Future Human
Posts
Isomorphic Labs

Isomorphic Labs

AI Designed Drugs

Andrew Kuzemczak
March 25, 2025

In partnership with

Receive Honest News Today

Join over 4 million Americans who start their day with 1440 – your daily digest for unbiased, fact-centric news. From politics to sports, we cover it all by analyzing over 100 sources. Our concise, 5-minute read lands in your inbox each morning at no cost. Experience news without the noise; let 1440 help you make up your own mind. Sign up now and invite your friends and family to be part of the informed.

Hi friend,

Welcome back to the seventh edition of Future Human! I hope you all enjoyed my dive into the total artificial heart from BiVACOR last week. I, for one, was wildly inspired and spent a chunk of the week researching other cardiothoracic surgeons developing life-changing tech, so look out for more editions on the heart (shocking).

With that said, we are going to venture elsewhere with this edition. I have been reviewing my deep dives since we launched, and I cannot help but notice my hardware bias. With only slight variation, we are essentially six for six with hardware. To offer some explanation, I have always pushed to understand tech that is as directly linked with the patient as possible. At this stage in my career, I get most excited by products which directly save lives. Ideally, there would be zero degrees of separation between a life saved and the product working in, on, or around the patient.

Software is different.

The discoveries made by software go on to save lives or the hours saved by software allow providers to go on and save lives. It is always a degree or two separated from the life saved, but that’s okay. I forgive you, software.

So to challenge myself and give all of healthtech a look, we are going on our inaugural venture into software.

With that, let me ask you:

How comfortable would you be accepting a therapeutic knowing full well humans had little to do with its discovery and formulation — it came to market sooner and may be cheaper — but it was left in the hands of computers?

The Story

Few pursuits are more noble and globally supported than healing of others and treating of disease. That, my friends, is much easier said than done. To treat is far from simple, quick, or affordable.

Therapeutics, specifically, are notoriously challenging to formulate and perfect for human trial. Us mere mortals struggle to accurately predict what properties a molecule will have just by assessing its formula. As a result, we look to lab experiments to understand such characteristics, but this is neither cheap nor rapid. It can take nearly $3 billion and 10-13 years to bring a drug to market in 2025.

British artificial intelligence researcher, Demis Hassabis, recognized this and soon decided to dedicate the might of his team to the problem. Critically, the team also had the might of Google — yeah that Google — behind it.¹^,²

Demis is the founder and CEO of Isomorphic Labs, a startup launched in 2021 to apply AI and computational methods to drug discovery. Prior to Isomorphic Labs, Demis founded DeepMind in 2010, a wildly successful company developing general AI systems. The team had early success, pioneering the field of deep reinforcement learning - a combination of deep learning and reinforcement learning - and using actual video games to test its systems. Early on, it developed the program DQN, which learned to play 49 Atari games from scratch simply by observing the pixels on the screen and teaching itself to maximize the score. Seemingly ‘simple’ work like this rapidly accelerated, attracting Google, which acquired DeepMind in January of 2014 for somewhere between $400 and $650 million (somehow they are still keeping this number a secret).¹^,³

Newly titled, Google DeepMind, would go on under Demis’ leadership to build AlphaGo, which defeated world champions in the game Go, and then, critically, release AlphaFold.

Launched in 2020, AlphaFold was a global wonder — an AI system that accurately predicts 3D models of protein structures — catalyzing a new wave of progress in biology.

Armed with the support of Alphabet, the $2.1 trillion internet behemoth, and a proven technology in AlphaFold, Demis set out to improve drug discovery with Isomorphic Labs.

As mentioned above, the traditional process of drug discovery is too long, expensive, and risky. Isomorphic Labs believes AI can improve it substantially, and they have the talent to execute.

Sir Demis Hassabis, PhD (you read that right…Sir) is now leading not just Isomorphic Labs but also Google DeepMind. Imagine having the leadership and subject matter capabilities to secure Google/Alphabet’s blessing to lead two spin off teams when their options for talent are endless. Oh yeah, and he won last year’s Nobel Prize in Chemistry.

He is joined by multiple former Google DeepMind leaders, such as Colin Murdoch as President, who served in the same role at DeepMind.

Through my research, I have gathered that their team views biology as an information processing system – one that transmits information and maintains structure. From this foundation, they are confident there is a basic underlying structure for the system — an ‘isomorphic map’. If this all proves true, their machine learning algorithm could soon model the map and the principles of biology, leading to drug design.

