Real and Virtual Sequences Supporting "Preference-Based Fine-Tuning of Genomic Sequence Models for Personal Expression Prediction with Data Augmentation"

This dataset contains real and virtual genomic sequences used in the study“Preference-Based Fine-Tuning of Genomic Sequence Models for Personal Expression Prediction with Data Augmentation”(bioRxiv preprint: https://doi.org/10.1101/2025.11.09.687505).

The initial release (v1) of the dataset included mixed-case nucleotide characters (lowercase/uppercase).These inconsistencies affected sequence preprocessing pipelines; therefore, we provide a corrected version (v2) with fully normalized uppercase sequences.All replication and downstream analyses for the above manuscript were performed using v2, and users should likewise rely on v2 for running the official code associated with the preprint.

🔹 Real Sequences

Real individual sequences were derived from the GEUVADIS cohort(E-GEUV-1),which provides paired whole-genome sequencing (WGS) and RNA-seq data for 421 phased individuals.Following the approach of Huang et al. (2023), Nature Genetics 55(12), 2056–2059,we extracted 196,608 bp (±98,304 bp around TSS) windows for selected chromosome 22 genes from the hg19 reference genome,substituting individual-specific variants to create personalized haplotype sequences.

Reference:Huang, C. et al. Personal transcriptome variation is poorly explained by current genomic deep learning models.Nature Genetics 55(12), 2056–2059 (2023). https://doi.org/10.1038/s41588-023-01574-w

🔹 Virtual Sequences

To mitigate data scarcity and expand genetic diversity, we generated 1,000 virtual individuals usingthe sim1000G framework (Dimitromanolakis et al., BMC Bioinformatics 20(1), 26, 2019).sim1000G simulates recombination and mutation events based on 1000 Genomes haplotypes, producing realistic population-level variant structures.

References:

Dimitromanolakis, A. et al. sim1000g: a user-friendly genetic variant simulator in R for unrelated individuals and family-based designs. BMC Bioinformatics 20(1), 26 (2019). https://doi.org/10.1186/s12859-019-2611-1

Huang, C. et al. Nature Genetics 55(12), 2056–2059 (2023). https://doi.org/10.1038/s41588-023-01574-w

本数据集包含用于题为“基于偏好微调基因组序列模型以实现数据增强的个性化表达预测”的研究（bioRxiv预印本：https://doi.org/10.1101/2025.11.09.687505）的真实与虚拟基因组序列。
本数据集的初始版本（v1）包含大小写混用的核苷酸字符（小写/大写）。此类不一致性会干扰序列预处理流程，因此我们推出了经过修正的v2版本，其序列均已统一为大写格式。上述手稿的所有复现与下游分析均基于v2版本完成，用户在运行该预印本配套的官方代码时，也应使用v2版本数据集。
🔹 真实序列
真实个体序列源自GEUVADIS队列（E-GEUV-1），该队列包含421个单倍型分型个体的配对全基因组测序（WGS）与RNA测序（RNA-seq）数据。参考Huang等人2023年发表于《自然-遗传学》第55卷第12期第2056-2059页的研究方法，我们从hg19参考基因组的22号染色体选定基因中提取了196,608 bp的窗口（转录起始位点[TSS]上下游±98,304 bp区域），并通过替换个体特异性变异生成个性化单倍型序列。
参考文献：Huang, C. 等人. 当前基因组深度学习模型难以解释个体转录组变异. 《自然-遗传学》, 55(12), 2056–2059 (2023). https://doi.org/10.1038/s41588-023-01574-w
🔹 虚拟序列
为缓解数据稀缺问题并拓展遗传多样性，我们使用sim1000G框架生成了1000个虚拟个体（Dimitromanolakis等人, 《BMC生物信息学》20(1), 26, 2019）。sim1000G基于千人基因组单倍型模拟重组与突变事件，可生成符合真实群体特征的变异结构。
参考文献：
Dimitromanolakis, A. 等人. sim1000G：一款面向无关个体与家系设计的R语言易用型遗传变异模拟器. 《BMC生物信息学》, 20(1), 26 (2019). https://doi.org/10.1186/s12859-019-2611-1
Huang, C. 等人. 《自然-遗传学》, 55(12), 2056–2059 (2023). https://doi.org/10.1038/s41588-023-01574-w