Type Name Latest commit message Commit time
Failed to load latest commit information. reversion and mpnn-s Jul 11, 2017 lowering indentation level #1232 Sep 21, 2018

Message Passing Neural Networks

MPNNs aim to generalize molecular machine learning models that operate on graph-valued inputs. Graph-Convolutions [] and Weaves
[] (among others) can be recast into this framework []

The premise is that the featurization of arbitrary chemical multigraphs can be broken down into a message function, vertex-update function, and a readout functi
on that is invariant to graph isomorphisms. All functions must be subdifferentiable to preserve gradient-flow and ideally are learnable as well

Models of this style introduce an additional parameter T, which is the number of iterations for the message-passing stage. Values greater than 4 don't seem
to improve performance.

##MPNN-S Variant MPNNs do provide a nice mathematical framework that can capture modern molecular machine learning algorithms we work with today. One criticism of this algorithm class is that training is slow, due to the sheer number of training iterations required for convergence - at batch size 20 on QM9, the MPNN authors trained for 540 epochs.

This can be improved significantly by using batch normalization, or more interestingly, the new SELU activation []. In order to use SELUs straight through the system, we dropped the GRU unit [] the authors used in favor of a SELU activated fully-connected neural network for each time step T. This modified approach now achieves peak performance in as little as 60 epochs on most molecular machine learning datasets.

MPNN-S sets new records on the Delaney & PPB datasets:

Dataset Num Examples MP-DNN Val R2 [Scaffold Split] GraphConv Val R2 [Scaffold Split]
Delaney 1102 .820 .606
PPB 1600 .427 .381
Clearance 838 .32 .28

Run Code

$ python