generate computational surplus, and the system immediately captures it, crystallizing a former technological breakthrough into a mandatory, zero-margin cost of doing business.

We are currently witnessing an unprecedented influx of venture capital determined to apply generative AI to drug discovery, confidently promising to break the 91% attrition rate. The capital pouring into computational biology over the next decade will indeed produce computational miracles. But those miracles will not be used to flood the market with cheap, abundant cures. The administrative state will simply use them to demand mathematical proof that your novel oncology drug does not slightly elevate the risk of arrhythmias in a simulated cohort of ten thousand digital mice. The 91% attrition rate will remain stubbornly intact, serving as a permanent monument to the exact equilibrium point where the pharmaceutical industry's cost of capital perfectly intersects with a regulator's intolerance for political embarrassment.