Is machine learning overhyped?

Machine learning is totally overhyped. Machine learning is completely not overhyped. Machine learning is transformative. Machine learning is a tool like any other.

Ask 10 chemists what they think about the promise of machine learning, and you’ll get 10 different answers. That might make for a lively Twitter debate, but the discussion also has serious implications.

If machine learning is less valuable than it’s been claimed to be, says George Schatz, a physical chemist at Northwestern University, “people end up wasting time and effort” testing it in their labs. Scientists who invest training, lab time, and money into machine learning could find themselves in a difficult position if the tool doesn’t solve a problem as promised.

On the other hand, if machine learning is the wave of the future, chemists who aren’t using it risk falling behind their peers.

It probably won’t be possible to definitively answer the question “Is machine learning overhyped?” without the benefit of hindsight. But after conducting dozens of conversations with chemists, C&EN has found that a consensus about the current state of machine learning emerges.

Machine learning is a category of artificial intelligence that describes a computer’s ability to train on a set of data and then create rules or knowledge from that data. Chemists are often interested in the tool’s predictive power. For instance, if you give a machine-learning algorithm a list of 100 metal alloys and their melting points, can it predict the melting point of an alloy it hasn’t encountered before—potentially even one that’s never been synthesized?

Despite all this promise—or perceived promise—one thing that machine learning isn’t is magic. “Let’s be realistic,” says George Dahl, a computer scientist at Google. “Machine learning is nonlinear regression,” a simple type of statistical analysis in which collected data are “fit” with model parameters. Dahl won a Merck & Co. machine-learning competition while a graduate student in Geoffrey Hinton’s group at the University of Toronto.

Making machine learning sound like something it’s not yet could be bad for the technique itself. If it can’t live up to the bar that’s been set, funders and scientists may decide machine learning isn’t worth their time. “We need the most brilliant minds to feel enticed” to study it and explore its benefits for it to be successful, says Nuno Maulide, an organic synthetic chemist at the University of Vienna.

To explore the space between what some have promised and what machine learning might actually deliver—and to discern among chemists a consensus about the much-ballyhooed tool—C&EN has examined some of the fields where it’s generating the most enthusiasm and skepticism.

Drug discovery

Pharmaceutical companies were natural early adopters of machine learning for a few reasons. Drug companies could afford computing power when it was still out of reach for many chemists, and they had reams of data about small molecules and biological targets with which to train algorithms.

The industry had motivation in addition to means. Machine learning offered the same promise to speed up and improve drug discovery as combinatorial chemistry or high-throughput screening, whose stars have since waxed and waned. “Pharma is very hopeful for new technologies because drug development is extremely difficult, and practitioners are always seeking new ways of making success more likely,” says Robert P. Sheridan, a computational chemist recently retired from Merck & Co.

Precisely when drug companies first adopted machine learning depends a bit on how you define it. They’ve used artificial neural network (ANN) algorithms, a simple form of machine learning, in drug design for almost half a century. In 1973 a group of researchers in the U.S.S.R. demonstrated that an ANN could predict the bioactivity of substituted 1,3-dioxanes (Comput. Biomed. Res. 1973, DOI: 10.1016/0010-4809(73)90074-8).

Starting in the 1990s, medicinal chemists used ANNs in quantitative structure-activity relationship (QSAR) models. QSAR models make predictions about a molecule’s properties according to the known properties of other molecules to help decide whether it’s worth pursuing as a possible drug. Random-forest algorithms and support vector machines—also types of machine learning—have now largely replaced ANNs in QSAR modeling.

From that perspective, Sheridan says, it’s not quite right to say that machine learning is overhyped. “ ‘Machine learning in chemistry’ in the sense of QSAR has been used for decades and is demonstrably useful,” he says. QSAR models aren’t perfect, but companies like Merck continue to use them because they can help chemists prioritize which sets of molecules to spend their time on, and that saves money and effort.

Pharma has already been through the machine-learning “hype cycle.” Consulting firm Gartner introduced that term in 1995 to describe the way people view new technologies. At first, the innovation rapidly gains attention on the way to a peak of inflated expectations, then it sinks into a valley of disillusionment. As people sort out its limitations and actual abilities, the technology levels out on a plateau of productivity. People familiar with the pharmaceutical industry may recognize those ebbs and flows from hyped innovations of the recent past, like nanotechnology and combinatorial chemistry.

