hazemessam/scl

FLIP SCL Dataset: SubCellular Localization SCL = SubCellular Localization What It Is This is a multi-label classification task where the goal is to predict which cellular compartment(s) a protein resides in within eukaryotic cells. The 10 Subcellular Locations

Note: DROP any example that has split=nan when training. The reason for leaving them is to keep this dataset identical to the original one.

Based on the DeepLoc dataset that FLIP appears to be derived from:

Nucleus Cytoplasm Extracellular (secreted proteins) Mitochondrion Cell membrane Endoplasmic reticulum (ER) Chloroplast (plants only) Golgi apparatus Lysosome/Vacuole Peroxisome

Why Multi-Label? Some proteins localize to multiple compartments (e.g., shuttling between nucleus and cytoplasm, or dual-targeted to mitochondria and chloroplasts). This is biologically important - around 20-30% of proteins have multiple localizations. The Splits Explained Looking at the directory structure, FLIP provides these splits:

1. balanced.csv Standard split with balanced representation across localization categories Avoids class imbalance issues Likely uses random or cluster-based splitting

2. human_hard.csv / human_soft.csv Human-only proteins (Homo sapiens) Hard split: More challenging - possibly using lower sequence identity cutoff between train/test, or testing on rare/underrepresented localizations Soft split: Easier - higher sequence similarity allowed between train/test sets

3. mixed_hard.csv / mixed_soft.csv Cross-species proteins (eukaryotes: yeast, plants, animals, fungi) Hard split: Tests generalization across phylogenetically distant organisms or rare localization patterns Soft split: Tests on more similar proteins/organisms

What Makes Splits "Hard" vs "Soft"? Based on subcellular localization literature:

Soft splits likely have: Higher sequence identity between train/test (e.g., 30-40% homology allowed) Similar organisms in train/test Balanced representation of all localization types

Hard splits likely have: Lower sequence identity between train/test (e.g., <20-30% homology) Phylogenetic distance: Train on yeast/bacteria, test on human Rare localizations: Test on underrepresented compartments (peroxisome, chloroplast) Multi-localization complexity: Test on proteins with multiple localizations

Key Challenge for FLIP Unlike other FLIP tasks that test extrapolation in fitness values (e.g., low→high fitness), SCL tests: Sequence-based generalization: Can models predict localization for proteins unlike any in training? Multi-label complexity: Handling proteins that go to 2-3 compartments Cross-species transfer: Learning universal localization rules vs. species-specific signals