AI in Drug Discovery
Discussion w/ Thrasyvoulos Karydis - DeepCure co-founder, building in European defense tech, zero-knowledge proofs, modernising agriculture, jobs, and more
This is Startup Pirate #102, a writing about technology, entrepreneurship, and startups every two weeks. Made in Greece. If you haven’t subscribed, join 6,260 readers by clicking below:
AI in Drug Discovery
The following is a conversation with Thrasyvoulos Karydis, co-founder & CTO of DeepCure, a company in AI drug discovery that has raised over $70 million to develop a drug generation engine starting from chronic inflammatory diseases. We covered a number of topics:
His journey from the University of Patras to MIT
Bringing AI and robotics into drug discovery
Why traditional drug development has failed to address autoimmune diseases
DeepCure’s product development pipeline
Limitations of AI in drug discovery
Is a PhD a prerequisite to starting an AI company
Running a startup requires an all-in commitment
and much more.
Thrasyvoule, it's exciting to talk to you today. One thing I like to do with these interviews is get more into the background of guests first.
TK: Happy to be here, Alex! I did my undergraduate in Electrical Engineering & Computer Science at the University of Patras. I wanted to study physics, but since it wasn't so reputable in Greece (which was counterintuitive given the top reputation of the degree abroad), I went into electrical engineering. I got obsessed with computer science and the physical aspect of electronics, so quantum computing became my passion. Later, I also spent time at Télécom-Paristech with an applied group using quantum optics for quantum cryptography.
My next stop was MIT, where I worked with a professor with a track record in quantum computing who had recently switched focus towards gene editing and advanced biology. The same unknowns at the forefront of quantum computing are analogous to molecular biology. Hence, my core research ended up in the intersection of machine learning, physics, and molecular biology to design proteins: How can we design protein therapeutics? Proteins with specific functions like binding to DNA or other proteins to cure diseases. That was when I first collaborated with pharma companies discovering the next generation of drugs.
When does DeepCure come into the picture?
TK: Fast-forward to 2018… A good friend and teammate in our lab, Kfir Schreiber (now CEO of DeepCure), was thinking of leaving MIT to commercialise the research we had been doing. I dropped out of the PhD in late 2018 to join Kfir and start DeepCure, and since then, I have been leading the platform side of things, which includes everything from computational tools to robotic automation.
Give us a 1,000ft view of what DeepCure is.
TK: DeepCure is a biotech company working on autoimmune diseases caused by inflammation. We go from drug discovery to clinical development, figuring out the next generation of small molecule therapeutics for various targets. Our key value proposition is bringing AI and automation into drug discovery so we can target previously intractable or undruggable targets much more effectively.
Drug discovery is the process of finding new candidate medications — a molecule with some effect on a disease — that will later undergo preclinical work, safety and toxicity controls, manufacturing in larger quantities, etc. DeepCure’s focus is drug discovery; it's making sure you have the right molecule, and we think this is the best way to increase the success rate in clinical trials.
Our company's structure is also unique compared to biotech or AI companies. Half of us are biologists, pharmacologists, medicinal chemists, and translational biologists, and the other half are engineers or computer and machine learning scientists.
Chronic inflammatory diseases are the most significant cause of death in the world today, and it's great to hear that you put your time and energy there. Do you think traditional chemistry has failed (to a large extent) to address them properly?
TK: I think the answer is not a clear Yes or No. There are drugs for chronic inflammatory diseases on the market, most of which are antibodies. Antibodies are the immune system's way of protecting you from infections, allergens, and toxins, and are produced naturally. Lab-made antibodies are used to treat certain inflammatory diseases, for example, targeting the cellular response of a broken inflammatory pathway or gene.
However, there are several issues. Antibodies are hard to administer (intravenous injections that typically require regular visits to a clinic). Additionally, the population can develop antibody resistance; hence, they are becoming less effective. And, they have a huge production cost and tend to be expensive for insurance companies and patients.
