Model Comparison

PARTIAL — BASELINES NOT RUN

PaTz vs baseline models (ESM-2 linear probe, sequence MLP). Baseline experiments have not been executed yet — only PaTz results are available from the held-out test set.

Note: Baseline model results are not yet available. The comparison table shows PaTz real results and baseline architecture descriptions only. Running baselines requires implementing and training ESM-2 linear probe and MLP models using the same train/val/test split (see data/processed/split_manifest.json).

📐

0.3212

PaTz Fmax

CAFA evaluation protocol

📈

0.8568

PaTz AUROC

held-out test (n=525)

⚖

Not run

Baselines

need separate training

Model Architecture Comparison

Model	Architecture	Fmax	AUROC	AUPRC	Status
ESM-2 Linear Probe	Frozen ESM-2 650M + mean-pool → single linear layer (256→354). No structure, no physics.	—	—	—	NOT RUN
Sequence MLP	ESM-2 mean-pool → 3-layer MLP (1280→512→256→354). No pairwise reasoning.	—	—	—	NOT RUN
PaTz (ours)	Pairformer (4L×8h) + Physics loss + Score diffusion + MC Dropout. Full model.	0.3212	0.8568	0.2178	✓ DONE

Fmax Comparison (PaTz only available)

Loading chart…

AUROC Comparison (PaTz only available)

Loading chart…

PaTz vs Baselines — Component Contributions

↑ Over Linear Probe

Pairformer adds triangle-update residue–residue reasoning. Physics loss adds differentiable structural constraints. Expected Fmax improvement: +0.05–0.12 (to be measured).

↑ Over Sequence MLP

Pairwise z_ij track captures co-evolutionary contacts impossible in independent residue processing. Score diffusion provides structural conditioning. Expected improvement: +0.03–0.08.

→ Ablation (separate page)

Individual component contributions (physics loss, diffusion, Pairformer depth) are measured on the ablation study page — each component removed one at a time.