While anxieties about the roles of humans in easily automated processes have been festering for some time, the startling boom in artificial intelligence’s ability and accuracy must compel us to recognize that even scientists are not exempt from replacement. Although computers have been able to surpass humans in storage capacity and speed for many years, human ingenuity has never been fully replicated. In the realm of research, computing increases efficiency by providing the brute force behind complex calculations, database extractions, and statistical predictions, among other things, while human investigators provide experimental direction and theory. What is it that makes humans especially equipped for this purpose? Is it feasible that computers could eventually replicate this skill set, too?
The secret of human knowledge and intuition lies in the architecture of the mind. While most computers have emulated abstract functionalities of the brain, such as its capacity for memory and calculation, computers are not equipped with the ability to learn in the way that humans do. However, in recent years, researchers have become increasingly successful in the construction of artificial neural networks – systems designed with the intricate workings of biological neural networks in mind. As computational power grows and research progresses, neural networks continue to become more relevant.
According to Fortune magazine, as little as five years ago, Google was developing “two deep-learning projects;” today, it houses “more than 1,000.” These neural networks have begun to match and surpass human ability in some arenas, such as in the notoriously difficult board game Go. The current world champion of Go is no longer human, but instead AlphaGo Zero, the machine brainchild of DeepMind technologies. It is not unreasonable to expect that artificial intelligence could be applied to other problems as well, especially those in the natural sciences, whose rules and parameters are inspired by the universe itself. As with Go, chemistry-centered artificial intelligence could potentially lead to entirely new spaces of chemical thought not yet explored by humans.
In fact, structural determination problems, often described as ‘puzzles,’ remain an essential component of almost every chemical research endeavor, and thus a potential target for computational improvement. In order to claim that a molecule holds significance, researchers must verify that they have indeed identified the correct molecule, and a variety of technologies, like nuclear magnetic resonance, can be applied to yield unique, analyzable spectral patterns for unknown structures. Structural determination problems can consume a considerable amount of valuable time, and while some spectral prediction and analysis technology exists, it generally has not been robust enough to replace human analysis.
However, researchers at the University of California San Diego have developed a much more reliable “Small Molecule Accurate Recognition Technology,” nicknamed SMART, to “[streamline] the discovery pipeline for new natural products.” Unlike other technologies, SMART can factor in effects like those of solvent, compound structure, and functional group interplay on a wide variety of compound structures, allowing it to classify unique, specific groups of molecules given characteristic spectral data. How do SMART’s capabilities rise so far beyond those of other networks? It utilizes a deep-learning neural network which, like our own neural ensembles, can learn and parameterize associations in an unspecified way specific to each piece of information. SMART has the potential to increase the efficiency and accuracy of a vast array of chemical inquiries, spanning fields from drug discovery to chemical ecology. This technology was trained on a thorough subset of natural products, such as medicinally active extractions from rhizomes and marine starfish. SMART was eventually able to differentiate novel compounds, including significant natural products such as terpenoids, which are important in the production of diverse biological compounds ranging from steroids to some proteins.
As with Go, chemistry-centered artificial intelligence could potentially lead to entirely new spaces of chemical thought not yet explored by humans.
If machine learning techniques can solve complex problems involved in exploring research questions, can they begin to master those driving the questions themselves– i.e., can they play the role of a researcher? (Human) researchers at Harvard University recently developed a neural network-based “system which, given a set of reagents and reactants, predicts the likely products,” a typical task of chemical experts. This machine learning approach has achieved an accuracy of 80 percent on common textbook problems and has an advantage over previous models in its “flexibility,” allowing it “to learn the probabilities of a range of reaction types.” Like SMART, similar machine-based approaches have the potential to enhance the efficiency and innovation of many chemical labs. However, what happens when, like AlphaGo, they begin to surpass the abilities of experts? Although this model is far from achieving expert ability, the overwhelming speed of the recent boom in deep learning suggests that such a level may not be far off.
Paralleling its current state in other spheres, artificial intelligence is both exciting and ominous for chemical scientists. As mentioned, neural network systems already have the ability to greatly simplify and expand the process of compound and reaction discovery; their capacity to do this, of course, should be expected to increase with time. As artificial intelligence technologies continue to progress, we must increase the dialogue about how we want change to happen before it happens to us.