Optimizing drug discovery at scale
Drug discovery today is enormously expensive and rarely successful. "Our mission is to fix that," says Thrasyvoulos (Thras) Karydis, chief technology officer and co-founder of DeepCure.
The MIT STEX25 startup aims to deliver highly effective small molecule drugs that can't be found with conventional discovery tools. DeepCure's radical next-generation platform for scaling up and optimizing discovery is based on three pillars: medicinal chemistry, predictive analytics and an automated robotic lab. Each of these pillars leverages machine learning and other rapidly evolving technologies.
Big pharmaceutical companies typically begin their medicinal chemistry programs by searching through a few million molecular candidates, via high-throughput wet lab screens or algorithm-driven virtual screens, says Kfir Schreiber, DeepCure chief executive officer and co-founder.
"You go through millions of different molecules to find your best starting point," Schreiber says. "That sounds like a lot, and it is a lot. But the number of potential drugs that you can develop is estimated to be 10 to the power of 60. So a few million is a tiny, tiny fraction of that. And throughout 120 years of pharmaceutical development, we have been searching pretty much this same space."
DeepCure set out to break that barrier with MolDB, the world’s largest medicinal chemistry database. MolDB covers more than a trillion compounds and can be expanded to a quintillion (10 to the 18th power) molecules, says Schreiber.
We know exactly how to synthesize every molecule in this database in the lab.
Moreover, "we know exactly how to synthesize every molecule in this database in the lab," he says. Until now, that synthesizing step has been a huge obstacle in virtual screening, because after the computer designs a candidate molecule, it might take years for a synthetic chemist to make the molecule.
No existing computational tools were accurate and efficient enough to sort through a drug discovery catalog with 10 to the 18th molecules and give a full profile of each molecular candidate. DeepCure built one, based on machine learning algorithms trained on reactions in organic chemistry.
Full profiles for each drug candidate include 50 properties, such as how well the molecule binds to the target protein, how well the molecule stays away from other essential proteins, how soluble it is and how well it avoids toxicity or unacceptable side effects.
"In traditional discovery, you go through these properties one by one," says Schreiber. "You start with binding and you optimize the potency of the molecule. You then move to the next set of properties and try to optimize those. This is a very inefficient process, because each iteration means that a chemist needs to design a few molecules, synthesize them, test them, get the data back and design the next step. Each iteration can take a few weeks, if not months. In most cases, this process ends up in a dead end, when you realize that you have reached a point where you cannot optimize it anymore."
DeepCure instead employs a multi-parameter optimization strategy to predict the relevant fitness of each molecule. The machine learning software produces a very long list of molecules with 50 properties for each molecule, and then can select the candidate compounds with the best overall profiles.
This discovery process is much quicker than the traditional approach, because it selects compounds that need much less optimization to become a final drug candidate, and the likelihood of success is much higher, Schreiber says.
Most importantly, this multi-parameter optimization paradigm can identify drugs that are practically impossible to find with conventional approaches, he emphasizes.
Traditional drug discovery picks the best candidates at each step. But, for example, the most potent drugs might be too toxic. "We make the best combination discoveries and we end up with drugs that are fundamentally different and fundamentally better," says Schreiber.
We make the best combination discoveries and we end up with drugs that are fundamentally different and fundamentally better.
DeepCure is building a fully automated wet lab in Israel to bring these candidate molecules into testable physical form. The goal for this "molecular foundry" is to generate 5,000 molecules each month. "By its own, this lab is a huge undertaking," says Karydis. "There are not many places in the world that can synthesize molecules at this scale and scope, and collect full data profiles on each molecule."
Presently it takes DeepCure about two months to select molecules, synthesize them, purify them and test them. The lab will cut this process down to a week, which will also allow production to scale up significantly, Karydis says. The foundry also will speed iteration—the company can take results from the lab, add a new property of interest to the search and come up with new molecules within a few weeks.
Additionally, the foundry will deliver a continuous feedback loop of data for improving the machine learning algorithms. "This means that our platform a year from now will be infinitely better than it is right now," he says.
DeepCure initially is focused primarily on cancer drugs, which can address huge unmet needs and take well-known paths towards the clinic. "Within oncology, we try to identify opportunities that traditionally were considered to be undruggable," says Schreiber. The company has five oncology programs in the pipeline, with some compounds progressing from cell studies into animal tests.
The startup's most advanced program goes after a protein that has long been known to be associated with many types of cancer but has eluded all potential drugs so far. Hitting an exact protein target while bypassing other extremely similar targets in that family of proteins is extremely difficult but DeepCure has found multiple candidate compounds that appear to do so.
As graduate students, Karydis and Schreiber studied ways to apply machine learning and data workflows for discovering therapeutics under Joseph Jacobson, associate professor of mechanical engineering. "In 2017 we had some really exciting results applying one of these methods to the Hepatitis C virus," Karydis recalls. "It was the first time that someone could demonstrate success with an algorithm searching a billion molecules."
Schreiber and Karydis launched DeepCure in 2018, with Jacobson as co-founder and chief science officer. Their goal was to build the pharmaceutical company of the future, Schreiber says.
As outsiders to traditional drug discovery, the DeepCure team started with a blank slate in designing a radically rethought process with fast-evolving tools such as machine learning and cloud computing. "I don't think anyone in the world knows how to solve the problems that we're trying to solve," Schreiber says. "Every step, every week, we have to find our way of doing things."
The startup has raised $47 million, with a $40 million round in November 2021 that is funding the automated wet lab and additional cancer programs. Despite the pandemic, this year DeepCure expects to double its staff of scientists and engineers, now numbering 25 around the world.
Joining MIT's STEX25 startup accelerator program has strengthened DeepCure's connections within the Institute's entrepreneurial ecosystem, giving opportunities to engage with other entrepreneurs facing similar startup issues and with pharmaceutical firms that might be partners down the road.
"The next step for us will be to get our first compounds into clinical trials," Schreiber says. "We want to be in a place where our innovation, our technology and our hard work actually help patients."