Introduction
Reducing the number of lethal analyses is fundamental in fish research, therefore alternative non-lethal approaches need exploration. For this purpose, the investigation of blood transcriptomics is of great interest. Blood transports molecules throughout the body and can be sampled even multiple times without sacrificing of the animal. RNA is relatively stable, and small quantities are enough for parallel massive gene expression profiling through high throughput techniques (i.e. RNA-seq). In fish, blood transcriptomics is still underutilized, although the few studies available show how it can be informative in a wide range of assessments. In this work, we focus on the European eel Anguilla anguilla, which is a critically endangered species with a key aquaculture interest. In nature, yellow eels may live in different salinity habitats (i.e., fresh-, and seawater), and this information is typically obtained through the analysis of otolith microchemistry, which requires the sacrifice of the fish. We combined blood transcriptomics and machine learning (i.e., random forest) to test the ability of transcriptomic blood-based analysis to predict salinity habitat history guided by their otolith-based classification.
Material and methods
A total of 60 eels were caught between July and August 2020, in different freshwater, brackish water and seawater sites in Norway. From each eel, otoliths and 600µl of whole blood were collected. Otolith thin sections were analysed for 24Mg, 43Ca, 55Mn, 88Sr and 137Ba to classify eels into three different salinity habitat behaviours: freshwater resident (FWR), seawater resident (SWR) and inter-habitat shifter (IHS; i.e. eel that switched one or multiple times between freshwater and seawater habitats).
RNA-seq from RNA of whole blood was performed, generating ~43M of paired-end reads for each sample. Reads were mapped against the European eel reference genome with tophat2 [1]. Read count, the step to determine gene expression on the ~31,000 eel genes, was performed with htseq-count [2]. Differential expression (DE) was performed with Deseq2 [3] comparing SWR and FWR eels to assess the pool of genes to use for machine learning. Genes with adjp-value <0.05 were considered as significantly DE.
Random forest with the DE genes was run over normalized gene expression data, using the R package randomForest (www.stat.berkeley.edu/~breiman/RandomForests/). For this analysis, we considered the whole sample set of eels, which had been previously classified into FWR, SWR and IHS according to their otolith microchemistry. From the first round of analysis with all the DE genes, a further reduction was performed based on Mean Decrease in the Gini Index to consider only genes with the highest classification power. RF was then repeated with the top 150, 100, 50 or 30 genes with the highest Gini values.
Results and Discussion
The DE analysis between FWR and SWR detected 3,451 DE genes. Figure 1 shows the principal component analysis of each eel based on the blood gene expression of these DE genes and labelled according to the otolith microchemistry analysis. FWR and SWR eels are well separated on the plot, with IHS samples between the two groups.
Random forest performed using the complete set of DE genes, classified eels based on their salinity habitat with an average error of 21.07%, with mis-assignments mostly for FWR and IHS (Table 1). The further subset of genes selected, considering the Mean Decrease in the Gini Index value improved the overall results, especially in the correct assignment of FWR individuals, reaching a minimum of 6.9% using 50 or 30 genes, where only four misplaced animals belonging to IHS were misclassified (Table 1).
Conclusions
The combination of random forest and blood transcriptomic profiling allowed the assessment of the salinity habitat history of European eels with high accuracy, showing potential to replace its lethal assessment. This approach is promising with respect to replacing or reducing other lethal approaches in fish aquaculture and monitoring, and suitable for both controlled and in-field experiments, as well as for longitudinal studies.
References
1. Kim D, Pertea G, Trapnell C, Pimentel H, Kelley R, Salzberg SL. 2013. doi: 10.1186/gb-2013-14-4-r36
2. Anders S, Pyl PT, Huber W. 2015. doi: 10.1093/bioinformatics/btu638
3. Love, M.I., Huber, W., Anders, S., 2014. https://doi.org/10.1186/s13059-014-0550-8