Isomorphic is Greek for “same shape”, which they say speaks to the parallel between computational models and life itself (I’m not so sure, but we’ll go with it). With the assistance of the models Isomorphic Labs is creating, they aim to remove much of the experimental work from the drug discovery process and conduct the search for new medicines in-silico (through virtual computing environments).

Positively bananas.

“Just as mathematics turned out to be the right description language for physics, we think AI will prove to be the right method for understanding biology.

Demis Hassabis, Founder and CEO

The Tech

Let’s begin this section with a question — why proteins?

For those less in the biological know, proteins are complex and dynamic molecules, encoded by our genes. They are the building blocks of life. From the molecules that detect light so that we can see, to antibodies that fight off viruses, and motors that drive motion in microbes and our muscles — proteins lead the charge. From there, a protein’s structure, the three-dimensional coordinates of all the atoms in the chain of amino acids, can be a key to understanding its function.

That is why we care and why Isomorphic Labs remains so invested.

Isomorphic draws technologically from the AlphaFold model. Recall, it was built to assist researchers in determining the structure and sequence of proteins. All 300,000,000 proteins on Earth are just a combination of ~20 basic amino acids. With that said, the sequence of those amino acids and the respective folding completely alters the function.²

Since their launch in 2021, the team has partnered with Novartis and Eli Lilly to work on new drugs. Then, in May of 2024, they announced with Google DeepMind the release of AlphaFold 3 (what happened to 2? It was released, but simply offered more structures without any details on potential interactions).

What’s the big difference?

Think structure vs. interactions. AlphaFold 3 is available for free for non-commercial research and has been trained on nearly 100,000 known proteins. Where as AlphaFold could tell you how proteins would fold, AlphaFold 3 can predict how proteins fold AND interact with molecules typically found in drugs such as ligands and antibodies (the key next step in drug discovery).⁴

A drug is ‘just’ a small molecule that attaches to a protein inside the body in some configuration to trigger a cascade of events to target a very specific pathology in the body. The AlphaFold 3 iteration can help scientists better understand these interactions, so they can even better grasp side effects and venture into new arenas of protein-drug interactions that may not have been previously possible.

So how does this all work?

Here’s my best stab at it. Isomorphic Labs collects and uploads a massive and varied dataset of molecules to Google Cloud, where their AI researchers continuously train their machine learning models to label every atom and bond and record the relationships between them (think bond angle, atom position in space, etc.). All of that insane amount of data then moves to a team of medicinal chemists who refine the information by visualizing the potential hits (promising products noted by the algorithm) and then running large-scale biophysics simulations on each option. The most promising results are then fed back into the system to improve that initial filter, leaving the medicinal chemists with less and less work with each round.⁴ Below is a visualization that may help:

Illustration by Campbell Medical Illustration

Even with a refined list of potential products and targets, questions remain:

Where on this target can a therapeutic compound bind to achieve the desired effect?
Will a given molecule bind tightly enough?
Will the molecule go where it’s supposed to go within the body?
How long will it take the body to break the molecule down, and will it be able to do so in a safe manner?

When dealing with these questions, human experts often adhere to familiar patterns based on their own experience, design programs they have participated in, or already approved drug molecules. This is helpful, but also limiting when exploring the ridiculous quantity of new molecules. The computational models are quite literally a means to explore the entirety of chemical space.

Despite the massive advancement in protein interaction prediction, there is still a lot to improve on. Therapeutically, we must know that a product is non-toxic, that it gets absorbed by the right tissue, is small enough to enter the intended cells, and will get safely flushed out of our bodies.

Isomorphic Labs has made incredible leaps to identify molecules that could work and will bind appropriately. Next, it becomes a question of accurately predicting side effects and unintended consequences for the body.

The Market

I write a lot on LinkedIn. Monday through Friday, I post about recent healthtech breakthroughs, investments, and acquisitions. Of all the startups I cover across all categories, second only to AI medical scribes, AI drug discovery is the most common focus of startups launched in the last year.