The alternative is chemistry-driven drugs — small molecules made by chemists. While there’s great interest in small molecule drugs for inflammation, it has been challenging to discover them following traditional medicinal and computational chemistry approaches. In autoimmune diseases, you want to target two proteins coming together, so-called protein-protein interactions, proteins and small peptides, or proteins and DNA. And using traditional tools to find a drug for a target with dynamic phenomena is extremely hard.
As a result, we have a very high failure rate for such big pharma programs. There's the famous ROI number for pharma: 5% of what goes into the clinic makes it to the market. Even beyond that, these diseases require additional risk in primary R&D, which makes them non-viable from a business perspective compared to other programs in a big pharma portfolio.
Therefore, I can't say traditional approaches have failed, but they certainly have yet to generate the impact the market would expect.
I’d love to understand the building blocks of your technology and how you are different from other AI drug discovery companies.
TK: DeepCure develops AI tools to work on autoimmune diseases — mapping proteins, designing novel molecules that interact with them, and creating these molecules. We combine molecular dynamics and quantum mechanical simulations with a generation engine that develops new drugs.
Our fundamental philosophy is building collaborative AI tools and models that work alongside chemists or biologists to design drugs together. We are in the game of building explainable models where every model prediction has an explanation behind it. We have 3D visualisations and tools to ensure users review the output of models, which, in turn, helps us improve our platform, generate more relevant molecules, and allow users to extract insights.
At the same time, we build physics-based AI. We combine the speed of AI models with the accuracy of molecular dynamic simulations and quantum mechanical simulations to generate molecules in silico (experiments performed on a computer). Even if you search or screen (as it's called in drug discovery) the largest available molecule databases, you probably won't find much since these have been designed for non-inflammatory disease targets. This is why it was imperative for us to develop custom molecules.
To that end, we have also built a robotic system that creates on-demand custom synthesis molecules much faster and cheaper than the alternative, which is using chemists in China. This is drug discovery on steroids as the human chemist can generate 1.2 molecules/week while our robot can generate 100 molecules/week because it's a parallelised synthesis on 96 well plates. We currently cover a good part of the most common chemistries used in approved drugs. Still, to fuel further growth and allocate even more resources, we decided to spin out this part of our platform and make it a separate business entity.
What does your product development pipeline look like at the moment?
TK: We now have three drug discovery programs. A program usually focuses on a target molecule, typically a protein, associated with a specific disease that can be addressed by a drug to produce the desired result. For all of them, we have identified novel starting molecules to work with while our major program has moved to pre-clinical development and is on its way to the clinic.
We have another high-profile program in the making that involves a target called STAT6. STAT6 is an important drug discovery target due to its pivotal role in mediating immune responses and its potential in treating a variety of inflammatory diseases. Every pharma company we have talked to wants to license it. We're still in the discovery phase, right before proceeding to animal model efficacies.
Do you see DeepCure bringing these products to market, or is there a plan to cooperate with big pharma later on?
TK: We plan to maintain an internal discovery pipeline for as long as possible, meaning we will not license assets in the early discovery phase. We can partner and co-develop a drug but will only license our technology and assets if they're already in clinical trials.
The most interesting decision we must make soon is how far we drive this ourselves. For a company like DeepCure, I think it wouldn't make sense to build marketing and distribution, and the entire organisation needed post-regulatory approval to have our drugs used worldwide. Big pharma is really good at that. The ideal scenario would be that we are a platform that keeps creating drugs. We will go as far as possible in the clinic to prove that it works, but then, we’d be happy to partner with another company to license and watch them used in the market.
What do you think are the limitations of AI in drug discovery beyond solving hard chemistry problems?
TK: One of the most significant limitations is translation to human biology. Our entire discovery process is based on animal models or in vitro data. You take the biological system out of context (out of the human) and put it on petri dishes or animal models. This is amplified with AI as you use that data to train the models; therefore, we build a pipeline that can design the best drug for a mouse but have no idea whether this will work on humans.
Of course, having a faster and cheaper discovery allows you to continuously experiment, which is very positive. However, you are training your model on a completely different type of data than the one on which you will test the molecule in real life. Especially in the immune system and inflammation, there are huge differences between mice, rats, pigs, and humans.
Multiple attempts are made to address this through engineered tissues from human organs or growing organoids (organs similar to the ones humans have) and testing drugs there. However, getting training data from thousands of people to train a model will never be as easy.
