hazemessam/meltome

FLIP Meltome Dataset Explanation What is the Meltome Dataset? The Meltome dataset is a thermostability prediction dataset derived from the Meltome Atlas, a large-scale study that measured the melting temperatures (Tm) of proteins across the tree of life.

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.

1. mixed_split.csv

• Purpose: Cross-species diversity split
• Description: Uses MMseqs2 clustering with >20% sequence identity
• Clusters proteins at 20% identity threshold
• 80% of clusters → training set
• 20% of clusters → test set
• Goal: Avoid information leakage between train and test (ensures proteins in test are sufficiently different from training)
• This tests the model's ability to generalize across diverse protein families and species

2. human.csv

• Purpose: Human-specific proteins
• Description: Contains only protein sequences from humans
• Likely includes data from multiple human cell lines, tissues, and body fluids
• The Meltome Atlas includes human TPP data from 14 cell lines, primary cells, tissues, and 5 body fluids

3. human_cell.csv

• Purpose: Human cellular proteins
• Description: Subset of human proteins specifically from intact cells (not lysates)
• The Meltome Atlas distinguishes between:

Cell-based measurements can show different Tm values due to cellular context effects

Labels/Target Variable Each protein sequence has an associated Tm value (in °C):

This is a regression task - you predict the continuous melting temperature Example: A protein might have Tm = 55.3°C

Why This Dataset Matters

Protein engineering: Thermostable proteins are valuable for: Industrial enzymes (operate at higher temperatures) Pharmaceutical development Starting points for directed evolution

Machine learning challenge: Tests whether models can predict thermostability from sequence alone Cross-species prediction is particularly challenging Models must learn what sequence features confer thermal stability

Key Challenge Research has shown that models achieving high Spearman correlation on cross-species data (mixed_split) may be misleadingly good - they often learn to distinguish global amino acid composition differences between species rather than the specific sequence features that determine thermostability within a species. This is why having both mixed_split (cross-species) and human / human_cell (species-specific) splits is important for proper evaluation.