Stardust: Modular Coordinate Refinement Library

Stardust is a modular library supporting coordinate refinement against experimental data: cryo-EM maps, crystallographic structure factors, and beyond.

Stardust is based on pytorch.

Structure and Vision

Stardust implements two primary abstractions:

1. losslab. These are likelihood functions that compute the probability of some structure given a set of experimental data: p(x|D). A common interface to these losses is enforced by an abstract base class, BaseLoss.

2. The refinementlogger, a gradient decent manager and logger. Many of the outputs of refinement are common to all refinement strategies: structures as a function of iteration, compute metrics, etc. The RefinementEngine class implements these common features and provides a foundation which specific refinement implementations can extend.

3. A structure module that helps manage topology, coordinate, B-factor, and occupancy information. It contains a powerful Structure object in its own right, as well as code that is crucial for converting and interoperating with different coordinate representations in use.

Out of scope

Stardust does not generate or sample structures/coordinates. Stardust simply provides a likelihood (and, via torch, likelihood gradients) and a generic system for tracking progress as one seeks to optimize that likelihood.

Stardust assumes your are working with a discrete list of cartesian coordinates that represent atomic positions. Models that use densities, continous distributions, etc. are out of scope.