Changing Your Martini Can Still Give You a Hangover.

The Martini 3.0 coarse-grained force field, which was parametrized to better capture transferability in top-down coarse-grained models, is analyzed to assess its accuracy in representing thermodynamic and structural properties with respect to the underlying atomistic representation of the system. These results are compared to those obtained following the principles of statistical mechanics that start from the same underlying atomistic system. To this end, the potentials of mean force for lateral association in Martini 3.0 binary lipid bilayers are decomposed into their entropic and enthalpic components and compared to those of corresponding atomistic bilayers that have been projected onto equivalent coarse-grained mappings but evolved under the fully atomistic forces. This is accomplished by applying the reversible work theorem to lateral pair correlation functions between coarse-grained lipid beads taken at a range of different temperatures. The entropy-enthalpy decompositions provide a metric by which the underlying statistical mechanical properties of Martini can be investigated. Overall, Martini 3.0 is found to fail to properly partition entropy and enthalpy for the PMFs compared to the mapped all-atom results, despite changes made to the force field from the Martini 2.0 version. This outcome points to the fact that the development of more accurate top-down coarse-grained models such as Martini will likely necessitate temperature-dependent terms in the corresponding CG force-field; although necessary, this may not be sufficient to improve Martini. In addition to the entropy-enthalpy decompositions, Martini 3.0 produces an incorrect undulation spectrum, in particular at intermediate length scales of biophysical pertinence.

Coarse-Graining with Equivariant Neural Networks: A Path Toward Accurate and Data-Efficient Models.

Machine learning has recently entered into the mainstream of coarse-grained (CG) molecular modeling and simulation. While a variety of methods for incorporating deep learning into these models exist, many of them involve training neural networks to act directly as the CG force field. This has several benefits of which the most significant is accuracy. Neural networks can inherently incorporate multibody effects during the calculation of CG forces, and a well-trained neural network force field outperforms pairwise basis sets generated from essentially any methodology. However, this comes at a significant cost. First, these models are typically slower than pairwise force fields, even when accounting for specialized hardware, which accelerates the training and integration of such networks. The second and the focus of this paper is the need for a considerable amount of data to train such force fields. It is common to use 10s of microseconds of molecular dynamics data to train a single CG model, which approaches the point of eliminating the CG model's usefulness in the first place. As we investigate in this work, this "data-hunger" trap from neural networks for predicting molecular energies and forces can be remediated in part by incorporating equivariant convolutional operations. We demonstrate that, for CG water, networks that incorporate equivariant convolutional operations can produce functional models using data sets as small as a single frame of reference data, while networks without these operations cannot.