Drug companies don’t seem to have developed any resistance to hype despite these repeated exposures. Deep neural networks are having their moment in drug discovery right now, approaching peak hype, according to Sheridan. Like ANNs, deep neural networks are built to resemble the brain: Information passes through a series of interconnected nodes akin to neurons. Each node analyzes pieces of information in a particular way, then passes them on to its neighbors. It’s how image-recognition software identifies shadows and shapes, then eventually eyes and ears, and finally an individual face. Deep neural networks have more layers of these nodes than their predecessors, and they’ve made possible huge advances in fields like image recognition and natural language processing.

Deep neural networks’ ability to learn even from very complex data makes them especially attractive in pharmaceutical chemistry. Abraham Heifets is cofounder and CEO of Atomwise, which makes deep-neural-network-based software for predicting binding affinities of drug candidates to targets in the body. He calls the introduction of deep neural networks a fundamental change in machine learning.

Some have accused Atomwise of overhyping machine learning’s capabilities. A 2015 TechCrunch article quoted the firm’s cofounder Alexander Levy as saying the Atomwise software allowed him to predict a cure for measles from his living room. In a response to a post about the article on the drug discovery and pharma blog In the Pipeline, Levy suggested the blame lay partly with the reporter: “Wouldn’t it be nice if Atomwise works even half as well as it sounds on TechCrunch?” he wrote. Levy left Atomwise earlier this year.

Heifets seems cognizant of the reputation companies like Atomwise have gained. He’s clear that Atomwise is not predicting a cure for anything. “Our focus is binding affinity. Binding is not a drug,” he says.

Others in the field echo that sentiment. Even if machine learning can accurately predict molecules that have high binding affinities to targets—and even if it can do that better than humans—there are still many other steps to creating a profitable new drug. “The drug discovery pipeline is long,” Google’s Dahl says. Adding machine-learning tools to the discovery process could be like developing a better microscope or assay, he adds: It would have a small but real impact.

Big pharmaceutical companies are nonetheless enthusiastic about deep neural networks, though like Heifets they’re careful to temper expectations. “To imagine AI or machine learning would solve all of the problems is not the way we think about it. It is powerful at discrete tasks,” says Jeremy Jenkins, head of chemical biology and therapeutics data science at Novartis.

Still, chemists are seeing deep neural networks as a way of taking drug discovery to a new level, by unraveling complex data collected from the biology happening inside the human body. “Often, biology is so complicated that it’s difficult to wrap one’s head around,” says Vijay Pande, a general partner at venture capital firm Andreessen Horowitz and a computational chemist at Stanford University. He thinks we humans may have reached our limits of understanding biological data but that machine learning will be able to make sense of the interface where drug chemistry meets the body’s biology.

Jenkins says that ability could let machine learning suggest a molecule or molecules for medicinal chemists to focus their efforts on and get them to potential drugs faster. That sounds a lot like what proponents of earlier iterations of machine learning have promised, and some are skeptical that deep learning will be any different. Sheridan says when his group compared deep neural networks with other machine-learning methods, they found statistically significant improvements in predictive ability. But he says the benefits were almost always modest in the context of the entire drug discovery pipeline.

It may be that while machine learning can benefit drug discovery, it will have greater impacts in other areas of chemistry.

Materials research

“The reason machine learning for drug discovery is hard is because the question is ill posed,” says Leroy Cronin, a chemist at the University of Glasgow. Machine learning has proved it does well with a defined goal, like recognizing a face. But Cronin explains that because humans don’t fully understand what makes a drug successful—as opposed to what constitutes, say, a face—we don’t know what data to give a machine-learning algorithm to make it successful. It’s possible that deep neural networks’ ability to deal with more-complex data sets could set them apart, but that remains to be seen. Cronin and others think machine learning is likely to have a bigger impact sooner in materials research.

“How does the chemistry affect the material’s property? There’s a direct connection,” says Jillian Buriak, a materials chemist at the University of Alberta and editor in chief of Chemistry of Materials. A crystal’s hardness is a consequence of how its atoms bond to one another, for example. That means machine learning is likely to be more immediately useful for materials research, she explains.

