Drugmakers embracing AI’s utility to uncover hidden opportunities in discovery

For anyone not hip-deep in the hot field of artificial intelligence (AI), its role in drug discovery — or in any endeavor, for that matter — can feel vague, especially relative to the concrete problems drugmakers face daily. The very definition of AI can seem fuzzy, touted at times as capable of doing everything from organizing your photos to driving your car. Yet, in the biopharma space, as dollars flow in and candidates rise up, clarity is surfacing, too, program by program, bringing with it a more down-to-earth explanation of how the technology is reshaping the discovery enterprise.

Andrew M. Radin, chief marketing officer and co-founder of the AI-driven startup Twoxar Inc., told BioWorld Insight that the way his team likes to think of its use of AI is as a decision-making tool tuned to take care of one important step in a bigger process: efficiently identifying promising molecules. “It’s like writing an email instead of sending a letter,” he said. “The message still gets there but empirically it’s faster.”

In that vein, Twoxar’s platform is streamlining the analysis of a wealth of data on existing drugs and small molecules, protein/target interactions, systems biology, ‘omics, genetics and clinical experience. The company’s efforts are filling not only its own pipeline, but those of a growing roster of partners faster than human scientists might otherwise be able to. (See BioWorld Today, Jan. 25, 2016.)

Feeding the software at the heart of its drug discovery platform are data from disparate sources, which is important to create findings in which Twoxar and its partners can have confidence. “If you use one type of data to make a prediction, it’s okay, but there are a lot of false positives,” Radin said. Combining an initial prediction with analyses of independent data found to point in the same direction, however, can deliver a much higher level of confidence, he said.

With $14.4 million in funding so far from Softbank Ventures, Andreessen Horowitz, OS Fund, CLI Ventures and Stanford University’s StartX accelerator, the Mountain View, Calif.-based company has identified preclinical candidates focused on hepatocellular carcinoma and type 2 diabetes. In the diabetes program, its team identified and shortlisted candidates and then generated positive preclinical results in three months as opposed to the one to two years such work can sometimes take.

Alongside its internal endeavors, the company has inked collaborations with some of its first external partners: In February 2017, it agreed to work with Santen Pharmaceutical Co. Ltd.’s U.S. subsidiary, Santen Inc., to identify new candidates for glaucoma. More recently, this month, it agreed to work with San Francisco-based Adynxx Inc. to generate a new understanding of the biology of endometriosis. Adynxx will select candidates and then look to Twoxar to generate “a sensitive prediction of the therapeutic benefit” to help the company arrive at the best one.

Dispelling any expectation that technological wizardry might magically bridge the gap from hypothesis to market, Radin said every indication in which the company works requires an existing animal model to support its forward progress once the AI-enabled candidate identification is complete. “A lot of companies come to us because they have a belief in AI. However, the ones that really move forward into partnering are the ones who are kind of indifferent about it,” he said. “They’re just focused on the efficiency of the program.” Counting Twoxar itself in that camp, he said, “I hope the story I can tell in one to three years is, ‘We launched this program in 2018 and two years later it’s in the clinic.’”

A visual approach

Another company driving discovery by piecing together diverse datasets with AI in new ways is Salt Lake City-based Recursion Pharmaceuticals Inc. Recursion’s co-founder and CEO, Chris Gibson, told BioWorld Insight earlier this year that over the last two decades of target-based discovery, efforts have largely been “based on the premise that you can target one piece of biology that will fix everything.

“I think we’ve found a lot of the diseases where that’s the case, and there are certainly more that we can find. But it’s harder and harder to find cases where there’s that silver bullet.”

Founded at the end of 2013, Recursion is taking a visual approach to its AI-enabled work. The company’s team generates hundreds of thousands of cellular images each week, leveraging computer vision and machine learning to determine when candidate compounds make model diseased cells look more like healthy cells.

“If the cell looks healthy again, it’s potentially telling you that that’s an interesting place to spend your time,” Gibson said. Though the idea of phenotypic screening isn’t new, at Recursion, he said, the company’s approach puts it “in a much higher dimensional space, where I think there’s a lot more of the system of biology being taken into account.”

The company’s most advanced internal candidate, REC-994, is under evaluation for the potential treatment of cerebral cavernous malformation. It was cleared to enter phase I studies in July. Recursion has also established numerous partnerships since its founding, including research collaboration agreements with Takeda Pharmaceutical Co. Ltd. in October 2017 and with Sanofi SA’s Genzyme in April 2016.

Inferring pathways

Another early adopter of AI was Framingham, Mass.-based Berg LLC. The company’s chief analytics officer, Slava Akmaev, told BioWorld Insight that right from the start, its founders sought to build a “very different pharmaceutical company,” one in which research and discovery would take their cues almost entirely from data, ideally freed from dependencies on prior hypotheses. “Often times, we don’t know what we don’t know,” he said.

That thinking was the genesis for the company’s Interrogative Biology platform, which, instead of hypothesizing the mechanism of a disease and focusing on only a few related compounds, tries to profile the entire disease by analyzing various biofluids and cell models as well as clinical information from electronic medical records. The data are then fed to its Berg artificial intelligence engine, which is designed to infer molecular pathways directly from data that may be specific to a certain disease, cell type, or interaction of multiple cell types.

“We try to understand it at a very small, contextual level, where we cannot spend decades trying to decipher these things one gene or protein at a time,” Akmaev said.

Among the clearest examples of the platform in action is the way in which it has shaped development of the company’s lead program, BPM-31510 (ubidecarenone). Though the candidate got its start as a potential topical treatment for epidermolysis bullosa, using its Interrogative Biology platform, Akmaev’s team was able to show it also may have significant potential in oncology. A new formulation of the drug is currently being tested in phase II for the treatment of pancreatic tumors.

The going has not always been easy, though, Akmaev said. When he joined the company about eight years ago, “in many ways, it was extremely challenging to talk about our approach in the community,” where conservative attitudes to the science of drug discovery minimized the company’s approach, he said. Within the last couple years, however, the company’s reception has improved, he said. It could become even warmer as the Berg team gets closer to launching the first clinical program with roots solely planted in its AI-based system biology approach, a feat Akmaev said will likely be achieved in late 2019 or 2020.

Plenty left to discover

With the many ways in which AI-based systems are being used to speed and ease drug discovery, attention to the field is unlikely to wane soon — nor is funding. In April, London-based BenevolentAI Ltd. raised $115 million to scale up its drug development activities, broaden its disease focus and hire new talent. In September, Goldman extended the financing, adding an undisclosed additional amount to the company’s coffers. (See BioWorld, April 20, 2018.)

There’s plenty of new drugs left to discover, but plenty left to discover about AI’s contribution to the endeavor, too. Will it ultimately make drug discovery more effective? Will the therapies to which AI-based platforms point prove to be safer or of higher quality that what can be produced by human effort? Will drugs developed through the cold logics of statistical modeling eventually win a lighter touch or added favor from regulatory authorities?