In this Founder Stories talk, Kfir Schreiber, co‑founder and CEO of DeepCure, outlines the systemic bottlenecks that keep small‑molecule drug discovery slow, costly, and failure‑prone, and argues that artificial intelligence alone cannot solve them. He breaks down the discovery pipeline into five stages—library, hit identification, hit‑to‑lead, lead optimization, and candidate nomination—highlighting five core problems: limited and biased libraries, inability to distinguish good from bad hits, reliance on manual custom chemistry, sequential screening cascades that restrict data generation, and too few development candidates. Schreiber then critiques current AI approaches, noting they focus mainly on hit identification and depend on commercial libraries that lack depth and synthetic feasibility. To unlock AI’s full potential, he presents DeepCure’s “molecular foundry,” a fully integrated platform comprising (1) a state‑of‑the‑art AI engine that performs multi‑parameter optimization of ~26 ADME/Tox properties, (2) MolDB, a virtual library of ~10^18 enumerated, synthesizable molecules with built‑in ADME/Tox predictions, (3) an automated robotic synthesis system capable of producing ~5,000 diverse compounds per month, and (4) assay‑automation that screens each compound across ~15 assays, generating hundreds of thousands of data points monthly. He explains how these components close the design‑make‑test‑learn loop, enable prospective model validation, and generate proprietary data to continuously improve the AI models. In the Q&A, Schreiber discusses MolDB’s scaffold diversity (≈5 billion unique scaffolds) versus commercial libraries, the library’s focus on drug‑like properties rather than hits, and the provenance and cleaning of the ADME/Tox training data, emphasizing that prospective, in‑house data generation is a key differentiator. The talk concludes that AI must be paired with enabling technologies—large, synthesizable libraries, high‑throughput synthesis, and automated assays—to truly accelerate drug discovery.
Questions this video answers
MolDB contains about 5 billion unique scaffolds and is built to ensure every enumerated molecule is synthesizable with high confidence. Unlike commercial hit libraries that focus on single‑step chemistries, MolDB is directed toward drug‑like properties and multi‑step synthesis, enabling queries for specific PK, toxicity, and selectivity profiles.
The models start from public databases, but DeepCure spends years cleaning, normalizing, and splitting the data to reflect real‑world performance. Crucially, the company generates its own proprietary data at scale using the molecular foundry, allowing prospective validation and continuous model improvement, which sets it apart from firms that rely solely on literature data.
Chapters
So it's my pleasure to introduce Kfir Schreiber. Kfir is co-founder and chief executive officer at DeepCure Incorporated. DeepCure is on a mission to give every patient hope by accelerating the discovery of better small molecules therapeutics. He leads a team of exceptional scientists and engineers that have developed a machine learning algorithm trained to learn organic chemistry to create a collection of trillions of synthesizable molecules. They use AI to predict and optimize a drug's most important properties, such as potency, selectivity, toxicity, to name a few. They also leverage automated synthesis and robotic testing to create a continuous feedback loop between the AI and the wet lab. Title of his talk is What's Holding Back AI in Drug Discovery. Kfir?
Thank you, Eric. That's not the presentation. Jesus, sorry. Oh, I don't have it. Yeah.
One sec.
You have it? Yeah.
Got it.
All right, so while we get the slides on, first of all, pretty high bar to step up after Pat. And while I completely agree with pretty much everything that was said, we'll have to focus this talk on how do we move from ACD 0, 1, 2, where we are right now, much closer to level 5. So one second. Yeah, should be on. Nope, not this one. This one. Great, thank you. Awesome. So I'm Kfir Shreiber. As Eric said, I'm the co-founder and the CEO of Deepure, which is a company we launched about 4 years ago right here at MIT on the premise of leveraging novel technologies in order to drive better outcomes for patients. And What I would like to talk to you today about is what I think, or at least what my take is on what's currently holding small molecule drug discovery. How can AI provide those solutions that we need and basically make the point that it is not a standalone solution and ask what kind of AI-enabling technologies are necessary in order to unleash the full potential of AI in this space. So let's start by talking about some of the key challenges in our industry at the moment. So as we've all heard time and time again, drug discovery is very expensive. It takes a very long time, but why? So first, we can break it down to the discovery, preclinical development, and development. And while we can all agree that there is so much we can improve in the clinic, I would like to make the case that a radical change is needed already in the discovery phases if you are ever to drive real outcomes for patients. Right now, about 49% of discovery programs will fail already at that stage. And not only that, 92% of the products will fail later on, which means that we are not only failing way too often in discovery, we are also providing a product that is not good enough. It is way too likely to fail in the clinical stages, which means that we have to address this problem. So let's look a bit deeper into that. Every discovery program can be broken down into few— into 5 main steps: a starting library, a hit identification phase, hit-to-lead, lead-opt, and development candidate nomination. So if we start with the first two and look at some of the challenges, we look at the libraries. Typically, we start with a physical library, whether it's HTS, DEL, fragment-based, you name it, and we screen those through a primary binding assay, typically triaging for a secondary one. This gives us the starting points, the hits. Now, it's key to understand that The hits we select at the very beginning pretty much determine the entire fate of the full program. A good hit might lead to a quick, efficient, and eventually more likely to succeed clinical trial, while a bad hit will inevitably lead us down a long, laborious, expensive, and