A rash of deals between biopharma and artificial-intelligence companies signals drugmakers’ growing interest in machine intelligence.
In October, Microsoft signed a deal with Novartis to deploy AI and cloud computing across their entire drug development. The five-year deal will see the Swiss pharma use AI tools, initially, in the program focusing on macular degeneration therapies, to improve cell and gene therapy manufacturing, and for drug discovery. Novartis recently signed another R&D collaboration with UK-based BenevolentAI to uncover new drug targets across multiple programs. Also in October, Hong Kong-based AI drug discovery company Insilico Medicine entered a $200 million pact with Jiangsu Chia Tai Fenghai Pharmaceutical to discover compounds against triple-negative breast cancer and use machine learning to hunt for new, affordable treatments for rare diseases. While pharma’s dalliance with AI is still incipient, several converging factors are moving machine learning further onto the drug discovery stage.
“It’s not machine learning on its own, it’s the fact that all these things have converged contemporaneously that have created these exciting opportunities,” says Juan Alvarez, associate vice president of computational and structural chemistry at Merck. It’s the convergence of on-demand molecular synthesis technology to generate leads combined with large-scale cellular screens using nuclease-editing technologies. Furthermore, the large datasets arising from these and other approaches are being fed into rapidly optimized AI algorithms, which in turn are being run on increasingly powerful computers.
To drug hunters, the potential gains are considerable. AI techniques can make predictions about lead properties, molecular absorption and pharmacokinetic and pharmacodynamics distribution. “Our machine learning models predict many of these properties so our chemists spend time making molecules that are more likely to pass hurdles downstream,” explains Alvarez. In practice, AI predictions allow scientists to work with a more manageable set of molecules, about a hundred, rather than the hundreds of thousands molecules from a high-throughput screens. As a result, the turnaround time to make and test molecules shrinks, says David Weitz, head of Takeda Pharmaceutical California and global research externalization. Takeda’s alliances with Numerate and Schrödinger “enable us to progress programs without doing a high-throughput screen,” he says. “It’s a smaller haystack and lets us take a different approach,” and these may now be bearing fruit. Takeda recently expanded its Schrödinger collaboration, increasing the number of targets and drug discovery programs and boosting the payments the computational platform company could receive for each program.
AI won’t be a panacea, but it is more likely to “augment what we do in drug discovery,” Rosana Kapeller said during a June panel at the Biotechnology Innovation Organization conference in Philadelphia convened by Endpoints News. Kapeller, an entrepreneur in residence at Google Ventures and former chief scientific officer of computational drug-discovery-focused Nimbus Therapeutics, compared the current moment to what film giant Kodak faced with digital cameras. Drug companies should embrace AI “instead of trying to push it away,” she said.
That appears to be happening. In June, 17 European industry and academic players came together to form MELLODDY, a consortium to accelerate drug discovery using machine learning. The public–private partnership funded by the Innovative Medicines Initiative and one of the partners, Janssen Pharmaceuticals, with €18.4 million (over $20 million) budget, will establish a decentralized database containing the chemical libraries of 10 participating pharma companies to train algorithms and create models to predict promising lead compounds. “Machine learning has fantastic potential to improve drug discovery efficiency, and we can all benefit from that, because it can digest data in a way that could not be done before,” says Hugo Ceulemans, scientific director of discovery data sciences at Janssen and project leader of MELLODDY. Participants will use this dataset to boost efficiency of target pharmacology and predict a drug’s side effects and pharmacokinetic properties. The consortium will use a blockchain-enabled platform to ensure data security while still sharing “unprecedented volumes” of data, says Ceulemans. That data volume is close to a billion data points: “This is a scale that finally becomes relevant,” he says.
Another emerging trend in industry is to use AI to process images of cells and their responses to potential drugs. Recursion Pharmaceuticals runs 400,000 cell assay experiments per week, says co-founder and CEO Chris Gibson; “And rather than low-dimensional readouts where you get a single piece of information, we take pictures of the cells,” he says. Recursion’s AI tools can then classify the cell’s biological state. Because it has run more than 20 million of these experiments, “we can use that data to ask ‘which molecules should we screen?’ That way we can look at 5,000 compounds that the algorithm believes will work best.” The model improves regardless of whether the screen succeeds or fails, he says: “Iteration and correction is where deep learning algorithms get really good.” Sanofi and Takeda Pharmaceutical have bought into the idea, and Takeda in January 2019 exercised its option to license two Recursion drug candidates and extended the collaboration, originally established in 2017. In July, Recursion raised $121 million in venture capital, bringing its total amount of funding to more than $226 million since 2016.
