Today, we announce LabGenius’ £35 million Series B (ref). With this capital, we’ll be doubling down on our antibody discovery platform and therapeutic pipeline, but the benefits of the round go well beyond the cash.

This Series B has provided an opportunity to bring on board some of the world’s most experienced financial and strategic investors. I’m grateful for their backing and confident that, alongside our existing investors, their support will significantly strengthen the business.

In this post, I want to share my personal perspective on why it’s imperative that we develop radically new ways to push the boundaries of scientific innovation and why I believe that the LabGenius team’s work will prove to be an endeavour of great consequence.

Reinventing the scientific discovery process

Our species’ success can be attributed to the capacity that we have to hypothesise and invent. But even this superpower has its limits. Concretely, hypothesis-driven innovation requires an inventor to understand the intricacies of the system with which they’re working. Now, here’s the rub. There are many fields in which hypothesis-driven innovation performs poorly because the human brain simply isn’t wired to grapple with the complexity or dynamics of the underlying system. Protein engineering is one such domain.

When it comes to the discovery of new therapeutic proteins, our in-built capacity for deep understanding, intuition and invention can at best be described as ‘limited’.

This matters. In fact, it matters a lot.

Proteins are the nano-scale actuators of life. They drive almost every cellular process in our bodies and unless we become significantly better at engineering them, we’ll continue to see costly clinical failures and suffer sub-optimal patient outcomes.

To find the protein therapeutics that will give us longer and healthier lives, we need to first break through the ‘cognition barrier’ presented by hypothesis-driven innovation. To do this, we must develop new forms of innovation that do not require a human to understand the underlying system.

I founded LabGenius with the conviction that the future of scientific innovation will be shaped by smart robotic platforms capable of autonomously discovering new knowledge, technologies, and sophisticated real-world products.

At the time, this worldview was considered by many as bordering on heretical but over the intervening period, it’s become clear that this future is all but inevitable.

Over several years, the LabGenius team has built and refined EVA™ — a closed-loop ML-driven robotic system capable of generating hypotheses, testing them in the lab and then using the resulting data to iteratively refine its understanding of how a protein’s design determines its performance.

The beauty of this approach is that it allows you to massively open up the area of design space that can be searched. With many more shots on goal, protein engineers can be empowered to take more risks, explore designs that go beyond our intuition and challenge our biases to the limit.

Through both internal (ref) and partnered programs (ref), we’ve seen firsthand the power of this approach and its ability to consistently deliver high-performing molecules with non-intuitive designs.

But when it comes to what’s theoretically possible for antibodies, we’re barely scratching the surface.

Antibodies and|are robots

Since the advent of the first industrial robot, Unimate, in the 1950s, engineers have produced increasingly complex integrated systems capable of sensing, processing and actuation.

Within the world of antibody engineering, the arc of innovation is showing every sign that it will track a similar trajectory.

Indeed, attend any antibody engineering conference today and you’ll be hard-pressed not to hear terms like ‘AND gate’, ‘OR gate’ and ‘conditional activation’.

This feels somewhat fitting given that Čapek, who coined the term ‘robot’, intended the word to describe chemically synthesised biological machines and was appalled by the popularisation of the term to describe “plate-metal dummies stuffed with cogwheels” (ref).

This trend of antibody complexification is seen most clearly through the steady rise of multispecifics. These antibodies combine several binding units into a single molecule and when designed well, they are capable of orchestrating exquisitely complex biology.

The engineering of new multispecific antibodies is however non-trivial. This is because proteins are flexible dynamic structures and a molecule’s constituent parts are rarely insulated from one another. Tiny, seemingly innocuous, changes to a multispecific antibody’s design often result in large and unpredictable changes to the molecule’s ability to function within the context of a complex biological system.

To grapple with this complexity, LabGenius’ smart robotic platform uses a form of active learning, called Multi-Objective Bayesian Optimization (MOBO), to navigate antibody design space efficiently and find high-performing molecules that are co-optimised for both function and developability.

What’s so powerful about this approach is that it provides a systematic methodology for reliably identifying diverse panels of high-performing multispecific antibodies. For example, in the context of our internal pipeline, we’ve demonstrated how this approach enables the efficient co-optimisation of T-cell engagers for best-in-class killing selectivity, potency, efficacy and developability.

Delivering therapeutic impact through our pipeline

Solid tumours account for 90% of newly diagnosed cancer cases, but few drugs produce durable therapeutic benefits in patients.

