InpactorDB: A Plant classified lineage-level LTR retrotransposon reference library for free-alignment methods based on Machine Learning

LTR retrotransposons are mobile elements that make up the major part of most plant genomes. Their identification and annotation via bioinformatics approaches represent a major challenge in the era of massive plant genome sequencing. In addition to their involvement in the variation in genome size, these elements are also associated in the function and structure of different chromosomal regions and in the alteration of the function of coding regions, among others. Several plant retrotransposon sequence databases of LTR retrotransposons are available with public access such as PGSB, RepetDB or restricted access such as Repbase. Although they are useful for approaches to identify LTR-RTs in new genomes by similarity, the elements of these databases are not classified down to the lineage/family level. with great depth. Here, we present InpactorDB a semi-curated dataset composed of 130,511 elements from 195 plant genomes (belonging to 108 plant species), classified down to the lineage level. This data set has been used to train two deep neural networks (one fully connected and one convolutional) for fast classification of elements. Used in lineage-level classification approaches, we obtain a score above 98% of F1-score, precision and recall. In order to classify elements of the ‘LTR_STRUC’ and ‘EDTA’ datasets, we used the methodology proposed by Inpactor, which uses homology-based strategy with known coding domains belonging to LTR-RTs. We utilized the RexDB domain library as reference. LTR-RTs were classified into superfamilies, Gypsy (RLG) or Copia (RLC) and sub-classified into lineages according to the similarities of five different amino acid reference domains (GAG, AP, RT, RNAseH, and INT domains). In addition, we applied filters to remove keep only intact elements: 1) to remove predicted elements with domains from two different superfamilies (i.e. Gypsy and Copia), 2) or elements with domains belonging to two or more different lineages, 3) to remove elements with lengths different than those reported by Gypsy Database (GyDB) with a tolerance of 20%, 4) to delete incomplete elements which has less than three identified domains, and 5) to remove elements with insertions of TE class II (reported in Repbase). The final non-redundant version of InpactorDB consists of 67,305 LTR retrotransposons. Both redundant and non-redundant versions of InpactorDB are available in Fasta format in which sequences have identifiers with the following general Identification code: >Superfamily-Lineage-plant_family-specie-source-length-ID, Where Superfamily can is either RLC (for Copia) or RLG (for Gypsy), Lineage/family follows following the RexDB nomenclature, source (can be Repbase, RepetDB, PGSB, LTR_STRUC or EDTA datasets), length, and ID, is a unique number which identify each element inside the InpactorDB.长末端重复序列反转录转座子（LTR retrotransposons）是构成多数植物基因组主体的移动遗传元件。通过生物信息学方法对其进行鉴定与注释，是大规模植物基因组测序时代面临的核心挑战之一。除参与基因组大小变异外，这类元件还与多种染色体区域的结构与功能、编码区功能改变等密切相关。目前已有多款公开或受限访问的LTR反转录转座子序列数据库，例如公开访问的PGSB、RepetDB，以及受限访问的Repbase。尽管这些数据库可通过相似性分析用于新基因组中LTR反转录转座子的鉴定，但现有数据库中的元件并未在谱系/家族水平上实现高精度分类。 本研究构建了InpactorDB，这是一套经半人工整理的数据集，涵盖来自195个植物基因组（隶属于108个植物物种）的130511条序列，且已完成谱系水平的精准分类。该数据集已被用于训练两种深度神经网络（一种全连接网络与一种卷积神经网络），以实现元件的快速分类。在谱系水平分类任务中，该模型的F1值、精确率与召回率均达到98%以上。 为了对"LTR_STRUC"与"EDTA"数据集的元件进行分类，我们采用了Inpactor提出的分析方法，该方法基于同源性策略，利用LTR反转录转座子已知的编码结构域，并以RexDB结构域库作为参考序列集。LTR反转录转座子首先被划分为两大超家族：Gypsy（缩写RLG）与Copia（缩写RLC），随后根据GAG、AP、RT、RNAseH及INT这5种不同的氨基酸参考结构域的序列相似性，进一步划分为不同谱系。 此外，我们通过多轮筛选仅保留完整的功能性元件：1）移除同时包含两种不同超家族结构域的预测元件（即同时包含Gypsy与Copia结构域）；2）移除包含两个及以上不同谱系结构域的元件；3）移除与Gypsy数据库（GyDB）报道的长度偏差超过20%的元件；4）删除仅识别出少于3个结构域的不完整元件；5）移除包含第二类转座因子（TE class II，在Repbase中有记录）插入片段的元件。 最终的非冗余版本InpactorDB包含67305条LTR反转录转座子序列。InpactorDB的冗余与非冗余版本均以Fasta格式提供，序列标识符遵循统一编码规则：>超家族-谱系-植物科-物种-来源-长度-ID，其中超家族可为RLG（对应Gypsy）或RLC（对应Copia），谱系/家族遵循RexDB命名规范，来源可为Repbase、RepetDB、PGSB、LTR_STRUC或EDTA数据集，长度及ID分别为元件的核苷酸长度与InpactorDB内唯一的元素标识编号。