I want to switch gears now… You dropped out of PhD to pursue DeepCure. What's your advice for folks thinking a PhD is a prerequisite to starting a company in a cutting-edge AI field?
TK: A PhD gives you hard and soft skills. What is helpful is that it teaches you resilience and independent thinking. You conduct original research and follow unconventional directions while your papers get rejected, but you keep pushing. This attitude is 100% entrepreneurial. You could get exposed to that elsewhere, but certainly not working on a project with limited scope in a big corporation. In that sense, if you have completed your Master's and are unsure of your next steps, a PhD is a good option.
In many highly technical industries, a PhD is becoming increasingly important if you are acting as the Chief Technology or Science Officer. It allows you to dive deeper into a field, map the universe of possible solutions, and increase the chances of directing the technical team (which will likely include a lot of PhDs, too) in the right direction. Your role as a founder is not to do the hard science but to ensure you spend the company's resources in the right direction.
By all means, a PhD is not necessary to start a cutting-edge AI company. Yet, the most valid concern is answering the following: How can I hire the ultimate expert in the field who not only has a PhD but has been through a system that values the PhD so much? You need to build that trust and respect within your team from the early days, regardless of whether you have a PhD.
In one of your talks, you said you had previously started four startups before DeepCure, none of which took off. Do you think these experiences better prepared the ground to pursue DeepCure?
TK: It's not about success or failure. That is affected by so many external factors. For me, the #1 thing is that you should be 100% committed. You need to give your whole self, not pursue it as a side project or be the type of person with 20 founded startups on their CV. You need to go out there, execute, and see what failed and what didn't. Then, it will stick in your head, and it's the only way to learn and avoid repeating the same mistakes.
I also want to highlight the importance of building connections throughout the journey: teammates, vendors, investors, customers, etc. If you give your whole self, these people will appreciate it regardless of whether you fail. And who knows? You might work with them again in your next attempt.
That was really insightful, Thrasyvoule. Thank you
TK: Appreciate it, Alex.
Jobs
Check out job openings here from startups hiring in Greece.
News
At Marathon, we announced our investment in PolyModels Hub to help pharma deliver breakthrough therapies faster than ever using AI. The team brings a unique mix of chemical and software engineering and they already work with leading pharmaceuticals to model, simulate, and get access to insights across drug development. Read more here.
AI video startup Runway is in talks for a $450m round at a $4b valuation.
Saronic, maker of autonomous military boats, raised $175m led by Andreessen Horowitz.
Software supply chain security Endor Labs announced a strategic investment from Citi Ventures.
Seismos raised $15m for its AI-powered acoustic sensing technology for energy.
Fintech company Plum raised over £2.8m in crowdfunding.
Growth stage fund EOS Capital Partners launched its fund II with the first closing of €219m.
Greece taps AI-powered satellite tech to build wildfire defence system.
Intryc (fka Sentify) entered Y Combinator to build AI to help companies automate their Support QA.
Resources
Building in European defense tech with Dimitrios Kottas, CEO & co-founder at Lambda Automata.
State of cryptography and zero-knowledge proofs with Kostas Chalkias, co-founder & Chief Cryptographer at Mysten Labs.
Reimagining data anonymisation by Pavlos Petros Tournaris, Sr. Staff Software Engineer at Blueground.
Navigating the AI Hype in pharma with Dimitrios Skaltsas, CEO & co-founder of Intelligencia AI.
Modernising the agricultural sector with Ilias Sousis and Petros Sagos, founders of Wikifarmer.
Democratising data with Graph RAG by George Anadiotis, founder of Linked Data Orchestration.
Events
“World of Learning” by LearnWorlds on Jul 23-24
“Incubation Program 2024 Demo Day” by EIT Digital & Found.ation on Jul 25
“Biomedicine, Bioinformatics & Biotechnology Forum” on Sep 1-5
That’s all for this week. Tap the heart ❤️ below if you liked this piece — it helps me understand which themes you like best and what I should do more.
Find me on LinkedIn or Twitter.
Thank you so much for reading,
Alex