cPAS-based sequencing on the BGISEQ-500 to explore small non-coding RNAs
- 6.8k Downloads
We present the first sequencing data using the combinatorial probe-anchor synthesis (cPAS)-based BGISEQ-500 sequencer. Applying cPAS, we investigated the repertoire of human small non-coding RNAs and compared it to other techniques.
Starting with repeated measurements of different specimens including solid tissues (brain and heart) and blood, we generated a median of 30.1 million reads per sample. 24.1 million mapped to the human genome and 23.3 million to the miRBase. Among six technical replicates of brain samples, we observed a median correlation of 0.98. Comparing BGISEQ-500 to HiSeq, we calculated a correlation of 0.75. The comparability to microarrays was similar for both BGISEQ-500 and HiSeq with the first one showing a correlation of 0.58 and the latter one correlation of 0.6. As for a potential bias in the detected expression distribution in blood cells, 98.6% of HiSeq reads versus 93.1% of BGISEQ-500 reads match to the 10 miRNAs with highest read count. After using miRDeep2 and employing stringent selection criteria for predicting new miRNAs, we detected 74 high-likely candidates in the cPAS sequencing reads prevalent in solid tissues and 36 candidates prevalent in blood.
While there is apparently no ideal platform for all challenges of miRNome analyses, cPAS shows high technical reproducibility and supplements the hitherto available platforms.
KeywordsNext-generation sequencing miRNA Biomarker discovery BGISEQ
Currently, high-throughput analytical techniques are massively applied to further the understanding of the non-coding transcriptome . Still, the full complexity of non-coding RNAs is only partially understood. One class of well-studied non-coding RNAs comprises small oligonucleotides, so-called miRNAs [2, 3].
Among the techniques most commonly used for miRNA profiling are microarrays, RT-qPCR, and next-generation sequencing (NGS), also referred to as high-throughput sequencing (HTS). An excellent review on the different platforms and a cross-platform comparison has been recently published . A detailed examination of technologies, however, frequently reveals a bias. One reason for the respective bias is the ligation step, as, e.g., reported by Hafner and co-workers . For example, the quantification of miRNAs differs between NGS and microarrays as it is dependent on base composition . Especially, the guanine and uracil content of a miRNA seems to influence the abundance depending on the platform used. A substantial strength of NGS is the ability to support the completion of the non-coding transcriptome. Unlike microarrays and RT-qPCR, NGS allows the discovery of novel miRNA candidates. To this end, different algorithms have been implemented, with miRDeep being one of the most popular ones . A substantial part of small RNA sequencing data has been obtained using HiSeq and MiSeq platforms (Illumina) based on stepwise sequencing by polymerase on DNA microarrays prepared by bridge PCR , as well as the IonTorrent systems from Thermo Fisher Scientific using a different type of polymerase-based stepwise sequencing on micro-bead arrays generated by emulsion PCR, the first method proposed for making microarrays for massively parallel sequencing . Another approach is the ligase-based stepwise sequencing also using micro-bead arrays, applied for example by ThermoFisher Scientific’s SOLiD sequencing platform, and which has also been used to analyze and present novel miRNAs .
In the current study, we applied the new combinatorial probe-anchor synthesis (cPAS)-based BGISEQ-500 sequencing platform that combines DNA nanoball (DNB) nanoarrays  with stepwise sequencing using polymerase. An important advantage of this technique compared to the previously mentioned sequencing systems is in that no PCR is applied in preparing sequencing arrays. Applying cPAS, we investigated the human non-coding transcriptome. We first evaluated the reproducibility of sequencing on standardized brain and heart samples, then compared the performance to Agilent’s microarray technique and finally evaluated blood samples. Using the web-based miRNA analysis pipeline miRmaster and the tool novoMiRank , we finally predicted 135 new high-likely miRNA candidates specific for tissue and 35 new miRNA candidates specific for blood samples.
In this study, we examined the performance of three sample types using three techniques for high-throughput miRNA measurements (Illumina’s HiSeq sequencer, Agilent’s miRBase microarrays, and BGI’s BGISEQ-500 sequencing system, see details below). The three specimens were standardized HBRR sample ordered from Ambion (catalog number AM6051) and UHRR sample ordered from Agilent (catalog number 740000). UHRR and HBRR samples were measured in two and six replicates, respectively. As third sample type, we used PAXGene blood tubes. Here, two healthy volunteers’ blood samples were collected and miRNAs were extracted using PAXgene Blood RNA Kit (Qiagen) according to manufacturer’s protocol. The study has been approved by the local ethics committee.
