Xinzhiyuan Reports
Five years ago, in 2021, when DeepMind founder Demis Hassabis announced the launch of a brand-new AI drug-discovery company, Isomorphic Labs, the entire pharmaceutical world was filled with curiosity and skepticism: “Can AI really make drugs?”
Today, the five-year deadline has arrived.
Isomorphic Labs President Max Jaderberg personally put pen to paper and delivered the five-year report card: AI has not only unlocked entirely new therapies across multiple disease categories, but has completely upended drug discovery.
In the past, traditional pharmaceutical companies needed years to complete a search of the 10 to the 60th power (10^60) chemical space. Now, AI does it in just days.
After rigorous wet-lab validation, these brand-new molecules “conjured out of thin air” by AI not only have excellent drug-like properties, but have also revealed entirely new biological mechanisms and opened unknown target pockets, and are about to stride toward clinical trials.
Isomorphic Labs declares to all humanity: an industrial revolution that completely overturns the traditional rules of drugmaking has officially erupted.
Dimensional Strike: Tearing Through the 10^60 “Chemical Universe” in a Few Days
To understand how terrifying the report card Isomorphic Labs has delivered is, we must first understand the desperate “astronomical-level barrier” facing traditional drug R&D.
In theoretical drug chemical space, the number of compounds that could become small-molecule drugs is about 10^60. The total number of atoms in our entire solar system is only about 10^57.
It can be said that finding one perfect new drug molecule is harder than finding one specific grain of sand in the entire universe.
Facing such a vast “chemical universe,” what has traditional pharma done?
The answer is: high-throughput screening (HTS).
Pharmaceutical companies spend months or even years using massive robotic pipelines to test molecular libraries containing 100,000 to 1 billion (10^5–10^9) fixed compounds.
Yet these 1 billion molecules, placed before 10^60, are not even a drop of water in the Pacific Ocean.
Even if countless manpower and resources are spent to find a few “hits,” they must then undergo a long and painful multiyear iterative optimization process to balance toxicity, solubility, membrane permeability, and more.
This is why a new drug often takes 10 years, costs billions of dollars, and has a failure rate of more than 90%.
But under Isomorphic Labs’ magic, “years” have been compressed to the extreme into “days.”
Max Jaderberg proudly showed off their secret weapon—IsoDDE (Isomorphic Labs Drug Design Engine).
When facing the same or even harder targets, IsoDDE completely abandons the old clumsy methods.
It does not need to blindly screen that pitiful inventory of 1 billion molecules in reality. Instead, directly inside the computer’s super-brain, it performs an “intelligent search” across the boundless 10^60 chemical space.
In just a few days, IsoDDE can autonomously complete batch molecular generation and iterative evolution, designing “perfect molecules” that not only bind the target but whose physicochemical properties are already ready to face real-world tests.
The long valley of death, regarded as an impassable chasm in traditional R&D, is directly filled in by AI’s compute.
Dissecting Target D: One Graphic to See How AI Crushes Human Experts
To let the public intuitively feel AI’s terrifying power, Jaderberg in the article specially released an optimization comparison chart for “Target D.”
In this case, drug designers only need to input a “requirements specification” into the AI—for example, which target I want to hit, what modulation mode is needed, how high selectivity must be, and what standards solubility and cell-membrane permeability must meet.
The rest is all handed to the AI Design Agent.
In just 2 to 4 days, the AI agent, like a tireless super-explorer, begins wildly exploring chemical space.
If we observe this process on a visualization heatmap, we see an extremely stunning scene:
Traditional molecular library screening can cover only a negligible “small blob” on the heatmap; IsoDDE’s design agent, by contrast, spreads out a giant radar net and directly covers the entire most promising high-desirability region.
Even more stunning, AI understands dynamic balance.
In the early exploration phase, its attention is highly diffuse, broadly probing all kinds of never-before-seen chemical scaffolds; once it locks onto a high-potential region, it immediately narrows attention and begins deep directed exploitation.
In this process, AI precisely controls the holy grail of drug design—“Pareto Frontier.”
Anyone who makes drugs knows that molecular properties are often mutually exclusive.
For example, the stronger the efficacy, often the greater the toxicity; the better the solubility, the worse the membrane permeability may be.