Targeted therapies like immune cell engagers and ADCs have significant potential in this space but clinical success has been limited because many of these molecules struggle to differentiate between healthy and diseased cells expressing the same surface markers.

This issue, which is termed on-target off-tumour toxicity, remains a significant issue for many of the targeted antibody therapeutics in development.

For our internal pipeline, we’ve been tackling the challenge of on-target, off-tumour toxicity using an avidity-driven selectivity strategy.

The avidity-driven selectivity strategy is mechanistically intuitive but co-optimising molecules for killing selectivity, in conjunction with potency, efficacy and developability is a real challenge for conventional antibody engineering approaches.

Through the application of active learning, we have successfully demonstrated the ability to improve the killing selectivity of a solid tumour-targeted T-cell engager to a level that is 400 times greater than the corresponding clinical benchmark (ref). Furthermore, we’ve shown that the process yields diverse panels of high-performing T-cell engagers with non-intuitive designs.

The proceeds from this Series B financing will enable us to both build on these capabilities and progress our pipeline of highly selective multispecifics towards the clinic.

A final thought

Yes, this is a moment to think about the future of LabGenius but it’s also a time to recognise the hard work of team members past and present and the role that they have played in getting us to this important milestone. The resolve and resourcefulness that the LabGenius team has shown in the face of every obstacle has been a constant source of inspiration and it’s my rock-solid belief in them that makes me excited for what we can achieve in the years ahead.

For more information on LabGenius, please visit https://labgeni.us

I’m delighted to announce that we’ve raised an additional $15M for LabGenius, extending our Series A round to a total of $25M. Atomico led the round, with participation from our current investors at Lux Capital, Obvious Ventures, Kindred Capital and Inovia Capital.

To add a fund of Atomico’s calibre and reputation, and to have our existing investor base double down, reflects the tremendous technical progress that the team has made over the past 12 months.

This fresh capital will enable us to greatly expand the scope and scale of our operations, as we slash the time, cost and risk of discovering next-generation protein therapeutics.

Since the first protein therapeutic was approved almost 40 years ago, this class of molecules has been used to treat both life-threatening conditions like cancer and chronic diseases like diabetes. Protein therapeutics represent 30% of total pharmaceutical sales. Yet high costs, risks and long timeframes associated with their development have limited the ability to bring new treatments to market. Today, getting just one new protein therapeutic to market costs the industry in excess of a billion pounds and can take more than a decade. LabGenius is building a systematic, machine learning-driven platform that will make the discovery of protein therapeutics faster and yield novel drugs that would not have been found using traditional methods.

Irina Haivas, Partner at Atomico, who is joining our board, had this to say: “The human and economic toll of COVID-19 makes it painfully obvious that we need a faster and better way to develop drugs. R&D in the industry reached a productivity bottleneck many years ago; the only way forward is through technology. LabGenius is taking on the task of applying software and engineering to biology, and doing this in an area of high complexity: proteins. We believe they have the best technical approach, and have been impressed by the team’s ability to execute and attract incredible talent. We’re excited to work with James and the team on their journey to change the industry paradigm”.

Read more in Fortune.

she said : he said

8 classic tales for my daughter and me

The ritual of reading my 16-month daughter, Evie, her bedtime story is one of my favourite parts of the day. She and I are diligently working our way through my childhood books and it’s with great nostalgia that I’m reminded of all those nights where I drifted off to sleep imagining myself as the protagonist of one swashbuckling adventure after another. This is truly the stuff of which dreams are made.

I want Evie to experience the same feeling of immense possibility but like a sledgehammer to the head, I’m struck by the fact that almost all the characters (and invariably the protagonists) in our bedtime stories are male.

Over the past couple of days, I’ve been thinking about how best to dig out classic novels where female characters are actively represented and ideally the protagonists.

To cut a long story short, I cobbled together a quick and dirty script in R and hit up the first 2,500 (of 58,495) books in Project Gutenberg. From a subset of popular books (≥1000 downloads/month), I extracted those for which the term ‘she said’ appears ≥1.5 times more frequently than ‘he said’.

Here are the hits (ordered by date of publication):

1813: Pride and Prejudice by Jane Austen

1865: Alice’s Adventures in Wonderland by Lewis Carroll

1868: Little Women by Louisa May Alcott

1871: Through the Looking-Glass by Lewis Carroll

1905: A Little Princess by Frances Hodgson Burnett

1905: The House of Mirth by Edith Wharton

1908: Anne of Green Gables by Lucy Montgomery

1915: Anne of the Island by Lucy Montgomery

So here begins a father’s quest to uncover tales that, in years to come, will delight and inspire his daughter. For stories that carry us off to sleep are the sparks that ignite our dreams.

