Datasets:
license: afl-3.0
tags:
- biology
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.
FLIP GB1 Dataset - Complete Explanation What Is GB1? GB1 = Protein G domain B1
The IgG-binding domain of streptococcal Protein G Function: Binds to the Fc region of IgG antibodies Size: 55 amino acids (small domain) Use: Widely used in biotechnology for antibody purification
The Famous GB1 Fitness Landscape This is one of the most studied protein fitness landscapes in molecular evolution, based on two landmark papers:
Olson et al. (2014): All single + double mutants across entire 55aa domain
~3,135 variants (55 positions × 20 amino acids each, plus combinations) Mutations at all positions in the domain Maximum 2 mutations per variant
Wu et al. (2016): Complete combinatorial library at 4 specific positions
160,000 variants (20^4 = 160,000) Only 4 positions mutated: Val39, Asp40, Gly41, Val54 These 4 positions are non-consecutive but in contact with each other All possible combinations of amino acids at these 4 sites
sampled.fasta Standard 80/20 random split Baseline for comparison Not biologically motivated
low_vs_high.fasta Training: Fitness ≤ wild-type (fitness ≤ 1.0) Test: Fitness > wild-type (fitness > 1.0) Challenge: Extrapolate from low/neutral → high fitness Why it's hard: Model must learn what makes proteins better than wild-type from only seeing worse/equal variants.
one_vs_rest.fasta Training: Variants with EXACTLY 1 mutation from wild-type Test: Variants with 2, 3, or 4 mutations from wild-type
Example: Wild-type: VDGV (positions 39,40,41,54) Train: ADGV, VNGV, VDAV, VDGI (1 mutation each) Test: ADAV, VNGI, ANAR, FNLM (2+ mutations)
Challenge: Predict epistatic effects from single mutants only This tests: Can model learn higher-order interactions (epistasis) from only seeing single mutation effects?
- two_vs_rest.fasta Training: Variants with ≤2 mutations from wild-type Test: Variants with 3 or 4 mutations
Challenge: Predict triple and quadruple mutant effects This tests: Can model extrapolate from pairwise epistasis to higher-order epistasis?
- three_vs_rest.fasta Training: Variants with EXACTLY 3 mutations Test: Variants with 1, 2, or 4 mutations
Challenge: Interpolate/extrapolate from 3-mutants
What Makes GB1 Special for Benchmarking?
- Epistasis (Non-Additive Effects) Epistasis = When the effect of one mutation depends on other mutations present Additive model (no epistasis): Effect(ADGV) + Effect(VNGV) = Effect(ANGV)
But with epistasis:
Effect(ADGV) + Effect(VNGV) ≠ Effect(ANGV) Actual could be: much better OR much worse