Once Recursion has hits, the process resembles a traditional drug R&D approach. But Recursion and others, including Insilico, make the bold claim that AI-derived hits are more likely to modulate the core drivers of disease and avoid safety liabilities than those with lead compounds uncovered by traditional drug discovery; in short, they’re more ‘drug-like’ straight out of the gate, shifting attrition earlier in the drug discovery and development continuum, where it’s less expensive. “The human effort down the line is focused on programs impacting patients and not on ones that fail,” says Gibson. Insilico Medicine CEO Alex Zhavoronkov asserts that the molecules his company generates should be better than “something found randomly or optimized from a hit in a high-throughput library.”
For AI to succeed in drug discovery, however, it will require a critical mass of both traditional drug discovery and machine-learning expertise, says Mark Rackham, vice president of chemistry at BenevolentAI. “If you are only a traditional pharma, it’s tempting to do things the old-fashioned way. If you’re a machine learning company without exposure to drug discovery, you can find yourself developing the wrong tools and not answering key questions,” he says.
“AI gives you predictions, on what would be a good target, a good molecule, a good structure,” says Douglas Bonhaus, chief science officer of Neuropore Therapeutics, which allied with BenevolentAI in August. “But you need experimental feedback to refine the model.” Neuropore will run BenevolentAI’s neurodegenerative disease candidates through its in vitro and in vivo assays that model how cells clear misfolded proteins and damaged organelles.
Despite the opportunities AI may afford drug discoverers, “the technology is still pretty new,” cautions Eli Lilly chief scientific officer Daniel Skovronsky. “We lack proof points as an industry—like targets identified and drugs created with AI.” In June, Lilly partnered with Atomwise to apply deep-learning technology predicting small-molecule-protein binding affinity to ten difficult-to-drug targets nominated by the big pharma. “Some of these are targets we’ve been working on for years without much progress,” says Skovronsky. “It’s tough to get a toe-hold, a place to start on a new target. If we can do that [through AI], we know how to take it the rest of the way,” he says. What’s more, a platform like Atomwise’s can address a variety of drug discovery questions, says CEO Abraham Heifets. “Hit identification, potency optimization, off-target toxicity, pharmacokinetic and pharmacodynamics and other questions can be framed in this binding prediction,” he says.
AI technologies’ march into drug discovery and development, however, has had some high-profile stumbles. IBM no longer sells its Watson AI system for drug discovery, for example. The field’s progress was hobbled by datasets that weren’t always easy to combine—AI denizens often refer to the ‘shape’ of data, whether it’s broad and shallow or narrow and deep. Drugmakers say the ability of AI to perform consistently on different data types still needs improvement. In new chemical space—for example, macrocyclic molecules—existing AI models fail. Part of the problem is a general phenomenon in science where failure is less likely to be published than success, says Alvarez. “You need successes and failures in your data so the machine can distinguish between the two,” he says.
AI may indeed accelerate drug discovery, says Skovronsky, “or might be able to attack harder targets and diseases with the same speed and success rate that we have today,” with fewer resources. But “quality is harder,” says Rackham, and that can only really be determined with clinical trials. Today, companies are more apt to focus on leading indicators: fewer failed reactions in a chemical synthesis; fewer molecules failing a battery of assays; fewer and more informative animal studies; and the internal cultural acceptance among researchers of machine learning models.
Near-term AI drug discovery successes, like generating molecular hits, are the precursors of what’s to come. Ceulemans likens it to park-assist or automatic-breaking as predecessors to fully self-driving cars. Elsewhere, it will take much longer, he says, but in 10 years AI could touch many aspects of drug discovery. Meanwhile, AI can “be a little black-box-ish,” says Weitz. “We’re a data-driven, follow-the-science organization, so you need to couple that with a fair level of pragmatism,” he says.
Heifets agrees the refrain should be “show me the data,” regardless of whether a molecule is discovered with machine-learning tools or via more traditional routes.