Polypharmacy Side Effects Eval

polypharmacy-side-effects-eval

Predicting Drug-Drug Interactions using Deep Generative Models on Graphs — Ngo et al. (2022) (arXiv:2209.09941, 2022)

What this evaluates

Evaluates a model's ability to predict polypharmacy side effects (drug-drug interactions) in a multimodal biomedical graph. It probes the model's capacity to learn continuous latent representations for drugs and proteins and generalize to unseen drug pairs across 964 specific side effect types.

Datasets

• Polypharmacy Side Effects Dataset — total ?; splits: train (-1), val (-1), test (-1); repo https://github.com/HySonLab/drug-interactions

Metrics

• cross-entropy loss (primary) — range: [0, inf)
  • Cross-entropy loss over positive and negative sampled edges: -log p_e(v_i, v_j) for positive edges and -log p_e(v_i, v_n) for negative edges.

Input / output format

Input: Multimodal graph with drug and protein nodes. Node features are one-hot representations of node indices. Edges represent drug-drug, drug-protein, and protein-protein interactions.

Output: Probability distribution over edge types (or binary existence probability) for a given drug-drug pair.

Scoring recipe

for edge in test_edges:
    pred_prob = model.predict(edge.u, edge.v)
    true_label = 1 if edge exists else 0
    loss = -log(pred_prob) if true_label else -log(1 - pred_prob)
return mean(loss)

Common pitfalls

• Training edges are split 80/20 for message passing vs supervision, deviating from standard transductive link prediction.
• Negative sampling is type-specific (random node under specific edge type), not uniform across the graph.

Evidence (verbatim from paper)

The first term of L denotes the cross-entropy loss of the probabilities of positive and negative edges which is sampled by choosing a random node vn for each node vi under a specific edge type e

Citation

@misc{ngo2022predicting,
  title={Predicting Drug-Drug Interactions using Deep Generative Models on Graphs},
  author={Ngo et al. (2022)},
  year={2022},
  note={arXiv:2209.09941}
}

• arXiv: 2209.09941