The market is full and growing.⁵ Just to rattle off a few competing teams:

Atomwise: inked partnerships in 2020 with giants Bayer, Bridge Biotherapeutics, and Hansoh Pharma. Two years later, Sanofi paid Atomwise $20 million upfront to launch an up-to-$1 billion-plus collaboration designed to use its AtomNet platform to pursue up to five drug targets.
Generate Biomedicines: agreed in 2024 to apply its Generate Platform in a $1 billion collaboration with Novartis. Generate has also been partnering with Amgen to discover and create monoclonal and bispecific antibody drugs. Secured a $370 million Series B in 2021.
Insilico Medicine: lead AI designed candidate, ISM001-055 for idiopathic pulmonary fibrosis, performed well in a Phase IIa trial, showing improved forced vital capacity at 12 weeks. Pipeline includes 31 disclosed programs, 9 with IND approvals.
Relay Therapeutics: plans to launch a pivotal trial of its RLY-2608 as a second-line treatment for breast cancer in 2025, based on data showing that the drug plus fulvestrant led to clinically meaningful progression-free survival.

The market for computational biology startups is growing so fast that new words are being created for it. I’m not joking. Two years ago, I heard ‘techbio’ for the first time and thought it was a silly error. Turns out, for these computer/AI led startups, ‘biotech’ is too antiquated for them, placing biology too high above their critical computing power. Hilarious.

The market players are not just our Google spin off against some less resourced startups. That would be unfair. Scientists and developers globally have recognized that AI has immense potential for this field and are rapidly working to create their own products. Facebook’s parent company Meta has invested significant resources in this sector and developed its own basic protein-folding model named ESMFold (they should work on that). As of March 2023, the platform states that it can predict 772 million protein structures (okay, I stand corrected).²

Although no AI-designed drug is fully FDA-approved, some have made significant progress. Insilico Medicines, which I mentioned above, announced just this month that their drug candidate for idiopathic pulmonary fibrisius, ISM001-055, has been granted the official generic name Rentosertib by the United States Adopted Names (USAN) Council. Rentosertib is the first investigational drug in which both the biological target and the therapeutic compound were discovered using generative AI.

In December of 2024, Insilico also announced that its drug ISM5939 has received FDA Investigational New Drug (IND) clearance. ISM5939 is an innovative oral small molecule designed to target the enzyme ecto-nucleotide pyrophosphatase/phosphodiesterase 1 (ENPP1) and treat solid tumors.

With all of that said, it isn’t all sunshine and rainbows in the AI-designed drug universe. A more affordable and faster version of a ridiculously expensive and slow process can still require cash and time. Bumps will arise:

Atomwise recently hired a new, veteran biotech CEO - swapping out their tech-focused founding CEO of 12 years as they go through rounds of layoffs to cut costs¹¹
Insilico has many developmental candidates, but needs to consider an IPO now to raise more funds in the short term¹²
Relay Therapeutics laid off 10% of its workforece in October to streamline processes and shrink its current burn of $90+ million per quarter¹³

AI discovered molecules still have to jump through hoops, biologically and economically. The product must make it through the gut without being immediately broken down by the liver; work in a particular organ such as the kidney without disrupting other organs; avoid binding to and incapacitating any of the thousands of other proteins in the body that are important; and break down and leave the body before drug levels become potentially dangerous. AI is improving in order to predict all of this, but it is not perfect yet.

There is no doubt this is one of the most promising places to be in medicine. AI drug discovery is small but among the fastest growing sectors. The global AI in drug discovery market was valued at only $1.7 billion in 2023, but is growing at a CAGR of 21-30%, reaching a valuation of $12+ billion by 2033 (that CAGR dwarfs any market we have analyzed thus far).⁶^,⁷

Regardless of growth potential, it will not be easy to remain solvent, achieve commercial viability, and demonstrate a robust revenue model. Let’s be honest though. If anyone can do it, it’s probably the team led by a Nobel Laureate with Google on its side.

The Sick

Less than 22% of the 18,500 recognized diseases in the world have FDA-approved treatments. Of the 78% without a cure, half are rare diseases (meaning <65 out of 100k people are affected worldwide). This category includes conditions such as Huntington's disease, sickle cell disease, and many forms of cancer.⁸

The combined impact of rare disease is enormous, with 300 - 400 million people affected globally. Despite this, >95% of those conditions do not have a treatment and only 14% of NIH funding goes to their R&D. Something must be done.

The barrier is often the financial viability of the product which can only help a few thousand patients. Governments have attempted to incentivize this drug discovery (think Orphan Drug Act of 1983), but unless the drug development process becomes cheaper, the issue will never be truly solved.

