Release Highlights 2025.3

Small Molecule Discovery Suite Highlights

Structural Biology Floes

The new Structural Biology Floes are used to refine Cryogenic Electron Microscopy (cryo-EM) results to protein conformer ensembles. Cryo-EM captures information about the biomolecules’ native conformational ensemble. The floes in this package leverage weighted ensemble molecular dynamics (WEMD) to explore the full protein conformer space defined by the cryo-EM density maps, or mean maps and eigenmaps when available. Even low-resolution maps guide exploration via global shape, enabling the generation of diverse conformers consistent with experimental evidence.

Figure 1. A sample initial protein structure (purple) with cryo-EM density maps, which are used for input in the simulation. The mean map is in blue, and the two heterogeneous maps are in tan and green.

Figure 2. The energy landscape after running weighted ensemble molecular dynamics based on two provided eigenmaps. The metastable state was chosen and an example trajectory is shown for the transition from the starting state to that conformer.

Simulations can be analyzed to identify the best conformers which fit into any provided cryo-EM density maps, even those not used to drive the original WEMD simulation. The resulting energy landscapes can be used to identify metastable states and the transition trajectory between them.

These floes are an affordable way to sample large protein motions, such as secondary structure reorientations or domain motions, potentially requiring only a few WEMD iterations. Each additional density map will be screened against the same simulation to find the best fitting structures; up to dozens of maps can be used simultaneously with minimal added cost.

Output from these floes are all-atom protein conformers which are suitable as input for other calculations, including cryptic pocket detection or SiteHopper.

ML Model Building Floes

ML Build: Graph Convolution Model on Pregenerated Features for Small Molecules

Two new floes are introduced in this release to build graph convolutional neural network (GCNN) models and to use those models for predictions. GCNNs are desirable models primarily because they can learn directly from the molecular graph using atom and bond features from OpenEye toolkits rather than features derived from the molecule (e.g., fingerprints), resulting in fewer parameters for the model to learn. GCNNs are also not limited to interaction radii from complex structures, therefore building more generalizable models.

The new floes are ML Build: Graph Convolution Model on Pregenerated Features for Small Molecules and ML Predict: Graph Convolution Model Prediction. To support these new floes, the ML Build and Predict: Optimal Property-Predicting Graph Convolutional Neural Network Model tutorial and the How to Build and Predict Optimal Property-Predicting Graph Convolutional Neural Network Machine Learning Models by Tweaking Neural Network Architecture how-to-guide have been added to the documentation.

Output analysis in these floes now includes information about how the number of parameters in the model compares to the amount of data. Warnings are included if overfitting based on this ratio is a concern.