whiteh4t/esm2-150m-protein-localization

ESM-2 150M — Protein Subcellular Localization (Full Fine-Tune)

Fine-tuned ESM-2 150M for protein subcellular localization (10 classes). All parameters trained end-to-end. Best accuracy among full fine-tune models.

Results

Metric	Score
Accuracy	76.6%
F1 (macro)	0.696
F1 (weighted)	0.761
MCC	0.706

Architecture

ESM-2 150M (all trainable) → Mean pooling → LayerNorm → Linear(640→160) → GELU → Linear(160→10)

Training

• Dataset: DeepLoc 2.0 (17,266 train / 3,700 val / 3,701 test)
• Strategy: Full fine-tune (all 150M parameters)
• Epochs: 10
• Learning rate: 1e-5
• Hardware: NVIDIA DGX Spark

All Models in This Project

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

Links

• Demo: HuggingFace Space
• Code: GitHub