General Framework that accounts graph and edges embedding into message passing layers

[aditya0by0]

Description:
In Section 13.3.3 of the Deep Learning book ([link](https://issuu.com/cmb321/docs/deep_learning_ebook/438)) and in the corresponding paper [arXiv:1806.01261](https://arxiv.org/pdf/1806.01261), the authors incorporate both graph-level and edge-level embeddings directly into the message-passing layers, updating them at each layer. This differs from the current PyG implementations of GAT and ResGated, which use edge embeddings only to enrich node representations without updating them across layers.

By contrast, in our current approach, RDKit2DNormalized molecular features are employed only in the final linear layers. It may therefore be worthwhile to investigate whether integrating and updating global graph-level features (e.g., RDKit2DNormalized) and edge features directly within the message-passing phase could improve model expressiveness and performance.

[aditya0by0]

• If this give performance improvement over the current approach of handling molecule-level properties, then it will reduce the number of linear layer parameters (as mentioned in Linear layers constitutes almost half of the trainable parameters #17 ) too.

[aditya0by0]

The edges features can be updated and incorporated into message passing layers for ResGated and GAT as per the GatedGCN-E and GAT-E version specified in https://arxiv.org/pdf/2003.00982