eventually a failed optimization process. Now, unfortunately, even though we saw a few big numbers on the slides before, The current libraries that we have today are still very limited in scale, but they also lack the diversity that we need. Each library has its own chemistry biases that overrepresent specific chemotypes compared to others. And at the same time, we are lacking the fundamental ability to differentiate a good hit from a bad one. When we start to look at the hits we get from those screens, it's very hard to say which is the one that is going to get us to a good development candidate and which one will take us down the wrong way. Second, second to that, once we already identified a few hits, we start the optimization phase, typically involving synthesizing, custom synthesizing hundreds and thousands of analogs of our chemical series and progressing those through a cascade of biological assays going upwards in cost, which is still probably the slowest, most expensive piece of the entire discovery development phase. Now, the sad part is that custom chemistry hasn't changed much in the last 100 years. We are still fully dependent on human chemists synthesizing one molecule at a time in a very slow and very expensive process, like Pat said before. On top of that, the fact that our screening cascades are so sequential limits the type of data and the amounts of data that we can generate in this process. Finally, Even those that succeed to reach a development candidate will typically only nominate one or at best two candidates. As we said before, 92% of those candidates will fail later on, which means that we just don't get enough shots on goal. We need to have more attempts in order to amortize the risks of clinical development. So to recap, there are five fundamental issues. Our libraries are not big and diverse enough. We lack the ability to differentiate a good hit from a bad one. We're too reliant on custom chemistry, on human custom chemistry. The screening cascades are too sequential and limit the data that we generate. And we end up with not enough products. So when you think about that, AI should have been the solution to many of these problems, right? And we have all seen the progress in the field for the last decade or so, and the size of the room can evidence for that. But I would like to make the point that AI is not a standalone solution. The reason we haven't seen the full potential from AI is because it can never work as a single tool in a vacuum. We need to develop complementary technologies to enable AI to accomplish its full promise. So let's look at some of the challenges of AI today. First of all, most or almost all of the machine learning technologies that are being published today focus only on the HIT-ID stage. And while AI or machine learning can definitely help us screen much larger libraries, we are still dependent on commercially available libraries like Enamine Reel, WuXi Galaxy, and others that are still limited in scale and provide a very biased set of chemistries, typically one-step synthesis. But moreover, all of this ignores the biggest elephant in the room, which is still the optimization phase and custom chemistry. We have to address this if we're ever to lower the cost, improve the outcomes of drug discovery. This manual process is still extremely slow, extremely expensive, and has been and became much, much worse in the last couple of years with the pandemic, the shutdowns in China, the war in Ukraine, and so on. Finally, the fundamental impact of AI is limited by the design-make-test paradigm that we have today in drug discovery. As long as we depend on custom chemistry, shipping compounds worldwide, which creates very long cycle times to get data back. We are limiting the speed of those cycles and the ability to learn from one to the next. We also use a variety of different assays being conducted in different labs worldwide, which creates data reliability issues, increases the noise levels, and restricts the type of learning that we can get from those results. And finally, the high cost of downstream screening once we already have a hit prevents us from generating significant amounts of data and stops us from identifying those key limitations, the key liabilities of each molecule at the very beginning. So what can we do about all of this? I would argue that what we need is not another model. Those are AI-enabling technologies that will help us solve those fundamental challenges of AI in drug discovery. And that's pretty much exactly what we have been doing at DeepCure for the last 4 years. We have built what we call the world's first molecular foundry. That's a site fully designed to automate the complete design-make-test-learn cycle from start to finish. It contains 4 main components. Of course, we have our own state-of-the-art AI engine, but also 3 complementary components are MoldDB, the world's biggest virtual library containing 10^18 molecules designed specifically for AI consumption. An automated robotic synthesis platform that is capable of synthesizing 5,000 diverse compounds a month, and an assay automation capability to screen each compound in a wide panel of different assays generating hundreds of thousands of data points a month. So let's start with the AI engine. If at the beginning I said that one of the key challenges is differentiating a good hit from a bad one, this is exactly what we designed our AI engine to do. We leverage multi-parameter optimization in order to predict and optimize about 26 different ADME and TOX properties for each molecule on top of the traditional potency, selectivity, cell activity, and so on. This gives us the ability to identify those good hits at the very beginning. On top of that, we leverage machine learning in order to optimize the chemistry, the synthesis, the reaction planning in order to speed up the later on stages as well. This practically allows our scientists to start thinking about the complete target compound profile from the very beginning and start optimizing exactly for that. We don't have to think about a very potent hit at the beginning. We can jump straight to thinking about the complete profile. But as I said before, the best AI model in the world is not so valuable if you don't have the library to fit that. And for that matter, we built MolDB. MolDB is basically an enumerated library that was trained to learn the foundations of organic chemistry, making sure that those compounds are synthesizable. It contains 10 to the 18 molecules. 