ML at Kebotix
The process of discovering a new molecule in the material science field is not so different from trying on a million pair of shoes, only to find out that sandals would have been more appropriate. Whether searching for compounds to combat pollution or infection, material scientists spend much of their time analyzing new molecules with little promise. In an effort to speed up the process, MIT startup Kebotix has developed a “self-driving lab” that merges machine learning (ML) and robotics to rapidly prototype new materials1.
Although robots have been a part of large-scale chemistry labs for some time now, automated analysis has only recently become a reality due to the novel application of ML algorithms. Scientists at Kebotix have designed ML programs that work in tandem with their robotic lab partners, using feedback from the output of a previous experiment to modify the input of the next experiment. Essentially, the robotic arms can conduct tests unaided and tweak chemical parameters by “thinking” carefully about the results.
While the self-driving lab creates significant value for human chemists who would otherwise be analyzing test results most of the day, the management team at Kebotix must consider the trade-offs associated with automated material discovery. For example, Kebotix scientists have created a layer in their neural network “to weed out designs that stray too far from the original.”1 Though the layer keeps the range of possibilities sufficiently narrow for ease of iteration, it prevents the robot from taking creative, divergent paths to a discovery. To be clear, the tunnel vision of the robot-program pair may not limit its abilities at all but simply define where humans need to intervene in the discovery process.
Describing Kebotix as “the materials company of the 21st century,” CEO Jill S. Becker has set ambitious goals for the company’s growth over the next several years2. Though still in the early stages of VC funding, Kebotix has already secured a $5 million investment from One Way Ventures and counts Baidu, a leader in ML, among its investors. Interestingly, Kebotix has been operating in “stealth mode”—a reference to exceptionally disruptive projects conducted in secret within large corporations3. Based on its management structure of mostly scientists and engineers, Kebotix is still primarily focused on product development with less emphasis on commercialization in the short-term4.
Short-term and Medium-term Recommendation
In the short-term, Kebotix is developing self-driven electronics a proof of concept for its robot-ML pairing1. The electronics industry is an ideal target due to the lack of regulatory hurdles that are present in pharmaceuticals and other discovery-intensive industries. When considering medium-term expansion into other classes of materials, Kebotix will undoubtedly have to understand whether the robot-ML pairing provides a unique competitive advantage. In the case of pharmaceuticals, the discovery process is fairly advanced and could be cost-prohibitive for the Kebotix team1. However, one area in which the company can provide a competitive advantage is coupling the pre-existing ML algorithms with molecular simulations developed in academic settings.
The area of simulation presents a powerful opportunity for Kebotix, given that the materials industry is founded upon many years of chemistry advancements. With the robot-ML pair knowing almost nothing about chemistry from the outset, the inclusion of simulation results in training data could vastly accelerate and improve the materials discovery process. MIT professor Alán Aspuru-Guzik, the Chief Visionary Officer of Kebotix, has stated in his research that virtual libraries contain hundreds of millions of candidates, but when used alone, they are limited by the search strategy employed to explore the range of chemical options5. Seeing that simulation and ML can experience the same setbacks, namely the limitations to new molecular discoveries, it is not immediately clear that the two technologies would be complementary. However, the rich data proffered by simulations would help scientists at Kebotix refine their current learning algorithms. In other words, the copious amounts of simulation data could reduce the learning time for Kebotix’s algorithms and potentially make the robot-ML technology competitive with human-led molecular design in other industries relying upon chemical discovery processes.
- What other trade-offs should Kebotix consider when fine-tuning its ML algorithms? Or does the feedback loop completely remove human bias?
- What is the role of the scientist/principal investigator in the new robot-ML model of material discovery?
(Word Count: 715)
1Will Knight, “A robot scientist will dream up new materials to advance computing and fight pollution” MIT Technology Review, November 7 2018, https://www.technologyreview.com/s/612388/a-robot-scientist-will-dream-up-new-materials-to-advance-computing-and-fight-pollution/
2“Harvard Scientists Launch Breakthrough AI and Robotics Tech Company, Kebotix, for Rapid Innovation of Materials”, BusinessWire, November 7 2018, https://www.businesswire.com/news/home/20181107005913/en/Harvard-Scientists-Launch-Breakthrough-AI-Robotics-Tech
3Paddy Miller and Thomas Wedell-Wedellsborg, “The Case for Stealth Innovation”, Harvard Business Review, March 2013 https://hbr.org/2013/03/the-case-for-stealth-innovation
4Kebotix.com website, 2018, https://www.kebotix.com/#team
5Automatic Chemical Design Using a Data-Driven Continuous Representation of Molecules
Rafael Gómez-Bombarelli, Jennifer N. Wei, David Duvenaud, José Miguel Hernández-Lobato, Benjamín Sánchez-Lengeling, Dennis Sheberla, Jorge Aguilera-Iparraguirre, Timothy D. Hirzel, Ryan P. Adams, and Alán Aspuru-Guzik
ACS Central Science 2018 4 (2), 268-276