Biophysical Society Thematic Meeting | Hamburg 2022

Biophysics at the Dawn of Exascale Computers

Tuesday Speaker Abstracts

CONSTANT PH MOLECULAR DYNAMICS IN GROMACS USING LAMBDA DYNAMICS AND THE FAST MULTIPOLE METHOD Eliane Briand 1 ; Bartosz Kohnke 1 ; Carsten Kutzner 1 ; Helmut Grubmüller 1 ; 1 Max Planck Institute for Multidisciplinary Sciences, Department of Theoretical and Computational Biophysics, Göttingen, Germany The residue protonation state of biomolecules is usually treated as fixed in molecular dynamics (MD) simulations: this is equivalent to a time-varying pH. Numerous approaches are found in the literature to obtain a more realistic constant pH by dynamically altering protonation, however these tend to be too slow or too complicated for routine use. Building upon the established λ - dynamics method with Hamiltonian interpolation, we aim to make constant pH MD (CPH-MD) accessible to the non-expert by an intuitive interface, a user-oriented documentation, and a performance high enough for use beyond small proteins through FMM electrostatics. To illustrate practical usages of our implementation as well as sketch an accuracy profile, we present titration results for small histidine and glutamate-containing peptides with pKa shifted by their proximate environment, as well as the usual CPH-MD benchmark protein lysozyme. The advent of high repetition-rate XFELs is generating a torrent of data. Will machine learning conquer the deluge? Machine learning, a branch of Artificial Intelligence, perform tasks typically reserved for humans. Most machine-learning tasks involve some kind of “recognition”. Examples include recognizing individuals (facial recognition), obstacles (self-driving vehicles), or patterns (stock-market fluctuations).Recognition tasks are, in essence, labeling exercises. Recognizing a face, for example, involves attaching a name to it. Most machine- learning approaches, such as “Deep Learning”, provide little or no insight into the principles by which the labels are generated. The ability to perform a task does not require understanding the underlying processes. You do not have to understand the workings of the brain to recognize your spouse. Scientific knowledge, in contrast, entails understanding the underlying processes. A deep understanding of facial recognition, for example, must elucidate the structures and processes by which the brain recognizes faces. Traditionally, scientific understanding proceeds by assimilating a few experimental clues into a (mathematically sound) theory. This theory is then buttressed by a succession of carefully designed observations. Such discovery processes are designed to make the best use of limited data. The data deluge is undermining this approach. I will describe how machine learning can help extract scientific understanding from the data deluge. This work was supported by the US Department of Energy, Office of Science, Basic Energy Sciences under award DE-SC0002164 (underlying dynamical techniques), and by the US National Science Foundation under award STC 1231306 (underlying data analytical techniques). WHAT CAN WE LEARN FROM MACHINE LEARNING? Abbas Ourmazd 1 ; 1 University of Wisconsin-Milwaukee, Physics, Milwaukee, WI, USA

10