Next-generation sequencing using BGISEQ-500
We prepared the libraries starting with 1 μg total RNA for each sample. Firstly, we isolated the microRNAs (miRNA) by 15% urea-PAGE gel electrophoresis and cut the gel from 18 to 30 nt, which corresponds to mature miRNAs and other regulatory small RNA molecules. After gel purification, we ligated the adenylated 3′ adapter to the miRNA fragment. Secondly, we used the RT primer with barcode to anneal the 3′ adenylated adapter in order to combine the redundant unligated 3′ adenylated adapter. Then, we ligated the 5′ adapter and did reverse transcript (RT) reaction. After cDNA first strand synthesis, we amplified the product by 15 cycles. We then carried out the second size selection operation and selected 103–115 bp fragments from the gel. This step was conducted in order to purify the PCR product and remove any nonspecific products. After gel purification, we quantified the PCR yield by Qubit (Invitrogen, Cat No. Q33216) and pooled samples together to make a single strand DNA circle (ssDNA circle), which gave the final miRNA library.
DNA nanoballs (DNBs) were generated with the ssDNA circle by rolling circle replication (RCR) to enlarge the fluorescent signals at the sequencing process as previously described . The DNBs were loaded into the patterned nanoarrays and single-end read of 50 bp were read through on the BGISEQ-500 platform for the following data analysis study. For this step, the BGISEQ-500 platform combines the DNA nanoball-based nanoarrays  and stepwise sequencing using polymerase, as previously published [13, 14, 15]. The new modified sequencing approach provides several advantages, including among others high throughput and quality of patterned DNB nanoarrays prepared by linear DNA amplification (RCR) instead of random arrays by exponential amplification (PCR) as, e.g., used by Illumina’s HiSeq and longer reads of polymerase-based cycle sequencing compared to the previously described combinatorial probe-anchor ligation (cPAL) chemistry on DNB nanorrays . The usage of linear DNA amplification instead of exponential DNA amplification to make sequencing arrays results in lower error accumulation and sequencing bias.
Next-generation sequencing using HiSeq
Agilent microarray measurements
For detection of known miRNAs, we used the SurePrint G3 8×60k miRNA microarray (miRBase version 21, Agilent Technologies) containing probes for all miRNAs from miRBase version 21 in conjunction with the miRNA Complete Labeling and Hyb Kit (Cat. No. 5190-0456) according to the manufacturer’s recommendations. In brief, 100 ng total RNA including miRNAs was dephosphorylated with calf intestine phosphatase. After denaturation, Cy3-pCp was ligated to all RNA fragments. Labeled RNA was then hybridized to an individual 8×60k miRNA microarray. After washing, array slides were scanned using the Agilent Microarray Scanner G2565BA with 3-μm resolution in double-pass mode. Signals were retrieved using Agilent AGW Feature Extraction software (version 10.10.11).
The new sequencing data using BGISEQ-500 data are available in the Additional file of this manuscript (Additional file 1: Table S3).
The raw reads were collapsed and used as input for the web-based tool miRMaster, allowing for integrated analysis of NGS miRNA data. On the server side, mapping to the human genome was carried out using Bowtie  (one mismatch allowed). miRNAs were quantified similar to the popular miRDeep2  algorithm. The prediction of novel miRNAs was performed using an extended feature set built up on novoMiRank . For classification, an AdaBoost model using decision trees was applied. Novel miRNAs were cross-checked against other RNA resources, including the miRBase , NONCODE2016 , and Ensembl non-coding RNAs. The assessment of the quality of new miRNAs was carried out using the novoMiRank algorithm. A downstream analysis of results including cluster analysis was performed using R. For target prediction, we applied TargetScan 7.1 (http://www.targetscan.org/vert_71/) and predicted for all new miRNAs the targets. With the predictions, we extracted the context++ scores and used them for prioritizing the targets, miRNA-target interactions with context++ scores below 1 were considered as high-likelihood targets. Target networks were constructed using an offline version of MiRTargetLink  and visualized in Cytoscape. miRNA target pathway analysis has been carried out using GeneTrai2 . For the GeneTrail2 analysis, all available categories were analyzed, the minimal category size was set to 4 and all p values were adjusted using Benjamini-Hochberg adjustment.
Raw data analysis
We sequenced six brain, two heart, and two blood samples using the BGISEQ-500 system. The resulting reads were mapped to the human genome allowing one mismatch per read. The 10 samples had a median of 30.1 million reads. Of these, 24.1 million reads mapped to the human genome and 23.3 million reads to miRNAs annotated in the human miRBase version 21. The remaining 0.7 million reads per sample contain potentially new miRNAs.