Traditional pharmacists painfully adjust this seesaw for years, while AI in a few days directly pushes to the multi-objective Pareto frontier, finding those “elite molecules” that perfectly balance efficacy and bioavailability.
When researchers compare AI-generated molecules with reference molecules that human experts spent 3 to 5 years painstakingly deriving from literature, the result is crushing. Functional concentration-response curves prove that AI-designed molecules perform extremely well in every dimension, directly becoming the perfect starting point for subsequent development.
Farewell to “Shennong Tasting a Hundred Herbs”: Moving Pharma from Experience to “Precision Engineering”
In this fifth-anniversary article, Jaderberg proposes a highly disruptive core idea: we are moving from the era of “empirical medicine” comprehensively into the era of “exquisitely engineered molecules.”
Over the past hundred-plus years, human drugmaking has in essence still not escaped the empiricist model of “Shennong tasting a hundred herbs.” Whether it is penicillin discovered by accident or targeted drugs found through massive blind screening, all are filled with a large element of “luck.”
We merely “happen” to encounter a molecule in nature or in an existing chemical library that can treat disease, and then try to patch its flaws.
But in Jaderberg’s eyes, future drugmaking should not rely on luck; it should be a “systems engineering” endeavor like building airplanes or chips.
In the dictionary, Isomorphic gives an extremely bold definition of “exquisitely engineered molecule”: a compound designed from scratch using advanced AI with extremely high precision, capable of modulating a biological target according to predetermined specifications while minimizing toxicity.
To achieve this kind of dimensional-strike engineering capability, Isomorphic Labs has built a highly vertically integrated super-company.
Under this roof, the world’s top AI researchers, biologists, chemists, and clinicians sit together. They not only develop models, but also generate their own data, discover new biology, and personally push these potential drugs into the clinic.
In 2024, AlphaFold 2 won the Nobel Prize in Chemistry for solving the century-old problem of protein structure prediction.
As AlphaFold’s direct lineage force, Isomorphic Labs has, over the past five years, infinitely amplified this revolutionary achievement.
The IsoDDE they built is not merely a structure-prediction tool, but an ultimate engine integrating “World Models” and “Frontier Reasoning Models.”
This engine can deeply understand physical and biological environments, possesses powerful zero-shot reasoning ability, and can simultaneously solve complex problems of structure, binding, molecular properties, and toxicity.
Using AI to generate data, train models, lay infrastructure, and finally mass-produce new drugs—Isomorphic is turning drugmaking into a computable, predictable, replicable precision assembly line.
Even more astonishing, AI is not merely spinning within existing knowledge frameworks; it shows shocking “originality.”
Isomorphic Labs officially revealed that through IsoDDE, they have already:
1. discovered entirely new biological mechanisms;
2. designed brand-new molecules capable of opening “previously completely unknown protein pockets”;
3. successfully validated the unique biological functions of these molecules in real wet-lab experiments.
This means AI has gone beyond the stage of “humans teaching it how to make drugs” and begun to guide humans in reverse: “Look, this hidden corner of this protein is druggable, and I’ve even built you the key (molecule).”
These molecules, rigorously validated by real biological experiments, have shown excellent preclinical data and are pushing Isomorphic Labs, in an all-out posture, to charge toward the final battlefield—clinical developments!
Conclusion: The Door to a New Era Has Burst Open
Looking back at the history of technology, whether aerospace, automobile manufacturing, or architectural design, almost every engineering field has been completely overturned by computer-aided design (CAD) over the past 50 years.
Today, this fire of disruption has finally reached humanity’s most complex and most sacred domain: life and health.
This report card delivered by Isomorphic Labs belongs to all humanity.
When the vast chemical universe of 10^60 is calmly measured by AI within days, when drug R&D goes from luck-based “magic” to precise, repeatable “engineering,” a future without incurable diseases seems, for the first time, within reach.
The “iPhone moment” of AI drug discovery has arrived. Five years is only the beginning; the future pharmaceutical revolution will be even crazier and greater than we imagine.
References:
https://www.isomorphiclabs.com/articles/building-a-new-path-to-make-medicines-with-ai
https://storage.googleapis.com/isomorphiclabs-website-public-artifacts/isodde_technical_report.pdf
Editor: David