Machine learning applications towards automatized analysis and interpretation of HS-AFM imaging data

Holger Flechsig¹, Romain Amyot¹

¹WPI Nano Life Science Institute, Kanazawa University, Japan

While high-speed atomic force microscopy (HS-AFM) experiments play a leading role to observe functional dynamics of proteins, the interpretation of observations is generally difficult, mainly for two reasons: HS-AFM images have no atomistic resolution, and, automatized analysis of large imaging data sets acquired from experiments is practically absent. Both limitations can be overcome only by the methods provided by computational science. Our previous work including simulation AFM, development of fitting methods, and integrative modelling has significantly improved interpretation of HS-AFM observation. However, large-scale automatized analysis to fully exploit the explanatory power of HS-AFM observations requires application of machine learning methods. In this situation, synthetic AFM data computed from the enormous amount of high-resolution structural data of proteins (available from the Protein Data Bank and AlphaFold predictions) and atomistic modelling of functional dynamics is essential to infer atomistic information from resolution-limited imaging. In this talk we present our recently developed algorithms for GPU-accelerated calculation of simulated AFM imaging, which allows to efficiently generate a massive database of simulated AFM topographic images which can be correlated with measured HS-AFM images. We then present our attempts to employ synthetic AFM data of proteins in machine learning methods and demonstrate first applications to experimental HS-AFM data.