PaTz โ€” Physics-Aware Transformer with eXplainability
โ— API Offlinepatz_env

Performance Metrics

TEST SET ยท n=525

Held-out test set. MMseqs2 30% similarity split prevents sequence-based data leakage.

max_ฯ„ 2PR/(P+R)
โ€”
Fmax (CAFA)

Primary CAFA metric. Threshold-independent. Macro-averaged over all test proteins.

โˆซโ‚€ยน TPR d(FPR)
โ€”
AUROC

Area Under Receiver Operating Characteristic. Measures discriminative ability across all thresholds.

โˆซโ‚€ยน P d(R)
โ€”
AUPRC

Area Under Precision-Recall Curve. Preferred for class-imbalanced GO annotation.

Per-Namespace F1
Metric Interpretation
MetricValueContext
Fmax=0.3212CompetitiveCAFA3 median for pathogen proteins is ~0.30โ€“0.35. Multilabel GO prediction is inherently challenging.
AUROC=0.857StrongAUROC > 0.85 indicates the model discriminates positive/negative GO associations well at all thresholds.
AUPRC=0.218ExpectedAUPRC lower than AUROC is expected with class imbalance โ€” most proteins do not have most GO terms.