Technical reproducibility of the BGISEQ-500 and comparison to microarrays
The full list of miRNAs with raw and adjusted p values in t test and Wilcoxon-Mann-Whitney test comparing BGISEQ-500 and microarrays is presented in Additional file 3: Table S1. Overall, the results are well in-line with those obtained between HiSeq NGS and the same microarray platform . Reasons that explain differences between arrays and NGS include different sensitivity levels of the platforms, cross-hybridization of miRNAs with similar sequences on the microarrays or bias in library preparation. Further, effects of the normalization can lead to variations in miRNA quantification.
Biological replicates of blood samples and comparison to other platforms
Expression distribution of miRNAs
Comparison of the distribution and abundance of miRNAs on the microarray platform is difficult since microarrays show a saturation effect. This means that for two miRNAs expressed in a range above the saturation, no difference can be observed. We nonetheless performed the same analysis as presented above, assuming that the sum of all expression counts equals to 100%. In this analysis, miR-451a which is found in 0.8% of HiSeq reads and 45.9% of BGISEQ-500 reads is the highest expressed in microarrays (37.2% of all expression counts), followed by 17% of miR-486-5p.
Prediction of novel miRNAs
miRNA target analysis
For all 926 miRNAs, we predicted targets using TargetScan. To rank miRNA-target interactions, we used the context++ score (distribution of the context++ score across all predictions is provided in Additional file 5: Figure S2). Thereby, we observed an accumulation of high-likelihood targets for tissue-specific miRNAs. Of the 926 miRNAs, the tissue specific had an average 42.8 targets, the neither for blood nor for tissue-specific miRNAs 40.7 targets while for blood-specific miRNAs, only 34.5 targets were predicted. The complex miRNA-target network is presented in Additional file 6: Figure S3. It contains 6014 nodes (5088 genes and 926 miRNAs). Network characteristics such as degree distribution and shortest path length are presented in Additional file 7: Figure S4. The genes with largest numbers of predicted miRNAs targeting the gene were CYB561D1 (229 miRNAs), FBXL12 (174 miRNAs), PML (162 miRNAs), and VNN3 (154 miRNAs). The distribution of miRNAs in the different group is presented as Venn diagram in Additional file 8: Figure S5). Among the predicted target genes that were found only for candidate miRNAs being blood specific was, e.g., HMOX1, heme oxygenase 1, mediating the first step of the heme catabolism by cleaving heme to build biliverdin or HPX, coding for hemopexin. The complex nature of the in silico calculated miRNA-target network requires further analyses to understand whether target genes accumulate in specific biochemical categories such as KEGG pathways or gene ontologies. We thus applied GeneTrail2 separately to the set of genes targeted by blood specific miRNAs, targeted by tissue specific miRNAs and by all other miRNAs. As the background sets, all genes predicted to be targeted by at least a single miRNA were selected and the functionality to compare different enrichment analyses by GeneTrail2 has been used. Enriched pathways seem to be largely relevant for either blood or tissue miRNAs, as Additional file 9: Figure S6 highlights. Tissue specific miRNAs had target genes enriched for DNA damage response, the apoptosis, or RNA polymerase II regulatory region DNA binding while blood miRNAs target genes were, e.g., enriched for TP35 network. Interestingly, tissue miRNA target genes also clustered on specific genomic locations (e.g., 19p12 and 19.q13) while blood miRNA targets did not show such an enrichment. In contrast, blood miRNA targets were enriched for disease phenotypes such as carotid artery diseases. In sum, the enrichment analysis highlights very distinct patterns for blood and tissue miRNA targets. Of course, not only the new miRNAs themselves but also the predicted targets deserve detailed experimental validation.
The advent of next-generation sequencing reduced the costs of sequencing while simultaneously increasing the speed of throughput . Today, the costs for small RNA seq are almost equal to and even lower than miRNA microarrays, although small RNA-seq provides the additional possibility for detecting novel small RNA entities.
In the present study, we investigated two current sequencing approaches supporting massively parallel sequencing, which is of high relevance in small RNA research because of the high dynamic range of these molecules: DNA nanoball -based sequencing by BGISEQ-500 and PCR cluster -based sequencing by HiSeq. An important difference between these techniques is in that the first approach uses linear DNA amplification, and the second uses exponential DNA amplification to make sequencing arrays. The latter approach may in turn lead to amplification errors and some specific biases. Besides this fundamental difference, both approaches have their additional advantages and disadvantages. Specifically for the BGISEQ-500, the library preparation currently takes around three working days, the sequencing itself needs one or at maximum two working days. Each flowcell of the BGISEQ-500 has two lanes. On each of these lanes, 32 Gb data can be generated using single-end reads of length 50 bases. The cost of the reagent and material is around 200 USD for 20 million reads ensuring high-quality data at a reasonable cost.
