hazemessam/aav

---
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

许可证：AFL-3.0
标签：生物学

FLIP AAV 数据集划分
FLIP中的AAV（腺相关病毒，Adeno-Associated Virus）数据集专注于预测用于基因治疗的AAV衣壳变体的适配性与效率。

注意：训练时请丢弃所有split=nan的样本。保留此类样本的目的是为了与原始数据集保持一致。

- low_vs_high（低性能vs高性能划分）
- 训练集：适配值等于或低于野生型的序列
- 测试集：适配值高于野生型的序列
- 用途：验证模型能否从低性能序列外推以预测高性能序列

- one_vs_many（单突变vs多突变划分）
- 训练集：与野生型恰好存在1处突变的序列
- 测试集：存在多处突变（超过1处）的序列
- 用途：验证模型从单突变序列向多突变序列的泛化能力

- two_vs_many（≤2突变vs更多突变划分）
- 训练集：与野生型突变数≤2的序列
- 测试集：突变数超过2的序列
- 用途：验证在低突变数序列上训练的模型，能否预测高突变数序列的适配性

- seven_vs_many（7处突变vs其他划分）
- 训练集：与野生型恰好存在7处突变的序列
- 测试集：突变数不同的序列
- 用途：验证当模型在特定突变数的序列上训练时的泛化能力

- des_mut（设计序列vs突变序列）
- 训练集：设计序列（理性设计的变体）
- 测试集：随机突变体
- 用途：验证在设计序列上训练的模型，能否预测随机突变体的适配性

- mut_des（突变序列vs设计序列）
- 训练集：随机突变体
- 测试集：设计序列
- 用途：与上述任务相反——验证在随机突变体上训练的模型，能否预测设计序列的适配性

- sampled（随机采样划分）
- 采用随机80/20比例的训练测试划分
- 用作基线对比基准