Welcome to the age of empirical computation

Agents of innovation

We’re entering an age in which humans will no longer be the sole agents of innovation.

Instead, new knowledge, technologies and sophisticated real-world products will be invented by smart robotic platforms called empirical computation engines (ECEs).

To many in the tech community, it’s a given that empirical compute will cause a seismic shift in the status quo. To everyone else, the question most commonly voiced is “empirical what?”

I’ve written this piece to provide the wider community with a comprehensive overview of empirical computation and the future that lies ahead.

Nothing is invented – only discovered

It’s Christmas, you’re 5 years old and the proud recipient of 18 shiny red LEGO bricks and an accompanying manual.

“To hell with the manual,” you cry! “I’m going to invent something new.”

You diligently set to work and before long you’ve presented an underwhelmed but supportive mother with 3 entirely novel LEGO sculptures (each built from 6 pieces).

“How pedestrian,” Older Brother remarks as he leafs through the discarded LEGO manual before smugly pointing out each of your ‘entirely novel’ designs. With abject horror, you find that the manual describes all 915,103,765 theoretically possible 6 piece assemblies.

In a dizzying moment of clarity, you realise that everything that can exist physically already exists theoretically.

Solution space

Any problem has an infinite number of solutions which exist (albeit theoretically) irrespective of whether they’ve been manifested in the real world. If you can’t get on board with this, don’t read any further.

Ok, so you’re still with me. Good!

For a given problem, the corresponding set of possible solutions live in a hypothetical place called ‘solution space’. Every problem has its own solution space which in turn has its own unique topology. If, like me, you’re a visual thinker, you can picture the topology of this space as a fitness landscape where the best solutions are found at peaks and the worst in troughs.

Now, here’s the interesting bit. If you can map a solution space’s topology, you can rationally search it. As you might expect, the better your search algorithm, the more likely you are to discover a high-performing solution.

After I first adopted this mental framework, I started to realise that all problems are tractable provided you can build a search algorithm capable of intelligently traversing its solution space.

What’s incredible, is that many of these search algorithms already exist and have done so for billions of years! It turns out that our universe has birthed and subsequently been shaped by a menagerie of autonomous problem-solving algorithms. In fact, you and I are both the direct product of one of those algorithms (Darwinian evolution) but more about that later.

Every one of these search algorithms works by recursively generating and evaluating hypotheses. Any given search algorithm can be categorised based on whether it generates and tests hypotheses virtually (e.g. in a brain or on a computer) or empirically (i.e. in the real world).

Type 1 search algorithms

Darwinian evolution is a purely empirical (Type 1) search algorithm in that it both generates and tests hypotheses in the real world. Every new genome that arises through mutation or recombination can be thought of as a new genetic hypothesis. Once formulated, a genetic hypothesis is compiled by cellular machinery and its impact on the corresponding organism’s fitness is measured as a function of reproductive success. As a process, Darwinian evolution has been operating continuously for the past 3.8 billion years and has been responsible for autonomously filling all biological niches with high performing genetic designs.

Type 1 search algorithms are powerful but suffer from several limitations. To exemplify this, consider Darwinian evolution. Firstly, as a process, it never learns from its mistakes. For us, this means that the same harmful genetic mutations arise (de-novo) over and over again. Secondly, Darwinian evolution is unable to design solutions from scratch and instead has to incrementally improve what currently exists. This means that we have to endure a body replete with bizarre and sub-optimal design features. Take for example, the fact that we have to risk choking with every meal because both air and food enter the body through the same opening. Thirdly, the rate of Darwinian evolution (at the population level) is slow because it’s coupled to an organism’s generation time.

Type 2 search algorithms

Biological Type 2 search

The human brain is an organic super-computer capable of both generating and testing hypotheses virtually. We exercise this capability every day when we imagine, run thought experiments and plan. Every time we execute on an idea generated through virtual hypothesis testing, we’re running a Type 2 search.

Think back to the last time that you had a difficult or important conversation ahead of you. If you’re anything like me, you’ll have run through several dozen simulations of the conversation in advance so you were able to best represent your perspective in the face of anticipated counterarguments.

Type 2 search algorithms are fast and resource efficient but they also suffer from a couple of major limitations.

Firstly, the sophistication of rationally formulated hypotheses is constrained by the intelligence of the underlying hypothesis-generator.