Recently, we published a manuscript about bias in NGS and microarray analysis for miRNAs , highlighting that the expression of miRNAs on different platforms varies by, for example, the nucleic acid composition. In the validation by RT-qPCR, we focused on miRNAs discordant between the high-throughput platforms. Thereby, we observed cases where the RT-qPCR results were concordant with Illumina HiSeq, with microarrays or with none of the techniques. Therefore, we were especially interested how the BGISEQ-500 platform compares to the HiSeq platform and microarrays with the content from the miRBase for small RNA analysis.
Three miRNAs had high divergence between arrays and BGISEQ-500, among them hsa-miR-4454, which was high abundant in arrays but almost not detectable in BGISEQ-500. According to the miRBase, only 28% of users believe that this miRNA is real. Although such votes have only limited value, they at least indicate that this miRNA may be influenced by technological bias.
For high-throughput sequencing, the library preparation and the kits used play a crucial role for the quality of the sequencing results. Others and we noticed an overly abundance of the miRNA miR-486-5p when using the TruSeq kit (Illumina, San Diego), which seems to be independent of the source of the analyzed material [6, 29, 30]. Using the BGISEQ-500 platform, we observed lower read counts for this miRNA. However, in some cases, the miRNA abundance of BGISEQ-500 matches to the HiSeq sequencing results while microarrays show a different expression level, and in other cases, the BGISEQ-500 deviates from the other platforms and in several cases, all three techniques provide substantially divergent results. The more even distribution of reads of the BGISEQ-500 compared to the HiSeq results facilitates the discovery of new miRNAs, which are expected to be significantly less expressed as compared to the already known miRNAs, especially from early miRBase versions.
With respect to many miRNA currently annotated in miRBase and the rapidly growing number of new miRNAs, it is essential not only to have tools for filtering likely false-positives such as the NovoMiRank tool but also to carry out validation of miRNAs using other molecular biology approaches such as cloning and Northern blotting.
Focusing on the performance of the BGISEQ-500, we found a high technical reproducibility of sequencing results, which was however slightly below the technical reproducibility of microarrays. This fact can have different reasons, e.g., the different limit of detection of microarrays. In contrast to sequencing, microarrays have a saturation effect. With respect to the total number of discovered known miRNAs, performance of the BGISEQ-500 was comparable both to the Illumina and the microarray platform.
In sum, none of the mentioned platforms seems to provide the “ultimate solution” in miRNA analysis. All have their advantages and disadvantages and show some bias for the detection of certain sequence types.
We acknowledge the support of BGI-Shenzhen and Complete Genomics.
The study has been funded by internal funds of Saarland University.
Availability of data and materials
Following publication expression data are available in the gene expression omnibus (GEO).
Setting up the assay were done by CG, XS, AA, SD, CZ, DA, JL, and RD. Generating miRNA data were done by SR, CZ, NL, MH, ZZ, CX, AC, and MN. Evaluation of data was done by TF, CB, NL, YL, and AK. Drafting and revision of the manuscript were done by EM, AK. Study design and set-up were done by YL, CS, XX, EM, and AK. All authors read and approved the final manuscript.
Authors with affiliations 1 and 2 are employed by BGI-Shenzhen, Shenzhen, China, and Complete Genomics (a BGI company), Mountain View, CA, USA.
Consent for publication
Ethics approval and consent to participate
The study has been approved by the local ethics committee (Ärztekammer des Saarlandes).
- 8.Mayer P, Farinelli L, Kawashima EHUhwgcpUS. Method of nucleic acid amplification. In.: Google Patents; 2011Google Scholar
- 9.Drmanc R, Crkvenjakov R. Prospects for a miniaturized, simplified and frugal human genome project. Sci Yugosl. 1990;16(1–2):97–107.Google Scholar
- 14.Tsien RY, Ross P, Fahnestock M, Johnston AJUhwgcpCAAce. Dna sequencing. In.: Google Patents; 1991Google Scholar
- 15.Church GM, Mitra RDUhwgcpEPAce. Nucleotide compounds having a cleavable linker. In.: Google Patents; 2003Google Scholar
- 26.Keller A, Backes C, Haas J, Leidinger P, Maetzler W, Deuschle C, Berg D, Ruschil C, Galata V, Ruprecht K, et al. Validating Alzheimer’s disease micro RNAs using next-generation sequencing. Alzheimers Dement. 2016:12(5):565-76.Google Scholar
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.