Welcome to mythos’s documentation!
mythos is a Python library for differentiable simulation and optimization of coarse-grained molecular models. It enables fitting force field parameters to macroscopic experimental data using gradient-based optimization.
Core Features
Multiple simulation backends — run simulations with JAX-MD, oxDNA, GROMACS, or LAMMPS from a unified interface. Simple API for simulators makes extensions easy. See Simulators.
Differentiable optimization — compute gradients via JAX automatic differentiation (JAX-MD; https://github.com/jax-md/jax-md) or Boltzmann reweighting with DiffTRe (for non-differentiable simulators, such as oxDNA, GROMACS, and LAMMPS; see https://www.nature.com/articles/s41467-021-27241-4). See Optimization.
DNA, RNA, and lipid energy models — built-in support for oxDNA1, oxDNA2, oxRNA, oxNA (hybrid DNA/RNA), and MARTINI coarse-grained models, with a clear extension API for implementing custom models. See Energy Functions.
Parallel optimization with Ray — run multiple simulators and objectives in parallel across heterogeneous hardware, enabling hyperscale optimization campaigns. See Optimization.
Rich observable library — The extensible Observable API Observables turns simulation trajectories into values for loss computations. The library includes a wide variety of observable implementations for different types of simulations, including structural observables (helical rise, pitch, diameter, persistence length, melting temperature), membrane properties (thickness, area per lipid), and distance metrics (Wasserstein).
Optimization Lifecycle
Optimizations in mythos are organized as follows:
Simulators run simulations (typically molecular dynamics or Monte Carlo) and expose Observables (trajectories and summary statistics). Objectives consume observables and compute gradients of a loss function. The Optimizer orchestrates the loop — running simulations, collecting observables, computing gradients, and updating parameters.
For a full description, see Optimization.