Accelerating Breakthrough Therapies
Are we getting AI-powered drugs soon?
This is Startup Pirate #114, a newsletter about technology, entrepreneurship, and startups every two weeks. Made in Greece. Here’s what we recently explored:
If you’re not a subscriber, hit the button below and join us 🏴☠️
Last month, I launched a new show, Ad Astra. It’s a series of short documentaries to tell the stories of the most ambitious technologies made in Europe.
We desperately need a culture that promotes innovation and progress in Europe. A culture that unlocks ambition and optimism, celebrating those who venture into uncharted territories. And as the saying goes, culture is upstream of everything else. It’s time to change the narrative in European tech. It’s time for a vibe shift.
The first episode was the story of Concorde, the fastest commercial plane ever flown. If you haven’t watched it yet, check it out.
While the world was building better ad tech or social media algorithms in the early 2010s, a small team in London started a multi-decade Apollo-style project to crack artificial general intelligence. Stay tuned for the story of Deepmind in Episode 2.
Ad Astra, Per Aspera.
Accelerating Breakthrough Therapies
Bringing a single drug to patients may cost more than $1 billion and take over a decade. Can AI help take breakthrough therapies to market faster and with less cost? What are the limitations of AI in drug discovery? Are we getting AI-designed drugs soon?
We explore all these questions with Andrea Dimitracopoulos, co-founder of deepmirror, a spin-off from the University of Cambridge developing an AI platform that pushes the boundaries of what’s possible in drug design, offering a faster, smarter way to develop new drugs for biopharma companies.
Let’s get to it.
Alex: Andrea, I’d love to start by giving a quick introduction to drug discovery. What does a typical drug discovery process look like?
Andrea: A typical drug discovery process starts with the identification and validation of a biological target. The goal is to pinpoint a protein or molecular pathway linked to a disease. Validating that this target truly influences disease progression is a critical and often time-consuming step because of the complexity of biological systems.
Next is hit identification. This is where we screen large collections of compounds, sometimes tens or hundreds of thousands in high throughput screening, to identify molecules that bind to the target in a beneficial way. From there, we move to hit-to-lead and lead optimisation, where medicinal chemists systematically refine the chemical structure of a hit molecule to improve potency, selectivity, and other critical properties that make a molecule a good drug, such as absorption, distribution, metabolism, excretion, and toxicity. This refinement stage can involve making hundreds or thousands of bespoke compounds, where the price per compound can easily exceed $1,000.
Before testing in humans, potential drug candidates go through preclinical studies both in vitro (with biochemical or cell-based assays) and in vivo (animal models) to further assess safety and efficacy. Only then do we proceed to clinical trials, starting with Phase I for safety in a small group of healthy volunteers, Phase II for early efficacy in a few hundred patients, and Phase III for large-scale efficacy and safety with thousands of patients. Phase IV trials continue post-approval to monitor long-term effects.
The whole journey is very long and expensive, estimated to cost more than $1 billion and take longer than 10 years. Unfortunately, about 90% of clinical candidates fail somewhere in the process. Since discovering problems late is extremely costly, a lot of innovation is focused on identifying potential pitfalls as early as possible. Given that only dozens of new drug candidates enter clinical trials each year and even fewer are ultimately approved, improving the chances of success of a drug earlier on can make a massive impact on costs, timelines, and, ultimately, patient outcomes.
Alex: There’s a lot of noise around drug discovery and AI. Many companies raised tons of cash and seem to be struggling now. Which parts of drug discovery do you think are ripe for disruption by artificial intelligence?
Andrea: I see three major areas where AI can truly disrupt drug discovery.
First, target identification and validation. AI-driven tools are helping scientists sift through vast datasets, often incorporating multi-omics data (genomics, transcriptomics, proteomics, and more), to uncover new disease-relevant pathways or proteins. This can speed up the early decision-making on which targets are worth pursuing.
Second, hit identification and lead optimisation. Traditionally, medicinal chemists iterate through designing, synthesising, and testing many molecules, often dozens per month, to juggle properties like potency, toxicity, solubility, and selectivity. It’s a very complex balancing act, and AI can assist by suggesting promising chemical structures, predicting the interaction between compounds and targets, or directly the result of a biological assay, reducing the trial-and-error. This could save considerable time and resources in hit identification and lead optimisation.
Third, patient stratification for clinical trials. One major factor in a drug’s success is ensuring the right patients receive it. AI can help analyse complex patient data, from genomic markers to real-world evidence, to identify which patient subpopulations will most likely benefit. By selecting more appropriate trial participants, we can boost the chances of showing efficacy and reduce the risk of late-stage failures.
Overall, while there’s been a lot of hype around AI in drug discovery, I believe these three areas are where we’ll see more and more evidence of impactful innovation. Our progress will depend on validating these AI models in real-world settings and integrating them seamlessly into the day-to-day work of the scientists.
Alex: How does deepmirror fit the discovery spectrum, and what makes you different?
Andrea: deepmirror was created to address one of the biggest barriers to AI adoption in drug discovery: accessibility. Historically, only large pharma companies could invest in bespoke AI platforms, and smaller organisations often had to rely on external partners and sacrifice significant intellectual property.
Our platform changes this. We empower scientists to iterate rapidly, combining their experience and intuition with AI-driven insights and suggestions.
We’ve built three core pillars:
Structural predictions: Researchers can examine how molecules interact with target proteins in 3D, quickly testing and refining potential designs.
Foundation models for molecular property predictions: We draw on a large, curated database of molecular properties, such as potency, solubility, and metabolic stability, so our predictive models can pinpoint which candidate molecules are worth testing in the real world.
