Skip to content

🚀 Use Cases

DiffBiophys enables a wide range of applications by bridging the gap between structural models and experimental solution-state data through differentiability.

1. Experimental Structure Refinement

Traditional structure refinement often relies on stochastic sampling (e.g., Monte Carlo or Simulated Annealing). With DiffBiophys, you can use gradient-based optimization to refine a structure directly against experimental data.

  • Example: Start with a homology model and refine the backbone \(\phi/\psi\) angles to minimize the difference between calculated and experimental Residual Dipolar Couplings (RDCs).
  • Key Advantage: Faster convergence and the ability to handle high-dimensional parameter spaces.

2. Physics-Informed Machine Learning

Integrate biophysical "truth" into your AI models. DiffBiophys kernels can be used as differentiable loss functions in deep learning pipelines (e.g., training a model to predict protein conformations).

  • Example: Use the debye_saxs kernel as a loss term in a Variational Autoencoder (VAE) to ensure the latent space represents physically plausible, compact structures that match experimental scattering curves.
  • Key Advantage: Forces the model to learn physics-consistent representations without requiring massive labeled datasets.

3. Ensemble Weight Optimization

Experimental data like SAXS often represent an average over a conformational ensemble. DiffBiophys allows you to differentiate with respect to ensemble weights.

  • Example: Given a library of 100 possible protein conformations, optimize the population weights \(\{w_i\}\) such that the ensemble-averaged SAXS curve \(\sum w_i I_i(q)\) matches experimental data.
  • Key Advantage: Native JAX vmap support makes calculating intensities for large ensembles extremely efficient on GPUs.

4. Differentiable Molecular Dynamics Analysis

Use biophysical gradients to steer or analyze MD trajectories.

  • Example: Calculate the "experimental force" acting on a structure by taking the gradient of an experimental misfit (e.g., \(\nabla_\text{coords} \|I_\text{calc} - I_\text{exp}\|^2\)).
  • Key Advantage: Directly couples structural dynamics to solution-state observables.

5. Automated Tensor Fitting

In NMR, the alignment tensor (Saupe tensor) is often an unknown parameter. DiffBiophys provides built-in SVD fitting that is fully differentiable.

  • Example: Simultaneously solve for the optimal alignment tensor and the optimal structure by alternating between tensor fitting and coordinate optimization.
  • Key Advantage: Eliminates the need for manual, iterative tensor estimation.