Results
The sequence data recovered from INSDC databases ranged in average read
length, between 200 bp (animal microbiome) to 300 bp (aquatic
microbiome) of average read lengths and pre-processing steps (Figure 1).
Notably, the aquatic microbiome data was archived with merged paired
ends, for a total average length of 600 bp. On average across datasets,
read qualities remained high until 200 bp, but were lowest for the soil
microbiome data (Figure 1). Among the 15 samples included for each
dataset, the number of reads per sample varied between 14,088 and 97,218
reads per sample animal, between 73,058 and 31,237 reads per sample for
aquatic, and between 196,518 and 51,092 reads per sample for soil
microbiomes.
Processing sequence data across a gradient of read lengths had
consistent effects across datasets, with the average number of reads
conserved per sample decreasing gradually with read length in data
processed without chimera checking (Figure 2, top), consistent with the
quality profile (Figure 1). Importantly, for lower quality sequences,
trimming reads below 100 bp resulted in the removal of a large
proportion of the original reads during chimera checking, but this was
dependent on sequence quality (Figure 2, bottom). In the soil data,
chimera checking removed over 25% of the original reads when trimming
resulted in reads of 50-80 bp in length, but removed only 11.9 ± 8% if
the reads were trimmed to 100 bp. In the higher quality animal
microbiome dataset, chimera checking reads that were trimmed below 60 bp
in length resulted in a loss of 10.9±2%, while chimera checking reads
that were 70 bp or longer only resulted in a loss of 5.1% of the
original reads.
The percentage of taxonomically unclassified reads decreased sharply
with read length. As expected, this effect was most pronounced at the
lower taxonomic levels, but depended on the diversity and/or prior
characterization of the system (Figure 3). For the less diverse and more
well characterized animal microbiome, read lengths beyond 70 bp achieved
minimal improvements in taxonomic classification, while aquatic and soil
microbiomes achieved minimal improvements in taxonomic classification
beyond 90 and 110 bp respectively. Genus-level classification followed a
similar pattern, but required longer reads to reach saturation: in the
animal microbiome, 87.1% of the community was classified on average
when reads were 120 bp or longer, while in aquatic and soil microbiomes,
trimming to 200 bp achieved a classification of 78.8% and 69.3% of the
community, respectively.
The number of ASVs detected across the 5 control samples in each dataset
was assessed given the same read depth and increasing read length
(Figure S1). With increasing read depth, all datasets detected higher
numbers of ASVs. Trimming reads to 50 bp resulted in the detection of
180 animal, 228 aquatic, and 367 soil ASVs, which increased to 1029,
1833, and 5105 ASVs, respectively, when trimming to 200 bp.
To determine how trimming affected the detection of diversity, the
richness and inverse Simpson index in the 5 control samples of each
dataset were assessed. Within each dataset, alpha diversity increased
with read depth, but saturated more rapidly in less diverse environments
(Figure 4). Richness estimates exhibited a hump with increasing read
length in soil samples, decreasing after 160 reads in line with the
decrease in read quality (Figure 1). Importantly, while similar patterns
were observed for both alpha diversity estimates, Inverse Simpson’s
index was more robust to read lengths, saturating with 70, 100, and 160
bp in the animal, water, and soil samples.
The extent to which shorter read lengths affected the variance in
compositional metrics (i.e., beta diversity) was assessed by measuring
the pairwise Sorensen and Bray-Curtis dissimilarities between the five
control samples in each dataset. For all datasets, increasing read
lengths resulted in gradual, but saturating increases in dissimilarity.
The point of saturation depended on the expected diversity in each
system: animal and soil microbiomes approached saturation with 60 and 90
bp reads, respectively (Figure 5). Bray-Curtis dissimilarities were less
variable (standard deviation of 0.03, 0.04, and 0.03 for animal,
aquatic, and soil microbiomes, respectively) across read lengths than
Sorensen dissimilarities (standard deviation of 0.03, 0.09, and 0.05 for
animal, aquatic, and soil microbiomes, respectively). To further examine
information loss from shorter read lengths, Mantel tests between the
communities resulting from each read length and the dataset trimmed to
200 bp (i.e., the most information-rich version, Figure 6) were
performed using Sorensen and Bray-Curtis dissimilarities. While the
strength of Sorensen-based correlations increased with read length,
Bray-Curtis dissimilarities were more robust, and exhibited little
deviation from the 200 bp dataset. Importantly, this pattern was
consistent regardless of whether a chimera-checking step was included
(Figure S2).
Finally, the extent to which shorter read lengths recovered differences
in alpha and beta diversity between the control and disturbed samples (1
day after disturbance) was evaluated for each dataset (Figure 7) using
Wilcoxon rank sum tests and PERMANOVAs. In general, longer read lengths
were able to better discriminate between the alpha diversity of control
and disturbed samples, both in terms of richness and Inverse Simpson
diversity (Figure 7, top). Above 100 bp, only marginal differences in
discrimination between the richness in the control and disturbed
treatments were detected with increasing read length for both metrics,
with p-values ranging between 0.095-0.055, 0.012-0.010, and 0.151-0.309
for aquatic, animal, and soil microbiomes. respectively. Inverse Simpson
diversity was more robust to decreasing lengths, and discrimination did
not change with reads greater than 90 bp (p=0.008 all the comparison
between control and disturbed in all microbiomes). The discrimination of
beta diversity between control and disturbed samples exhibited similar
patterns, with the exception of the water dataset evaluated with
Sorensen dissimilarities, which indicated that samples became more
similar with longer reads (Figure 7, bottom). Sorensen dissimilarity
between the treatments in the aquatic communities decreased linearly
with read length and discrimination decreased, from p=0.074 with 50 bp
data to p=0.456 with 200 bp data). The abundance-weighted Bray-Curtis
dissimilarities were more robust to read lengths, exhibiting low
variation in general.