Generative molecule design: Our generative AI tailors novel molecules to each project’s unique requirements, uncovering ideas scientists might otherwise miss.
We are building deepmirror as a co-ideation platform where both scientists and AI can contribute ideas. The AI evaluates them based on the complex real-world data available and the scientists based on their experience and intuition.
Our B2B SaaS business model ensures that customers retain full control of their intellectual property, democratising AI in drug discovery without sacrificing confidentiality or ownership.
Alex: What do you think the platform's impact on real-world applications can be?
Andrea: Our overarching vision is to accelerate breakthrough therapies, compressing what might traditionally take five years of molecule design and testing into perhaps one or two. By guiding scientists to focus on only the most promising candidates, say, 50 compounds instead of 1,000, we can dramatically reduce the number of design–test cycles and save both time and millions in R&D costs.
For example, we worked with the Medicines for Malaria Venture, a nonprofit tackling malaria, to address a high risk of drug-drug interactions in their lead series. Within about an hour, deepmirror proposed an alternative structure that, once tested, showed a 10x reduction in off-target activity while maintaining similar potency against the main target.
We also helped a European biotech accelerate three months of hit identification to just two weeks by leveraging their available data to find alternative hits that could overcome issues with ADMET properties of their current series of compounds.
We’re still early in our journey, but these real-world results highlight the platform’s potential. By cutting down on trial and error and leveraging AI-driven insights, we aim to transform how scientists discover and optimise new drugs.
Alex: Where do you see most of the limitations of AI in drug discovery?
Andrea: One major limitation of AI in drug discovery is data volume. Biology and chemistry datasets simply don’t match the sheer scale of textual data that large language models, like ChatGPT, rely on. However, drug discovery can compensate with smaller but high-quality datasets, which pharma has access to. Finding ways in which different organisations can share their data while protecting their IP could help improve AI models that can benefit the whole industry.
Another challenge is effectively integrating multiple data modalities, from cell-based imaging and chemical assays to genomic and literature data. Companies like Recursion and Valence are pioneering this space, using AI to merge these diverse datasets. Given the sheer complexity of biological systems, approaching drug discovery from many angles can lead to deeper insights.
Alex: When will we see the first AI-powered drugs coming to market?
Andrea: There’s been a lot of excitement around the concept of an ‘AI-made drug,’ but in practice, every current program still involves humans in the loop. Compounds described as ‘AI-powered’ have already shown promising Phase II results, so we’ll likely see one reach the market before long.
That said, what qualifies as ‘AI-made’ can be ambiguous. Maybe AI identified the biological target via knowledge graphs, or maybe it accelerated hit identification. In all cases, however, medicinal chemists have guided the process, reviewed AI outputs, and made key refinements. These contributions aren’t always highlighted in publications. Still, reading between the lines, it’s clear that humans remain a vital part of the discovery workflow and will continue for some time.
Alex: Before we wrap things up… You came to entrepreneurship from academia. You have a PhD in Theoretical Physics & Cell Biology and did plenty of postdoc work afterwards. Do you think a PhD is a prerequisite to starting a company in a cutting-edge AI field?
Andrea: It’s definitely not a prerequisite, but a PhD can be very helpful in certain ways. In AI and other deep-tech fields, so much of your work builds on existing scientific research. Having strong domain expertise and a rigorous mindset can give you a head start, such as helping you understand the latest research or assess what technology is most relevant to your product. It also increases your chances of guiding the technical team in the right direction from the start.
Beyond the technical advantages, a PhD can also add credibility in the early days. For instance, attracting top-tier talent or convincing potential customers and investors can be easier if you have recognised expertise. In our space, many chemists and biologists hold PhDs themselves, so speaking their language and sharing a similar academic background can be a real asset.
Of course, there are successful founders without PhDs who simply partner with domain experts when needed. Ultimately, it’s about building a team with the depth of knowledge and skills required to solve tough technical and business problems.
Alex: I enjoyed our chat, Andrea. Thank you so much!
Andrea: Nice chatting, Alex.
Jobs
A list of 400 startup jobs in Greece from 90 companies hiring! Some great opportunities there. Here’s the list.
News
Keragon (AI-powered healthcare automation) secured a $7.5m Seed round led by Upfront Ventures.
Akina (AI-powered physiotherapy) secured CHF 2.8m led by Freigeist Capital.
Robenso (robotics for recyclable waste sorting) raised €500k led by Helidoni Group.
New Greek LLM launched built on top of Llama-3.1-8B by Athena Research Centre.
Resources
AI’s role in audio storytelling with Dimitris Nikolaou, co-founder and CEO at Wondercraft.
Internships and navigating the career ladder with Anastasios Andronidis, Senior Staff Software Engineer at Worldcoin.
Tracking US mobile satellite service spectrum by Argyris Kriezis, GPS Lab Graduate Researcher at Stanford.
Bringing back Greek scholars, a report by the Deon Policy Institute.
AI in the public sector with Viktoria Kalfaki, Head of Public Sector at Google Cloud.
Foundation models, modular AI architectures, fine-tuning, and the role of synthetic data in post-training with Alex Dimakis, co-founder at Bespoke Labs.
Events
DevOops Athens #3 on Feb 18
Athens WordPress Meetup on Feb 19
Women in Test by Ministry of Testing Athens on Feb 25
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.
Thanks for reading,
Alex
Really interesting take on AI in drug discovery. The way AI is speeding up target identification and lead optimization is exciting, but I can’t help but wonder,will regulatory agencies be able to keep up? If AI can cut years off R&D, but approvals still take forever, how much impact can it really have? Curious to see how companies like deepmirror push things forward