In the beginning of November, part of the Glowacki group flew over to Denmark to participate to the “machine learning and molecules” conference. The aim of this event was to highlight some important developments in the field of machine learning applied to biology and molecular modelling.
There were two particularly interesting talks on the theme of combining experimental and computational in a unified theoretical framework. Ken Dill who used the the challenge of modelling proteins to talk about MELD – a Bayesian funnel approach to incorporate experimental data through the use of probabilistic restraints into MD simulations. Simon Olsson also talked about incorporating experimental data to correct biases in the computationally derived conformational dynamics of proteins. Both approaches systematize the combination of physical intuition and chemical knowledge to guide and interpret computational experiments.
Klaus-Robert Müller presented work on how in his group they have managed to learn accurate force-fields with machine learning models that guarantee energy conservation.
In addition, he also presented work on how to better understand image classification with neural networks. He described a procedure that enables to visualise the parts of an image that have the largest influence on the decisions of neural networks. In our group, we are currently doing something similar, as we are trying to understand how neural networks ‘learn’ to predict energies from simple descriptors. Therefore, it was useful to hear how other researchers have approached a related problem.