Two new computational methods for universal DNA barcoding: a benchmark using barcode sequences of bacteria, archaea, animals, fungi, and land plants

Abstract

Taxonomic identification of biological specimens based on DNA sequence information (a.k.a. DNA barcoding) is becoming increasingly common in biodiversity science. Although several methods have been proposed, many of them are not universally applicable due to the need for prerequisite phylogenetic/machine-learning analyses, the need for huge computational resources, or the lack of a firm theoretical background. Here, we propose two new computational methods of DNA barcoding and show a benchmark for bacterial/archeal 16S, animal COX1, fungal internal transcribed spacer, and three plant chloroplast (rbcL, matK, and trnH-psbA) barcode loci that can be used to compare the performance of existing and new methods. The benchmark was performed under two alternative situations: query sequences were available in the corresponding reference sequence databases in one, but were not available in the other. In the former situation, the commonly used "1-nearest-neighbor" (1-NN) method, which assigns the taxonomic information of the most similar sequences in a reference database (i.e., BLAST-top-hit reference sequence) to a query, displays the highest rate and highest precision of successful taxonomic identification. However, in the latter situation, the 1-NN method produced extremely high rates of misidentification for all the barcode loci examined. In contrast, one of our new methods, the query-centric auto-k-nearest-neighbor (QCauto) method, consistently produced low rates of misidentification for all the loci examined in both situations. These results indicate that the 1-NN method is most suitable if the reference sequences of all potentially observable species are available in databases; otherwise, the QCauto method returns the most reliable identification results. The benchmark results also indicated that the taxon coverage of reference sequences is far from complete for genus or species level identification in all the barcode loci examined. Therefore, we need to accelerate the registration of reference barcode sequences to apply high-throughput DNA barcoding to genus or species level identification in biodiversity research.

Figure 1. Schematic illustration of the relationship between query and reference sequences.

A query sequence (filled circle) and reference sequences similar to the query sequence (open circle) are shown. The range of nucleotide variation of the genus (gray area) is shown with reference sequences of species and in the genus (A and B, respectively). Distance between the sequences represents genetic distance in the schematic two-dimensional space. (a) A case in which our new criterion works well. The query falls within the nucleotide variation range of genus . (b) A case in which our new criterion might produce misidentification. Because the genetic distance between a query sequence and the sequence similar to it (A) is smaller than the genetic distance between sequence A and sequence B, the query sequence will be assigned to the genus under our new criterion.

Figure 2. Schematic illustration of the NNCauto and QCauto methods.

The processes of the NNCauto method are summarized as follows: (a) By a BLAST-search of a query sequence (Q), a nearest-neighbor sequence (A) is retrieved. (b) By a BLAST-search of A, a borderline sequence (B) is retrieved. (c) By an additional BLAST-search of A, all neighborhood sequences (open circles) are retrieved. Finally, the query is identified at the lowest taxonomic level where the taxonomic information of all the neighborhood sequences including A and B is consistent with each other (i.e., lowest common ancestor algorithm [21]). In the QCauto method, the processes a and b are shared with the NNCauto method, but neighborhood sequences are retrieved by a BLAST-search of Q (d). After the search of neighborhood sequences, the query is identified by the LCA algorithm as in the NNCauto method. A bidirectional arrow indicates genetic distance between two sequences, and a dotted circle represents the range of nucleotide variation that meets the requirement of a BLAST-search.

Figure 3. Frequencies of correctness scores in the no-LOOCV of full-length query sets.

The number of correctly identified taxonomic levels is used as an index representing the degree of correctness of taxonomic assignment. This correctness index has the maximum value 6 when the taxonomic information at all the phylum/division, class, order, family, genus, and species levels is correctly assigned. On the other hand, the index has the minimum value 0 when taxonomic information at all the six taxonomic levels is erroneously assigned to a query or a query remains unidentified even at the phylum/division level. 1NN, 5NN, 97%, 99%, Bar, CNJ, RDP, NNC, and QC means 1-NN, 5-NN, 97%-NN, 99%-NN, Barcoder, ConstrainedNJ, RDPClassifier, NNCauto, and QCauto methods, respectively. (a) Animal COX1. (b) Bacterial/Archaeal 16S. (c) Fungal ITS. (d) Plant matK. (e) Plant rbcL. (f) Plant trnH-psbA.

Figure 4. Frequencies of incorrectness scores in the no-LOOCV of full-length query sets.

The number of incorrectly identified taxonomic levels is used as an index representing the degree of incorrectness of taxonomic assignment. This incorrectness index has the maximum value 6 when the taxonomic assignment of all the six taxonomic levels is incorrect. On the other hand, the index has the minimum value 0 when the taxonomic assignment does not return incorrect results at any taxonomic level: note that this includes the situation in which a query is unidentified even at the phylum/division level. 1NN, 5NN, 97%, 99%, Bar, CNJ, RDP, NNC, and QC represent the 1-NN, 5-NN, 97%-NN, 99%-NN, Barcoder, ConstrainedNJ, RDPClassifier, NNCauto, and QCauto methods, respectively. (a) Animal COX1. (b) Bacterial/Archaeal 16S. (c) Fungal ITS. (d) Plant matK. (e) Plant rbcL. (f) Plant trnH-psbA.

Figure 5. Frequencies of correctness scores in the LOOCV of full-length query sets.

1NN, 5NN, 97%, 99%, Bar, CNJ, RDP, NNC, and QC represent the 1-NN, 5-NN, 97%-NN, 99%-NN, Barcoder, ConstrainedNJ, RDPClassifier, NNCauto, and QCauto methods, respectively. (a) Animal COX1. (b) Bacterial/Archaeal 16S. (c) Fungal ITS. (d) Plant matK. (e) Plant rbcL. (f) Plant trnH-psbA. See the caption of Fig. 3 for the explanation of the correctness index.

Figure 6. Frequencies of incorrectness scores in the LOOCV of full-length query sets.

1NN, 5NN, 97%, 99%, Bar, CNJ, RDP, NNC, and QC represent the 1-NN, 5-NN, 97%-NN, 99%-NN, Barcoder, ConstrainedNJ, RDPClassifier, NNCauto, and QCauto methods, respectively. (a) Animal COX1. (b) Bacterial/Archaeal 16S. (c) Fungal ITS. (d) Plant matK. (e) Plant rbcL. (f) Plant trnH-psbA. See the caption of Fig. 4 for the explanation of the incorrectness index.