Statistical Machine Learning for Combinatorial, Geometric, and Discrete Structures

Data in the life sciences (such as genomes, proteins, gene networks, and molecular structures) are not just numerical but come with complex underlying structures. Our research focuses on designing machine learning methods, a form of statistical prediction, tailored to these structured data. The unique challenges of understanding life's complexity demand more than simply applying existing methods. Instead, we aim to redefine and rebuild machine learning and machine discovery from the ground up as an information science. We believe that only through this rigorous effort can we bring the kind of practical breakthroughs seen in image, speech, video, and natural language processing into the realm of science.

Machine Discovery and Machine Understanding: The Science of Scientific Methods

AI for Science is, at its core, the science of how we do science, i.e., a field inherently meta-scientific in nature. Recent advances in information science compel us to confront a fundamental question: What does it mean to truly understand something scientifically?

The unique nature of machine learning (or statistical prediction from data in general) has sharpened this question. Modern technologies like image recognition, speech processing, and natural language processing have reached practical, even transformative, levels thanks to large-scale data and machine learning. These systems are designed by humans, and their inner workings are known to us in full technical detail. Yet, despite this complete transparency, these technologies offer us surprisingly little in terms of helping us understand how we, as humans, perceive images and sounds, or how we acquire and use language.

This paradox challenges the constructivist assumption that we can understand a phenomenon if we can build a model that replicates it. Our lab sees this as a crucial opportunity: by using life itself (an extraordinarily complex real-world phenomenon) as our testbed, we can turn this abstract philosophical problem into a concrete, technical one. We seek to develop new tools that bring us closer to true scientific discovery and scientific understanding, from a broad and integrative perspective.

From Life as Chemistry to Life as Information

We believe that bioinformatics stands on two pillars: using information science to advance life science, and advancing information science for life science. Our lab focuses on the latter, as an information science laboratory.

Just as a pile of stones isn’t a house, simply gathering all the atoms that make up life doesn’t result in life. The difference between living and non-living things doesn't lie in the materials themselves. In fact, the atoms in living organisms are constantly being replaced through metabolism. In the continuous flow of time, these atoms are organized hierarchically into molecules, tissues, and organisms, and these structures interact in complex ways through chemical reactions, creating non-equilibrium dynamics. As a result, thousands of chemical reactions occur continuously and in a highly coordinated manner, even at room temperature, enabling both material and energy metabolism. In this way, life maintains a high level of order, defying the natural tendency toward increasing entropy.

Our core interest lies in understanding molecular interactions and non-equilibrium dynamics. By focusing on diverse domains (metabolism, transcription, catalysis, microscopy, chemical reactions, quantum chemistry, mass spectrometry, and molecular structures), we aim to rigorously test and demonstrate the power of the new tools we develop. These efforts, grounded in information science, are our contribution to understanding life not just as chemistry, but as information.