Not to now state the obvious, but a computer led discovery system from Isomorphic Labs could allow for:

Faster development of treatments: predicting which drug candidates are most likely to succeed, allowing researchers to focus on the most promising compounds
Targeting hard-to-treat conditions: certain cancers, neurodegenerative disorders, and rare genetic conditions, lack effective treatments, but AI can uncover novel approaches to treating these diseases or explore drug repurposing, which I am particularly excited about
Precision medicine: design drugs that are tailored to an individual’s genetic makeup, lifestyle, and disease characteristics by analyzing genomic data
Increased drug safety: simulate and predict how drugs will behave in the body, helping identify potential side effects before clinical trials begin

On the topic of repurposing existing drugs, this was one of the first ideas that came to mind when I began diving into Isomorphic weeks ago. Inflammation is the root of most diseases. What if an inflammatory disease and a certain cancer share a common mechanism with similar genes and proteins responsible? Could trained AI platforms link currently approved drugs for one disease with another condition that shares a pathway? Reduce, reuse, recycle, am I right?

Next, when I was looking for recent details from the Isomorphic team on their upcoming goals for treating patients, where do you think I found valuable information?

Where do all successful and rich people go in the winter to flex their achievements and plans to change the world?

Why, Davos, Switzerland, of course!

In an interview with the Financial Times at the World Economic Forum in Davos, Sir/Dr./Lord Demis Hassabis explained that he believed AI could shrink the typical five-to-10-year drug discovery process by as much as 90%.⁹ From there, he was a bit shady. We know they have had partnerships with pharma giants Eli Lilly and Novartis on six drug development programs, but on the question of timeline, Demis said Isomorphic’s first AI-designed drug is expected to enter clinical trials later this year, targeting one of the “big disease areas”.

Hmmmmm. I would venture to say he means oncology, cardiology, or neurodegeneration, but that’s all I got.

Regardless, there is little denying the impact even a 50% shorter development pipeline would have on patient’s lives everywhere, especially those with rare and currently untreatable conditions.

The Economy

Pharmaceutical companies are spending more and more on research and development. The 10 largest pharma giants now pay nearly $80 billion a year to come up with fewer and fewer successful drugs. Ten years ago, every dollar invested in research and development saw a return of 10 cents; today it yields less than two cents.¹¹

Not to be a downer, but this is in part because drugs that are easiest to find and that safely treat common disorders have been found. What remains are largely complicated products for complex and less common diseases. This all returns far less in revenue. The average cost of bringing one drug to market nearly doubled from 2003 to 2013. It has since stabilized slightly, but it does not look to be falling unless something major is introduced.

Enter artificial intelligence.

90% of drugs wash out in one of the phases of human trials. Imagine if a computer could automatically eliminate countless candidates like that in the early stages so we waste less time. Labeling compounds as potentially toxic, weakly binding, or quick to metabolize early on could reduce R&D costs unlike any innovation introduced before.

Even back in 2019, when AI was merely a figment of our imagination, Bristol-Myers Squibb was testing a machine-learning program that they trained to find patterns in data that correlate with CYP450 inhibition (CYP450 is a family of membrane-bound enzymes that metabolize drugs). They found that the program boosted the accuracy of its CYP450 predictions to 95%—a sixfold reduction in the failure rate compared with conventional methods.

Six years later, we are seeing these improvements further validated across the R&D path. Researchers at Johns Hopkins recently said a platform from startup Exscientia can cut the time spent in discovery from 4.5 years to as little as one year, reducing discovery costs by 80% and resulting in one-fifth the number of synthesized compounds as is normally needed to produce a single winning drug.

It is key to recognize, however, that some of these details above were from 6 years ago and we still are without an AI-designed drug. In my opinion, there is no denying the economic (read: savings) potential of AI in drug development, but we are realizing that the promise will take time to come to fruition.

My Thoughts

We are looking at a revolutionary innovation for all patients unlike anything seen before.

Okay, pause. Breathe.

When I pare down my excitement and imagination, I still see the value, but know it will come in time, not immediately. Demis said it best in a recent interview:¹⁴

”AI appears overhyped in the short term but the mid-to-long-term consequences remain underappreciated”

Demis Hassabis, Founder and CEO and Sir and Lord

We probably will not be getting a drug this year or even next that changes the world. With that said, in the next decade, this technology from Isomorphic Labs and its competitors could change all of medicine and pharmaceuticals. While I hope I have to eat my words with a revolutionary product approved in the coming 9 months, my research tells me otherwise.

So there you have it. A balanced take (I hope) on the work of Isomorphic Labs in the world of AI-driven drug discovery. As my first step into software, I know it is far from perfect. As always, I aim to improve and deliver an even better product the next time.

To more lives saved,

Andrew

I always appreciate feedback, questions, and conversation. Feel free to reach out on LinkedIn @andrewkuzemczak.