Matthew P. Greving, PhD, has been interested in the use of computer science for drug discovery for more than two decades. So much so that when Google first launched its email platform, he snagged an electronic address that’s the amalgam of computational science and biochemistry.
“That’s how long I’ve been thinking about how to integrate biochemistry with computer science!” Greving, who is vice president of platform technology and machine learning at iBio, told GEN.
Greving, accompanied by CEO and CSO Martin Brenner, DVM, PhD, met with GEN in a makeshift office—an area with a coffee table and some chairs walled off with two long, perpendicular shoulder-high bookcases near the back corner of their new open-concept start-up style digs in San Diego. The meeting takes place right in the open, and not for a lack of space.
A publicly traded company, iBio is going through a revamp and is being as transparent as possible. “We started this turnaround story maybe two years ago by hiring the right team, bringing the right technology on board, and finally last year moving into our new space,” said Brenner. “We are in the process and final steps to completely switch gears into a new biotech.”
In September 2022, iBio announced the acquisition of certain assets from RubrYc Therapeutics, the company co-founded by Greving, with the pinnacle of which is an artificial intelligence (AI) drug discovery platform and pipeline. Two months later, the company initiated a process to divest its CDMO business and cGMP biologics manufacturing facility, switching its investments over to a pipeline of immuno-oncology assets and AI-based drug discovery. A restructuring could potentially realize about 50% annualized cost savings. A month later, CEO Thomas Isett resigned as the CEO and a member of the board, and Brenner stepped in as interim CEO, in addition to his role as the CSO of iBio.
Drug discovery in three acts
Both Greving and Brenner have backgrounds in mathematics and computer science and are appropriately calculated and cautious in discussing their use of AI. They insist that while iBio is using machine learning to direct antibodies against specific protein epitopes, what they really want to be known for is the real-world application. The tech stack also allows iBio to optimize these antibodies in a way that reduces downstream risk and speeds up the process.
It feels as though they are being quite modest when it comes to their potential innovations in machine learning, which plug into almost every aspect of their three-part discovery process.
First, they’re harnessing an engine for engineering epitopes for elusive targets. “The efficiency, particularly for traditional challenging targets, goes way up,” said Greving, in reference to the computational tool that he developed about six years ago as the foundation for RubrYc. He attributed its success to a serendipitous discovery with Cody Moore—a young scientist who climbed up the ladder at RubrYc and joined Greving at iBio.
Second, by utilizing several publicly available databases and their own in-house generated data on human protein structure, iBio is using a model based on natural language processing. However, instead of working with written language, it’s being used on antibody sequences. Essentially, the processor spits out every computationally conceivable version of an antibody, enabling iBio to build antibody libraries with full human sequence diversity.
“Our libraries can be between many hundreds of thousands, even millions of antibody sequences,” said Greving. “You take one or more of your engineered epitopes, and you incubate that with the library. These engineered epitopes are fast and cheap to make. Now, you can fish out those antibodies that bind those epitopes.”
To make a hard problem harder, Greving and Brenner are going after multi-specific antibodies. “Maybe there are only 10 out of a billion [antibody sequences] that bind to that epitope, and if you use traditional methods, it’s going to be very hard to find 10 in a billion. But if you really focus on that with the engineered epitopes, you can pull those out more efficiently.”
Third, they’ve developed a mammalian display of these antibody libraries, which is critical for scaling up the production of antibodies. To do so, iBio’s pipeline integrates the cells used for production to express the antibodies early in the discovery process. Specifically, they’ve engineered CHO cells to display IgG antibodies to mimic a B cell, and they do single-cell sorting to find the cells that most optimally express a candidate antibody.
“It just accelerates what we can do downstream and allows us to kind of immediately judge the developability of an antibody, so we don’t have to go down one path and realize six months from today that an antibody is going to suck,” said Greving. “Within literally weeks, we can optimize and pull out an antibody that we believe has high developability.”
A mean, lean machine
Currently, iBio is in the preclinical stages, with its most advanced program in the IND-enabling stage. As part of the ramp-up, they’re utilizing outside partners so they can focus on the team and the pipeline.
“I’m a huge fan of lean R&D,” said Brenner. “By really focusing the team on what we’re good at, we can actually run a relatively broad pipeline. We follow the standard rules of drug discovery: screen for antibodies, test them in vitro, do an in vivo screening model to see if there’s proof of concept in vivo, optimize, and then go into more complex translational animal models. From that point on, we can make really quick decisions to put it into IND-enabling and process development. We can go really, really fast, from idea to in vivo proof-of-concept in six months.”
Brenner said in order to be so lean, they’ve tried to hire “bilingual” scientists proficient in both biology and machine learning or data science. “I think what makes us different and successful is we integrated basically both sides of the equation into a single person,” said Brenner. “For example, Alex Taguchi, our director for machine learning is not only designing epitopes; he’s also in the lab. Our head of the platform, Cody Moore, is not only basically running automation. He’s also very well versed in machine learning.” Unsurprisingly, yet remarkably coincidental, Greving said Moore tried to take the exact same Gmail address as him years ago and had to settle for a slightly different version.
According to Brenner, this type of cross-disciplinary scientist is essential for companies using machine learning. “It’s very intriguing to see the first-generation AI-enabled companies are still struggling,” said Brenner. “Big Pharma almost can’t seem to overcome this gap between the two disciplines. So, if you build this by design, you avoid having these silos. You can’t really explain the fuzziness of a biological experiment to somebody who’s used to zeros and ones unless they have grown up with it and have come up through an academic path with it. Maybe those scientists were not available five or six years ago, but they are now.”
This philosophy is literally built into the space at iBio, with one half devoted to dry lab computation and the other to wet lab biochemistry and molecular biology, which has been meticulously designed so that each lab bench is one step in the drug discovery process that feeds into the next. The only separation is floor-to-ceiling glass.
Despite how modest Greving and Brenner are, they are just as ambitious and resolute. They are trying harder and harder to prove what their platform is capable of. “So far, we have not yet reached a point where we have to raise the white flag—we’re at least eight for eight,” said Brenner. “That doesn’t mean we’re going to stay 100% successful. But so far, we haven’t hit a target that was too hard for the platform to solve.”
While it seems like the sky is the limit for iBio, the reality is resources are sometimes limited. Although Brenner and Greving would like to go after infectious diseases and neurodegeneration, the place where their technology can be best applied is immuno-oncology.
“At the end of this year, we want to be known by all of the antibody discoverers that if they have a hard problem to crack, they know where they come to, and that should be us because we have ample evidence in our pipeline that we can actually go after the really hard drug targets where others have tried and have not succeeded,” said Brenner.
Greving is anxious about the potential of machine learning in drug discovery. “My first job was during the internet boom, and I saw all these internet companies with so much potential,” he said. “Well, that bubble mostly burst. I don’t think we’re in a [machine learning] bubble now. Machine learning has really enabled us to do some high-efficiency novel target discovery, and when you’re in the middle of that, you almost get anxiety of ‘We’ve got to get moving!’”