Secondly, the ability to evaluate a hypothesis virtually is determined by the accuracy of the underlying real-world model. Building accurate virtual models of highly complex and poorly characterised real-world systems is immensely challenging. To illustrate just how terrible mental models are, just think how often we experience surprise.

I’m fairly confident that the brain instinctively knows that Type 2 search algorithms are often incorrect and that the emotional response we experience when surprised encourages us to refine the underlying world model. This phenomenon perhaps provides a mechanistic explanation for why we enjoy music and humour.

Artificial Type 2 search

Like the human brain, computers are also able to both generate and test hypotheses virtually. As a platform for Type 2 search, the computer has been breathtakingly successful.

The major weakness of this flavour of Type 2 search lies in hypothesis evaluation. This is once again because it’s incredibly challenging to computationally simulate highly complex and poorly characterised real-world systems.

Type 3 search algorithms

For a Type 3 search algorithm, hypotheses are generated virtually (e.g. in a brain or computationally) and then tested in the real world (i.e. empirically). We tend to call the virtual world ‘the world of bits’ and the real world ‘the world of atoms’.

Biological Type 3 search

The most successful entrepreneurs that I’ve met build and optimise their businesses using Type 3 search. This involves relentlessly maturing and re-formulating world models (hypotheses) based on experience (data). This dramatically contrasts the Type 2 search approach (simulate and execute) that dominated the era of the 30-page business plan.

Artificial Type 3 search

An empirical computation engine (ECE) is an artificial system capable of recursively executing a Type 3 search. ECEs have the following properties:

(1) Hypothesis generation — virtually generated hypotheses must have the capacity to accurately capture the dynamics of the underlying real-world system.

(2) Hypothesis testing — the empirical test must accurately capture the dynamics of the system of interest.

(3) Intelligence — the data generated with each iteration must inform the next round of hypothesis generation.

(4) Autonomy — the system must have the capacity to run continuously without human involvement.

Ethical & societal questions posed by the rise of ECEs

Empirical computation is a technology that’s set to deliver incredible benefits to humanity. At the same time, it also presents several important ethical and societal questions.

(1) How will the role of human scientists change?

(2) How will empirical computation impact on the balance of power at both the corporate and state level?

(3) How do we feel about living in a world where we are dependent on technologies and products that we are cognitively incapable of understanding?

Further reading

• Towards Robot Scientists for autonomous scientific discovery by Ross King (Aberystwyth University)
• Fast Forward: Radical Empiricism as the root of Innovation at Zymergen by Duane Valz (Zymergen)
• Modern AI and the fundamental undoing of the scientific method by Nan Li (Obvious Ventures)
• AI robot mixes chemicals to discover reactions by Adam Levy (Nature)
• Robot “Scientist” Helps Discover New Ingredient for Antimalarial Drug by Patrick Caughill (Futurism)
• Automating drug discovery by Gisbert Schneider (ETH Zurich)

Get involved now

Interested in building an ECE that will fundamentally change the nature of scientific discovery?

Great news, LabGenius is hiring.

Each month MedCity profiles an innovative, growing company from across The Golden Triangle of Cambridge, London and Oxford. This month, I gave MedCity’s Kim Watson a sneak peek into the work going on at LabGenius HQ.

LabGenius combines AI, DNA synthesis and robotics to engineer high-value proteins — what does this mean in reality?

Engineered proteins, for example antibodies & enzymes, provide huge economic and societal value in their ability to both treat disease and power important industrial processes. Despite this, the development of these molecules is largely artisanal and relies heavily on humans for both experimental design and execution. This dependence is often limiting because, as a species, we’re cognitively incapable of fully grasping the complexity of biological systems. As a consequence, the development of novel proteins is often both inefficient and prone to failure.

At LabGenius, we think that the key to improving the protein engineering process is to break through this ‘cognition barrier’. To do this, we’re developing EVA — the world’s first fully autonomous AI-driven evolution engine.

Tell us more about EVA.

EVA is our Silicon Scientist. It’s an agent capable of designing, conducting and critically, learning from its own experiments. The learning aspect means that EVA gets continuously smarter as it unpicks the genetic design rules that underpin life. Ultimately, we think that this approach will give us the ability to engineer proteins with both enhanced and entirely novel functionality.

You raised $3.66m seed funding last year. Tell us more about who was involved with this.

Our seed funding has massively accelerated the team’s work. In a few short months, we’ve been able to build our own state-of-the-art research facility and hire some of the best minds in the field. Given the somewhat unconventional nature of our work and the high degree of customisation required, we ended up building our new facility in a former biscuit factory in Bermondsey.

