This work was done in collaboration with Tobias Höppe, Emmanouil Angelis, Mathias Schreiner, Stefan Bauer, @andreadittadi.bsky.social and @smnlssn.bsky.social‬. We gratefully acknowledge funding from Chalmers, TUM, @www.helmholtz-munich.de, @mit.edu, WASP, and Fulbright

We’d love feedback and questions, especially from NMR/FRET/EM experimentalists, developers of generative ML for molecules, or MD folks working on conformational sampling and ensemble refinement.

TL;DR:
MEW guidance offers a principled way to adjust generated distributions to match experimental data and focus sampling on specific subspaces without the overhead of retraining or heavy data demands.
Check out the full paper here:
arxiv.org/abs/2505.13375

MEW guidance differs from other guidance methods in that no reinforcement learning is involved (no rewards or policy learning). It is related to Schrödinger bridges but specifically optimized for experimental data and is grounded in thermodynamics and information theory.

Key results from our experiments:
- Accurately match experimental constraints (shown on toy models and a chignolin ensemble generator)
- Ensembles remain physically realistic.
- Increase in diversity of guided samples
- Works robustly, even if constraints are noisy or limited.

In our paper, we introduce two specific ways to use MEW guidance:
a) Observable Guidance: adjusts the molecular ensemble to match experimental observables, like NOEs, RDCs, or ΔG
b) Path Guidance: enhance sampling in specific regions of conformational space (e.g., transition states)

Why MEW guidance? It’s physically intuitive, optimal from a control/information theory perspective, works well in sparse data settings common in science, and is computationally efficient as well as numerically stable.

Similarly, some regions in the conformational space, like transition regions between states, are severely undersampled.
MEW guidance helps by gently steering the generated distribution to match new constraints or to concentrate sampling in a region of interest.

Most generative models for molecular ensembles sample from a learned prior.
But due to bias in the training data, ensemble statistics often don’t match experimental results.

The core idea is to encode additional data, like equilibrium expectations or a set of reference samples, by modifying the score of the generative ODE/SDE. MEW guidance ensures we stay as close as possible to the original learned dynamics, i.e., minimizing the excess work of the perturbation.