whiteh4t/esm2-8m-protein-localization

ESM-2 8M — Protein Subcellular Localization (Linear Probe)

Fine-tuned ESM-2 8M for protein subcellular localization (10 classes). All ESM-2 layers frozen — only the classification head was trained.

Results

Metric	Score
Accuracy	69.6%
F1 (macro)	0.581
F1 (weighted)	0.686
MCC	0.614

Architecture

ESM-2 8M (frozen) → Mean pooling → LayerNorm → Linear(320→80) → GELU → Linear(80→10)

Training

• Dataset: DeepLoc 2.0 (17,266 train / 3,700 val / 3,701 test)
• Strategy: Linear probe (backbone frozen, head only)
• Epochs: 15
• Learning rate: 1e-3
• Hardware: NVIDIA DGX Spark

All Models in This Project

Model	Strategy	Accuracy	Link
ESM-2 8M	Linear probe	69.6%	This repo
ESM-2 35M	Full fine-tune	74.3%	whiteh4t/esm2-35m-protein-localization
ESM-2 150M	Full fine-tune	76.6%	whiteh4t/esm2-150m-protein-localization
ESM-2 650M	LoRA (r=16)	76.5%	whiteh4t/esm2-650m-protein-localization-lora

Links

• Demo: HuggingFace Space
• Code: GitHub