Statistical Analysis

n=525 TEST PROTEINS

95% confidence intervals via normal approximation (closed-form) and in-browser bootstrap resampling (B=5000). Real results from test_results.json — n=525 held-out proteins.

📐

0.3212

Fmax

CI [0.2813, 0.3611]

📈

0.8568

AUROC

CI [0.8268, 0.8868]

📉

0.2178

AUPRC

CI [0.1825, 0.2531]

🧬

525

Test Set

MMseqs2 30% split

Confidence Interval Methods

Normal: CI = p̂ ± z·√(p̂(1−p̂)/n) z₀.₉₅=1.96, z₀.₉₉=2.576

Bootstrap: resample n=525 proteins B=5000× → percentile method [2.5th, 97.5th] of bootstrap distribution

Bootstrap samples (B)

Runs in-browser (JavaScript) — no server call needed

Metrics with 95% CI (Normal approx.)

Loading chart…

Bootstrap Distribution — Fmax

Click "Run Bootstrap" to see the sampling distribution of Fmax.B resampling iterations → percentile CI

Confidence Interval Comparison Table

Metric	Point Est.	Normal 95% CI	Bootstrap 95% CI	Bootstrap SE	Width (normal)
Fmax	0.3212	[0.2813, 0.3611]	—	—	0.0799
AUROC	0.8568	[0.8268, 0.8868]	—	—	0.0599
AUPRC	0.2178	[0.1825, 0.2531]	—	—	0.0706
molecular_function	0.2101	[0.1753, 0.2449]	—	—	0.0697
biological_process	0.1933	[0.1595, 0.2271]	—	—	0.0676
cellular_component	0.3371	[0.2967, 0.3775]	—	—	0.0809

Statistical Tests — Design & Status

Test	Purpose	Implementation	Status
Normal approx. CI	Closed-form 95/99% CI on all metrics	p̂ ± z·√(p̂(1-p̂)/n)	✓ Live
Percentile bootstrap CI	Exact bootstrap CI via B resamplings	In-browser JS, B=1k–10k	✓ Live (click above)
McNemar's test	Per-protein correct/incorrect vs baseline	Requires baseline predictions	Needs baseline run
Wilcoxon signed-rank	Non-parametric per-protein Fmax comparison	Requires baseline predictions	Needs baseline run
ECE (calibration)	Expected Calibration Error for MC Dropout probs	Binning MC probs vs accuracy	Design ready
GO term frequency χ²	Are common GO terms predicted more accurately?	χ² test on freq × accuracy bins	Design ready