3Dprint

3Dprint is a transformer-based neural network that learns molecular descriptors from 3D structural information. The fingerprints are pretrained and then ridge regression is used to regress all observables.water

 Architecture:                                                                                                            

  - Uses MACE-OFF24 (medium) atomic embeddings as input features                                                          

  - Transformer encoder with attention mechanisms aggregates atom-level features into molecular-level descriptors          

  - Outputs fixed-size molecular descriptors                                                                    

 Training approach:                                                                                                      

  1. Self-supervised pretraining: The model was pretrained on a large molecular dataset using denoising/masking objectives  

  2. Descriptor extraction: For downstream tasks, 5 conformers per molecule are generated, and their descriptors are averaged                                                                                                                

  3. Ridge regression: A simple linear model (Ridge regression) is trained on top of the frozen REM3DI descriptors for      

 Performance Comments    

-              Model including graph embeddings show improved performance

-                 Linear models are worse than MLP ontop of descriptor.