Materials science came later than medicinal chemistry to machine learning, embracing it only in the past decade or so. Still, a number of papers have already shown that machine-learning algorithms can predict molecules or materials with desired properties, sometimes to humans’ surprise. Algorithms have discovered spin-crossover complexes, which are inorganic complexes that might act as switches and sensors (J. Phys. Chem. Lett. 2018, DOI: 10.1021/acs.jpclett.8b00170), and they’ve found metallic glasses made with alloys that hadn’t previously been theorized (Sci. Adv. 2018, DOI: 10.1126/sciadv.aaq1566).

Like in pharma, materials research has long been using techniques that now would be called machine learning. Heather Kulik, the Massachusetts Institute of Technology chemical engineering professor who led the spin-crossover complex research, says that when she started her lab four or five years ago, she told people her goal was to extend what worked in organic cheminformatics to inorganic chemistry. “These days I’d never say that,” Kulik says. “These days I’d say we’re accelerating inorganic discovery with machine learning because that’s something that funders and people who read papers are going to be more responsive to.”

Buriak agrees that materials researchers have been quietly using machine learning for years, even if they didn’t call it by that name. Buriak points to projects like the Materials Genome Initiative, a $500 million collaboration between several federal agencies that started in 2011 to find and produce new materials faster. MGI collected and made accessible materials data to accelerate materials science, especially by computer.

The materials genome concept—the idea that collecting and analyzing a large amount of data could lead to new insights—dates back only to 2002, but in 2016 researchers used it in one of the first demonstrations that machine learning could benefit materials research. The authors built a public database—which included results from failed reactions—and were able to predict vanadium selenite crystallization reactions more accurately than humans could (Nature 2016, DOI: 10.1038/nature17439).

A computer beating humans at some tasks is one thing, but a common refrain from researchers who use machine learning is that computers can’t replace human intuition. Bobby G. Sumpter, a physical chemist at Oak Ridge National Laboratory who’s been working on computational methods since the 1980s, says there’s every reason to believe machine learning can interpolate large amounts of data and make predictions that may be too subtle or complex for human scientists. But to think that machine learning can make predictions far beyond the chemical space of a particular training data set is to ignore fundamental rules of statistics, Sumpter says.

If this worked, we could, for example, predict a structure of a molecule from its observed properties. That’s not possible right now.

Bert de Jong, a computational chemist at Lawrence Berkeley National Laboratory, says that what we now refer to as machine learning is mainly a tool for speeding up computations on large data sets. It’s not capable of extrapolating beyond the knowledge contained in data sets, understanding the physics of the molecules, and truly learning, he adds.

Instead, many see machine learning as helping materials scientists by improving their experimentation. “Machine learning can aid significantly in analyzing and providing meaningful results,” Sumpter says.

He says machine learning is demonstrably good at interpreting images and spectra of compounds and materials, particularly in finding signals among noise close to an instrument’s detection limit. He says it can also help guide experiments in real time. Because machine learning can assimilate and interpret huge amounts of data in milliseconds, AI can adjust inputs and parameters to optimize an experiment as it happens, particularly in a flow-type reactor setup.

If it’s well built, machine learning doesn’t suffer from humans’ bias, and it’s more consistent. To many, that makes it a better fit for designing experiments as well. “Chemistry is messy and complicated,” Cronin says. “Machine learning can help design better experiments,” especially when the number of variables might overwhelm a human, like understanding solvent effects on a reaction.

Reaction discovery

The first thing that comes to many chemists’ minds when they hear the terms “machine learning” and “hype” is retrosynthesis. Nearly since the day that Harvard University’s Elias J. Corey described his concept for strategizing a synthetic route to a target molecule by thinking about key bonds and building blocks, he and others have been working toward computer programs that can plan synthesis.

“They claimed that for so long,” Sumpter says, “so everybody’s been like, yeah, sure.” Despite that past skepticism, Sumpter is among those who think machine learning may in fact put synthesis by computer in reach. But the way chemists working in this field talk about machine learning is more toned down than most might expect.

“I see machine learning as enabling human abilities, not making humans unnecessary, but making humans more efficient in everything they do,” says Matt Toussant, senior vice president for product and content operations at CAS, a division of the American Chemical Society. ACS publishes C&EN.

Toussant says CAS is introducing its retrosynthesis planner, called ChemPlanner, this fall. MilliporeSigma released Synthia (formerly Chematica), another prominent product in this space, in August.

