---
license: afl-3.0
tags:
- biology
---

FLIP AAV Dataset Splits
The AAV (Adeno-Associated Virus) dataset in FLIP focuses on predicting the fitness/efficiency of AAV capsid variants for gene therapy.

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.

- low_vs_high
  - Training set: Sequences with fitness values equal to or below wild type
  - Test set: Sequences with fitness values above wild type
  - Purpose: Tests if models can extrapolate from low-performing sequences to predict high-performing ones

- one_vs_many
  - Training set: Sequences with exactly 1 mutation from wild type
  - Test set: Sequences with many mutations (more than 1)
  - Purpose: Tests generalization from single mutants to multi-mutant sequences

- two_vs_many
  - Training set: Sequences with ≤2 mutations from wild type
  - Test set: Sequences with more than 2 mutations
  - Purpose: Tests if models trained on low-mutation sequences can predict fitness of higher-mutation sequences

- seven_vs_many
  - Training set: Sequences with exactly 7 mutations from wild type
  - Test set: Sequences with a different number of mutations
  - Purpose: Tests generalization when training on a specific mutation count

- des_mut (Designed vs Mutant)
  - Training set: Designed sequences (rationally designed variants)
  - Test set: Random mutants
  - Purpose: Tests if models trained on designed sequences can predict fitness of random mutants

- mut_des (Mutant vs Designed)
  - Training set: Random mutants
  - Test set: Designed sequences
  - Purpose: The reverse - tests if models trained on random mutants can predict designed sequence fitness

- sampled
  - Random 80/20 train-test split
  - Used as a baseline comparison