10 to the 12 of them are fully enumerated. And on top of the traditional properties you might find in pretty much every virtual library, like molecular weight, log P, and so on, you will find here all the 26 ADME and TOX properties that we predict using the AI engine. And this gives us the ability to select those compounds that are the closest to the TCP from the very beginning. But as I mentioned, maybe the most important piece of MolDB is the fact that it was designed with synthesizability as top of mind. We built the database in a way that guarantees that every molecule is fully synthesizable with very high success rates. And at the same time, we have those synthetic routes. We know exactly how to go and make it. Which is mission critical for us because the next step is to take those molecules selected from ODB by the AI engine and move them on to our robotic synthesis platform. This platform was designed to synthesize more than 5,000 unique, diverse, and complex molecules every month, including multi-step, including purification, everything else that is included in custom chemistry today. Which allows us to pretty much solve the biggest bottleneck, in my opinion, in drug discovery today, the reliance on manual custom chemistry at different CROs or internally. And finally, the assay automation helps us answer that early question of how do we generate data that will help the AI learn? Our assay automation capabilities take every compound synthesized on the synthesis platform and screen those through a panel of about 15 different assays, including tier 1 ADME, tier 1 TOX, and the typical biochemical binding, biochemical activity, cell-based activity, and so on. This means that we can progress the entire design, make, test, and learn cycle within this facility. So to summarize, if you ask me, drug discovery is currently limited by a few underlying issues. And AI has a ton of potential to help us solve and overcome some of these. But at the same time, it will be the AI enabling technologies that will help us drive the full impact of this technology. We need to solve those underlying issues first before we move on to develop a better— another machine learning model. This is pretty much exactly what we are doing at DeepCure, and this is what I think the future of drug discovery is going to look like. Thank you very much.
Thank you, Kafir. We have time for a few questions if anybody wants to come up to the microphone. Hi.
That was a nice talk. Thank you. So my question is about the synthesizability model specifically. So you made some bold claims about, you know, guaranteed, very high confidence of success. As Pat explained, you know, the folks at Enamine do a pretty good job already. We order from them all the time and, you know, we'll see 85, 90% of our list delivered in the expected timeframe. So clearly they're doing a pretty good job of optimizing for synthesizability as well. So I guess the real, what would be the edge that MolDB would have potentially is what you, your claim to diversity. Have you done scaffold analysis between those big libraries and MolDB, and how does it compare?
Yes. So we use multiple ways to analyze the diversity within MolDB. And we currently stand at about 5 billion unique scaffolds within the library. But I would argue that that's not the main differentiation. Absolutely, and I mean, it's doing a fantastic job and, in my opinion, developed a very, very unique know-how around the chemistry in their libraries. We use them all the time. Super valuable product. However, if you look at it, most of it is single-step chemistries, and it's well suited for heat ID stages. It doesn't allow you to look for a molecule that has a specific PK profile, a molecule that will avoid specific tox liabilities, and will be optimized for also the selectivity and potency for a specific target. So if you are adding all of those advanced queries, if you are looking for a molecule that is much more advanced in the discovery workflow, you are not going to find solutions in AnaminReel. And we have the full 32 billion in our database. It's roughly 1%, a bit less, of MolDB. We just don't find answers there because it was never designed to be a library of drugs. It was designed to be a library of hits. And that's the big paradigm shift with MolDB.
So it sounds like it's by construction that you've optimized for certain properties. So you have the handles for certain properties. Already in the library.
Correct. So basically, the way ModDB is being designed is as an enumerated library, but directed enumeration towards desired admin talks properties and constrained by the organic chemistry foundations that we implemented into these models.
Thank you.
Wen Gong, is there anyone online who is— nope.
OK. Really, really good talk there. On your admin talks models, could you elaborate where do you get your data to build the models and How does that differentiate from 10 other companies that use the same literature for that? Yes, absolutely. So we currently have about 26 different admin talks models. We started all of them from public databases. And we did— we probably spent a couple of years developing ways to clean, normalize, and just make this data usable. But I think the biggest change happened when we started thinking about what Regina spoke about today. How do you split data? Like we saw the papers a few years ago claiming 99% accuracy in drug discovery in us. Why isn't this problem solved then? And the answer is because that's not the real-life performance of these models. So that's first of all why every model here is being tested prospectively in a real-life setting. And the numbers I showed before are all prospective. Second, we start generating our own data. And one of the big benefits of having the foundry up and running is the ability to generate data at scale that we just cannot do with CROs, helping us improve those models. So yes, we do a lot of work on developing machine learning and different models and neural networks and whatever. I don't think that's the secret sauce. It's about understanding your data, cleaning it, splitting it so you have a good evaluation, and then going and generating some more.
Great. If there aren't any more questions for now, Thank you so much again, Kfir.
