This model utilizes a Graph Isomorphism Network (GIN) to predict the bioactivity and binding affinity ($K_i$) of small molecules against specific protein targets. By representing molecules as graphs where atoms are nodes and bonds are edges, the model captures complex spatial relationships crucial for pharmacological efficacy.
Model Architecture
The model implements a Message Passing Neural Network (MPNN) using the GIN convolution operator.
Node Features: Includes atomic number, chirality, hybridization, and formal charge.
Edge Features: Includes bond type (single, double, triple, aromatic) and stereochemistry.
Readout Layer: Global Mean Pooling followed by a 3-layer MLP.
Aggregation: The update rule for node $i$ at layer $k$ is defined as:
$$h_i^{(k)} = \text{MLP}^{(k)} \left( (1 + \epsilon^{(k)}) \cdot h_i^{(k-1)} + \sum_{j \in \mathcal{N}(i)} h_j^{(k-1)} \right)$$
Intended Use
Virtual Screening: Ranking massive libraries of compounds to identify potential lead candidates for synthesis.
ADMET Prediction: Estimating the solubility and lipophilicity of new chemical entities.
Target Profiling: Predicting potential off-target interactions to minimize clinical side effects.
Limitations
Stereoisomers: The model may struggle to differentiate between complex enantiomers that have identical connectivity but different biological activity.
Large Molecules: It is primarily validated on small molecules (MW < 800 Da) and may not generalize to biologics or large macrocycles.
Dataset Bias: Prediction accuracy is highly dependent on the chemical diversity of the training set (e.g., ChEMBL or PDBBind).
