实际上，chimera检测工具的性能评估很难做到公平统一，各工具有各自的适应范围。

   De novo策略的原理是De novo methodologies are generally based on the fact thatparents of any chimeric sequence have gone through at least onemore PCR cycle than chimeric sequences. 工具有Perseus 、de novo UCHIME 、de novo ChimeraSlayer，这些工具目前都已整合到了mothur中。近来the UPARSE pipeline was released, combining in one step chimera detection with clus-tering of sequencing reads into operational taxonomic units.

   reference-based和de novo两种策略各自具有优缺点：1）reference-based的优势在于In situations dealing with well-studied environments, the reference-based approaches werefound to be very effective in distinguishing between chimeras andchimera-free (parent) sequences. 2）reference-based的劣势在于efficiency is assumed to belower when dealing with less well-known environments.而这正好是de novo方法的优势所在。3）de novo方法的劣势在于most of the de novo approaches depend on redundancy differences between chimeras and parents, assuming that the number of parentsequences has to be at least one time more redundant than theircorresponding chimeric sequences. This requires data abundances to have been reported with high accuracy. （这个就涉及到多少数据量是能保证效果的）

   15年CATCh（Combining Algorithms to Track Chimeras）出现，其原理在于利用其他chimera的检测工具作为input，利用有监督的学习方法去进行分类模型构建，利用测试集验证分类模型的准确度，最终确定分类模型来鉴定嵌合体：which is able to discriminate betweenchimeric and nonchimeric sequences based on a specific set ofinput data (called features in the context of machine learning).在此工具中，输入数据不是测序reads，而是不同工具鉴定chimera的结果。With this tool, we use as input data not the sequence read charac-teristics but rather the results (e.g., scores) of different individual chimera detection tools mentioned above and integrate them intoone prediction. All different tools are run separately, and theiroutput values are combined and processed by the classifier in or-der to give a prediction of whether a read is a chimera or not.  此工具在处理时，主要分为3个步骤：（1）the necessary input features (i.e., output values of the different chimeradetection tools) are identified. （2）the classifier istrained via a supervised learning approach. In this step, the classifier learns to make a correct prediction based on example input data; in our case, training data consist of the output features of a set of sequences reads obtained from different chimera detection tools, together with their correct classification (i.e., whether thisread is a chimeric sequence or not).  （3）In the third step, the trained classifier can be used to predict chimeric sequences in new, previously unseen data (i.e., data that did not belong to the training data).  By feeding the outputs of the different individual chimeradetection tools into the classifier, CATCh is able to classify them into chimeric and chimera-free subsets. As two different types ofchimera detection tools exist, either reference based or de novo, wealso developed two different versions of CATCh. In order to illus-trate its performance, CATCh (reference based as well as de novo)was benchmarked against other chimera detection tools using var-ious publicly available benchmark data. （利用其他工具的检测结果作为输入数据，不同工具的结果出现不一致的情况时，对模型结果是否存在影响）

参考文献：

Mysara M, Saeys Y, Leys N et al. CATCh, an ensemble classifier for chimera detection in 16S rRNA sequencing studies. 2015, 81(5):1573-84. doi: 10.1128/AEM.02896-14