To many, the boom in AI tools born over the last several years has enabled an increase in personal efficiency. Thanks to its ability to sort through large data sets or to teach challenging concepts in simple terms, AI has shrunk the barriers to innovation and interdisciplinary thinking across the board. The benefits of AI extend to chemical research as well, where it has demonstrated the potential to expedite the discovery and optimization of new drug candidates and democratize the burgeoning field of data-guided chemistry.

SOLVE Chemistry is a company embracing the field of predictive chemistry, utilizing specialized high-throughput experimentation and AI tools to accelerate the advancement of scientific progress. Linden Schrecker, CEO of SOLVE Chemistry, reasons that “we’ve solved the low hanging fruit of what we can do building these narrow towers in the scientific landscape,” and now “we are filling in the rest of the landscape with data.” In his view, “the future of scientific discovery and development is going to be data-driven.” With AI and machine learning (ML) platforms now at the disposal of nearly every synthetic chemist, the time is now to fold statistical thinking into the chemist’s repertoire.

One reason why the transition from one-variable-at-a-time designs to statistical approaches is challenging is because some scientists “consider it unscientific when not looking at the effect of one at the time,” Schrecker says. Notably, something a scientist may feel they lose when transitioning to data-based chemistry is “the feeling of knowing what’s going on. It’s very obvious if you change one factor and it changes the outcome, then great, it changes the outcome,” he says. The fear that embracing ML in chemistry will relegate scientists away from the bench and away from experimental certainty, though, is unfounded in Schrecker’s view.

In fact, the converse may be true: if experimental data is to be fed into a model, that data may need to be of even higher quality than in one-factor-at-a-time testing, where outliers may easily be noted and excluded. As Schrecker says, scientists “need to collect diverse data across the domain space,” so that models can be as complete as possible when identifying patterns and interactions. He notes that high data quality from across the domain space will “become even more important as we move towards more of an AI-driven paradigm.” Good data going in becomes good insights coming out, but the converse is also true.

These innovations are two of the largest ways that building a data-centered workflow can advance chemical research. First, on the bench scale, slimming down the number of experiments needed, for example, to evaluate a process route to a drug candidate also trims the amount of time, effort, and chemicals needed to evaluate the process, which provides boosts in efficiency, cost savings, sustainability, and safety in pilot-scale research. Simultaneously, focusing more on data quality for modeling purposes offers more confidence to leadership down the line. As Schrecker puts it, “If you don’t have data you can really trust at the small scale, it makes it harder to make decisions on a larger scale. If the scientist can’t scientifically and statistically show that the data is reliable, why should the people up the value chain trust it?” If repeat data is incongruous with the established model, it may act as a self-check to ensure that experiments are replicable, potentially offering greater confidence in new results. The statistical framework of DOE is built upon interactions between variables, enabling a more complete exploration of the input space. These benefits have long been recognized by the biological research community; Schrecker notes that “biologists are very good at being statisticians, and as chemists, we need to catch up to them.”

In doing so, chemists in pharmaceutical and biologics research have the opportunity to build an AI IP portfolio that rivals the drugs themselves. Schrecker explains that the goal should not be to try and best the powers that be at creating AI models. Instead, “the competitive advantage is to have the proprietary data set that helps you use AI in the best possible way.” A key argument for interdisciplinary thinking between chemists and statisticians is that “the people who win these competitions in ML are the ones with the most domain knowledge,” Schrecker notes. He continues that this domain knowledge will remain crucial in advancing chemical research, no matter how advanced AI and ML tools become.

Finally, adoption of more rigorous data practices and statistical experimentation provides an opportunity to create new standards. Schrecker warns that while AI is a tool that largely democratizes data science and statistical thinking, “everyone is thinking they can create their own standards.” In lieu of this individual thinking, he argues “it’s going to be very important that these core tools retain wide usage, but we must also retain influence over how we build these standards going forward.” The adoption of unified standards for statistical exploration of chemical space will provide the firm foundation needed to reap the benefits of these modern techniques.

Data-driven thinking in chemistry has the chance to augment each rung of the drug development ladder. If chemists embrace multidimensional thinking by incorporating statistics into their bench workflows, then they can reap the benefits of more meaningful experiments. Simultaneously, leadership can enjoy heightened confidence in their R&D teams’ results, while cementing the adoption of unified standards for the sustainable, replicable use of AI and ML tools. It’s the kind of multidimensional progress that can only be built with data.