Deep learning model-based inference for molecular simulations offers a great speedup (orders of magnitude) when compared to reference quantum chemical methods such as density functional theory (DFT), along with evidence of increased accuracy compared to classical force field calculations on some systems. We demonstrate an entire atomistic simulation pipeline designed for protein systems to exploit the benefits of such models. The application of the MACE model architecture is combined with a physics-informed loss function inspired by PhysNet to improve the representation of molecular physics and account for long-range interactions explicitly. The model is trained on PhysNet's solvated fragments dataset. Our pipeline enables stable GPU-accelerated molecular dynamics (MD) simulations of small molecules within the same size as the molecules present in the dataset as well as generalisation towards larger peptides such as chignolin (175 atoms). Forces along the MD trajectories are assessed by comparison to a DFT reference. Furthermore, we present stable and accurate energy minimisations for a selection of six test molecules. Based on our results, we provide a discussion of the strengths and limitations of the approach including an outlook towards future improvements.