Our seed round itself was jointly led by Kindred Capital and Acequia Capital with participation from Backed VC, Beast Ventures, Berggruen Holdings Ltd, and System.One. The round also included former head of corporate strategy at Microsoft, Charlie Songhurst and prominent angel investor, Tom McInerney.

It was a record year for investment into spinouts last year, why do you think this is?

From a commercial perspective, a startup can be a fantastically effective vehicle through which dedicated teams of talented individuals can solve problems with which large incumbents have struggled for decades. The market recognises this which, in part, explains the growth in investment. Interestingly, much of this growth has been concentrated in and around cities like London where you find a magic mix of talent, capital and underlying infrastructure, such as co-working spaces and meetups.

What is your main aim for the future?

We’ve spent several years developing a powerful protein engineering capability which we’re now using to address some really high-value challenges. Through our industrial partnership, our aim for the future is to translate these novel engineered proteins products to market and in doing so, responsibly create both economic and societal value.

LabGenius seems to be doing something different — do you worry about competitors?

Protein engineering is a high-value sport and so, unsurprisingly, we’re competing against some extremely well-funded companies. This said, I am confident that in this game, smarts will win out over brute force. Ultimately, I believe that LabGenius has got the right team and technology to make a major impact in this space.

What is your advice for other start-ups?

The best advice that I can give to aspiring entrepreneurs is to dream big! I started LabGenius as a PhD student and at the time, my scientific and commercial ambitions were constrained by a perceived lack of resources. In retrospect, this was a big mistake. Over the years, I’ve realised that, that even if a venture is ridiculously ambitious, if you’re working on a meaningful problem with the right team, you have a good chance of finding the resources you need.

Why did you decide to set up in London?

I’m often asked why we went to the effort of building our Central London facility from scratch rather than renting bench space in a biotech incubator. For me, the decision was simple. To build a narrow AI for protein engineering, you need two things. Firstly, you have to recruit an insanely talented and multidisciplinary team of synthetic biologists, automation engineers, software engineers, data engineers and data scientists. Secondly, you have to create an environment in which your team can work together in a fully integrated fashion. For LabGenius, establishing our own purpose-built facility in Central London has enabled us to both recruit exceptional talent and provide a working environment that’s conducive to solving highly multidisciplinary challenges.

Which other companies working in AI excite you?

It’s certainly an exciting time for the field. I’m a huge fan of Zymergen and Recursion Pharmaceuticals both of whom have pioneered the application of machine learning within the biological domain. Here in London, I’m a huge fan of the research going on at DeepMind.

Find out more at https://www.labgeni.us/

The original interview was published here.

Last week, I had the privilege of exploring this idea at the TEDMED conference in beautiful Palm Springs, California. The brief for the talk was simple — share a single idea about which you are passionate in no more than 2 minutes.

Transcript (with minor edits) is also included below. Enjoy :) .

“We’re living in an age where the concept of curing cancer, infectious disease and even ageing is no longer considered absurd. We know that there are answers to these problems — we just don’t have them yet.

But the question that keeps me awake at night is: are we good enough to find the answers? What if our brightest minds are just not good enough? What if collectively, the 9.8 million metric tons of human brain currently residing on Earth are just not good enough?

What I’m going to say, is uncomfortable to hear. Human creativity — namely our ability to develop new art, ideas, materials and therapies is shackled by the limits of our intelligence. There are many problems that we as a species are mentally incapable of understanding and there are diseases that without help, we will not be able to cure.

Yet there is also reason for hope. At LabGenius, we are building autonomous machines with the ability to iteratively learn through conducting their own scientific experiments. Right now, these agents are unpicking nature’s design rules with unparalleled speed and dexterity. Once equipped with this knowledge, these Silicon Scientists will give us the ability to engineer life with unprecedented control.

With this capability we will, for the first time, be able to break through a cognitive barrier that we’re only just beginning to realise existed. In this future, we will harness new materials and therapies that we do not or even cannot fully understand.

In an age where the quest for understanding is our highest priority, depending so much on machines can feel uncomfortable. Yet across the entire breadth of human invention, I believe that this is our future and if we are to cure seemingly incurable diseases, I think we will need to redefine our relationship with artificial intelligence with both humility and hope.”

ABOUT LABGENIUS

LabGenius is a London-based synthetic biology company on a mission to develop radically new proteins by harnessing evolution. The company is run by a multidisciplinary team of synthetic biologists, data scientists, and engineers.