Both programs rely heavily on human experts who created the databases of rules that chemical transformations must follow, drawn from the literature and their own knowledge. Machine-learning algorithms let the programs navigate chemical space using these rules and suggest to the user possible ways to synthesize a target molecule.

Bartosz Grzybowski of the Ulsan National Institute of Science & Technology, Synthia’s creator, says machine learning is just one tool the software relies on. Synthia also uses molecular dynamics, quantum mechanics, and electronic properties to judge how favorable a transformation is or how stable an intermediate is along a synthesis route. Machine learning can’t do everything, Grzybowski says. “Some advanced aspects of organic chemistry require all these other tools. So what I’m advocating is be open, solve the problem, and don’t insist on a specific method.”

Some chemists remain skeptical that these products will offer a significant advantage over the traditional way of planning synthesis, which typically involves a graduate student accessing a database through tools like Reaxys or SciFinder, another CAS product, reasoning out a path forward, and experimenting. Toussant says what chemists will care about is whether a machine-learning algorithm like ChemPlanner makes them more productive and allows them to make more discoveries.

Grzybowski responds to the skepticism with his mantra: “Cook it.” Put the routes that Synthia predicts to the test in the lab and see whether the program can find a better route—or the same route faster—than a human using a database. He has published one paper doing just that (Chem 2018, DOI: 10.1016/j.chempr.2018.02.002), which showed that Synthia can find novel, efficient, experimentally valid routes to targets in just 15 or 20 minutes.

If chemists think they need to use buzzwords like “machine learning” to attract more eyeballs or dollars, Grzybowski doesn’t blame them. “Impact factor is god,” he says. But he says once the hype of machine learning dies away, valuable tools will remain, as previous fads like combinatorial chemistry or genomics have demonstrated. Each of those had its own ride on the hype roller coaster, and while neither lived up to what some people promised, they both remain in use.

Toussant believes we’re near peak hype in machine learning and about to fall into the valley of disillusionment. “But ultimately all technology recovers from the pit of despair,” he says. “I expect machine learning to do the same. I believe in its future.”

Human learning

Despite their differences, the chemists that C&EN interviewed agree: Yes, machine learning is overhyped. No, it won’t cure cancer. Nonetheless, it’s a valuable tool that’s here to stay.

How valuable is up for debate. But in order for chemists to get the most out of machine learning, another thing is clear. Chemists need to change their behavior. That starts with data.

It’s little wonder that pharma and materials science—two areas of chemistry that have made dedicated efforts to create useful databases of chemical properties—appear to be furthest along in using machine learning. Getting clean, comprehensive data to build good training sets has been a problem in applying machine learning to organic synthesis. Information about successful reactions is scattered across journals in all kinds of formats and notations, and information about failed reactions is hidden in old lab notebooks.

“The old rules still apply,” Oak Ridge National Laboratory’s Sumpter says. “Garbage in, garbage out.” Machine learning does best when trained on a large amount of organized data, preferably including negative results. The University of Glasgow’s Cronin says chemists have to learn how to build databases and create descriptors for data so algorithms can learn.

Chemists will also need basic coding skills. Javier Garcia Martinez, an inorganic chemist at the University of Alicante, says chemists’ training must change. “Every Ph.D. student knows NMR and X-ray diffraction,” he says. “The new tools will be machine learning and artificial intelligence.” For chemists who’ve completed their formal training, Garcia Martinez encourages them to educate themselves with free tools available online.

Collaborations will also become increasingly important, many say. Expert organic chemists will pair up with computational chemists or computer scientists to find ways to apply machine learning to their research. Dahl says he’d love to have more chemists come to Google with their data and questions: “I’m happy to try working on it.”

Even the fiercest machine-learning proponents don’t believe it can be useful for chemists without real effort on the molecular scientists’ part to learn new skills, change the way they think about data, and even ask questions differently. If they can, says Joshua Schrier, a computational chemist at Fordham University and Haverford College, “machine learning enables the ordinary chemist to have superpowers.”

Whether that’s overhyping it depends on your perspective. For those chemists who work most closely with machine learning, the excitement they see in press releases and casual conversation can get tiresome. These experts have great faith that machine learning will have a real and lasting impact on chemistry, especially if more people are trained to use it. At the same time, some worry that this tool can’t possibly live up to the highest expectations and that disappointment might hurt progress.

Cronin puts it this way: “Although I say machine learning is overhyped and annoying, I think it’s underused by chemists.”