← Lectures·L7

AI for 'Omics' Data

Blumenthal · 1 concepts · 7 questions

Concept 1 / 1

Key Concepts to Memorize

What is a graph (network)?


  • Mathematical object G = (V, E): V = nodes, E ⊆ V×V = edges
  • Attributes can be stored at both nodes and edges
  • Excellent for data where information is encoded in connections

Biomedical graph examples:


  • Protein-Protein Interaction (PPI) networks — nodes = proteins, edges = physical binding
  • Gene regulatory networks
  • Disease-disease networks

BIONETS Lab — 3 research areas:


1. Algorithmic network medicine — mine omics data for molecular disease mechanisms

2. Robust and reproducible AI — algorithms/standards for robust biomedical AI

3. Privacy-preserving federated biomedical AI — decentralized AI for cross-institutional studies on sensitive data


The core problem with PPI prediction benchmarks:


  • Many published deep learning results are inflated due to data leakage
  • Naive random data splitting introduces 2 types of "shortcuts":

- Shortcut 1 (topology-based): model memorizes which proteins appear only in positive pairs in training → predicts "interaction" for all their test pairs regardless of partner

- Shortcut 2 (sequence similarity): proteins in training and test sets are too similar → model exploits similarity instead of learning true interaction rules


Data leakage (definition from Blumenthal):


> "Performance scores for an ML-based predictor f are inflated due to data leakage if f uses illegitimate information to achieve the reported scores, that is, information in the training data that generalizes to the test data but not to the inference-time data."


Two intended use cases → different evaluation protocols needed:

Use caseAppropriate splitResult
Proteins WITH well-characterized homologs (in-distribution)Random splitting is fineDeep learning works
Proteins WITHOUT well-characterized homologs (out-of-distribution)Hard split requiredNo tested DL model outperforms random guessing

Key insight: High accuracy scores in the literature were often due to shortcuts, not genuine learning. When hard splits are used, models drop to ~random (0.50) performance.


Solution (DataSAIL): Algorithm for leakage-reduced data splitting + guidelines to avoid data leakage in biological ML.