Introduction to Graph Neural Networks for Drug Discovery

Graph Neural Networks: Learning Beyond the Grid

An intuition-first introduction into processing relational data, with a focus on Molecular Chemistry.

The Limitation of Traditional Deep Learning

Standard Neural Networks (CNNs, RNNs) expect data with a fixed structure, like grids (images) or sequences (text). They struggle when data has arbitrary size and complex topological connections.

The Real World is Relational: Molecular Example

Running Case Study: Drug Discovery. Molecules are natural graphs.

Nodes = Atoms (Carbon, Oxygen, Nitrogen).

Edges = Chemical Bonds (Single, Double).

Problem: Permutation Invariance

Using a standard matrix (like an image) fails because graph nodes have no fixed order. If you label atoms differently (Atom 1 vs Atom 2), the matrix changes, but the molecule remains identical. A robust model must be 'permutation invariant'.

The GNN Solution: Neural Message Passing

Instead of processing the whole graph at once, we update each node based on its neighbors. This is local, scalable, and order-independent.

1. MESSAGE: Neighbors send their feature information.

2. AGGREGATE: The node sums or averages these messages.

3. UPDATE: The node updates its own state using a Neural Network.

Under the Hood: The Aggregation Function

The state of node 'n' at time 't+1' depends on its previous state and the states of neighbors (ne[n]).

AGG: Sum, Mean, or Max pooling (permutation invariant).

W: Learnable weight matrix (shared across all nodes).

Scarselli et al. originally described this as an information diffusion mechanism to reach a stable state.

From Local Nodes to Global Prediction

Once message passing is done, every atom (node) has a vector representing its local chemical environment.

To classify the whole molecule (e.g., 'Is it Mutagenic?'), we use a READOUT function (sum or average of all node vectors) to get a single graph embedding.

Evolving Architectures

GCN (Graph Convolutional Network)

Uses a normalized weighted sum of neighbors. Simple and effective baseline.

GraphSAGE

Inductive learning. Samples a fixed number of neighbors (instead of all). Crucial for massive graphs (social networks) where nodes have thousands of connections.

GAT (Graph Attention Network)

Learns to weigh neighbors differently. A 'double bond' might be more important than a 'single bond'.

GNN Performance on Mutagenesis

Non-linear GNNs outperform other methods on graph-structured data sets (Mutagenesis). They successfully utilize the topological connections (bonds) that linear methods and standard feature-vector NNs struggle to capture fully.

Impact & Conclusion

GNNs have moved from theoretical curiosity to the backbone of modern AI in science.

Drug Discovery: predicting properties of molecules before synthesizing them.

Social Networks: Recommender systems (Pinterest/UberEats) use GraphSAGE.

Maps & Traffic: Predicting arrival times in Google Maps (ETA).