sRNARFTarget: a fast machine-learning-based approach for transcriptome-wide sRNA target prediction

Figures & data

Table 1. Studies from which we collected sRNA-mRNA interacting pairs

Bacterium	Data source
Escherichia coli	[12,17–20]
Pseudomonas aeruginosa	[16,21,22]
Burkholderia cepacia	[23]
Pasteurella multocida	[33]
Salmonella	[24,25]
Mycobacterium tuberculosis	[26]
Synechocystis	[34,35]
Multiple bacteria	[27–29]

Table 2. Training and benchmarking data characteristics

Data	No. of species	No. of sRNAs	No. of pairs
Training	37	176	745
Benchmarking	3	25	144

Table 3. Parameters per ML method used for grid-search CV

Method	Parameter	Values
RF	n_estimators	[500, 600, 800, 1000]
	max_features	[‘sqrt’, ‘log2ʹ]
	max_depth	range(1, 11)
GB	n_estimators	[400, 500, 700, 1000]
	max_features	[‘log2’,‘sqrt’]
	max_depth	range(1, 11)
KNN	n_neighbors	range(1, 50)
	weights	[‘distance’, ‘uniform’]

Table 4. Final benchmarking dataset used for all three programs. The table lists the genome accession used, the number of sRNAs, the number of mRNAs, the number of confirmed interacting pairs (P), and the number of pairs considered non-interacting (N) per bacterial species (from top to bottom: E. coli, Synechocystis and P. multocida)

Accession	sRNAs	mRNAs	P	N
NC_000913.3	22	4,240	101	92,348
NC_000911.1	2	3,179	20	6,324
NC_002663.1	1	1,804	22	1,781

Table 5. 10-fold CV AUROC for the best model per classifier trained on sequence-derived features (trinucleotide frequency difference and tetra-nucleotide frequency difference) of 1490 sRNA-mRNA pairs

	AUROC (mean $\pm$ standard deviation)
Models	Tri nt. difference	Tetra nt. difference
RF	$0.67\pm 0.03$	$0.31\pm 0$
GB	$0.66\pm 0.03$	$0.32\pm 0$
KNN	$0.63\pm 0.03$	$0.45\pm 0.01$

Figure 1. ROC curve for the three programs on Escherichia coli data. The plot shows the sensitivity (also called recall or true positive rate) as a function of the false-positive rate (FPR). The dash line indicates random classifier performance.

Figure 2. ROC curve for the three programs on Synechocystis data. The plot shows the sensitivity (also called recall or true positive rate) as a function of the false-positive rate (FPR). The dash line indicates random classifier performance.

Figure 3. ROC curve for the three programs on Pasteurella multocida data. The plot shows the sensitivity (also called recall or true positive rate) as a function of the false-positive rate (FPR). The dash line indicates random classifier performance.

Table 6. AUROC obtained on each bacterial species included in the benchmark for all three programs assessed

Bacterium	CopraRNA	sRNARFTarget	IntaRNA
E. coli	0.88	0.65	0.62
Synechocystis	0.95	0.63	0.48
P. multocida	0.65	0.44	0.40
Average $\pm$ sd	0.83 $\pm$ 0.16	0.57 $\pm$ 0.12	0.50 $\pm$ 0.11

Figure 4. Rank (lower = better) distribution of 102 Escherichia coli confirmed interacting pairs. The violin plot for each program shows the data density for different rank values and the horizontal line inside each box indicates the median rank of confirmed interacting pairs.

Figure 5. Rank (lower = better) distribution of 22 Synechocystis confirmed interacting pairs. The violin plot for each program shows the data density for different rank values and the horizontal line inside each box indicates the median rank of confirmed interacting pairs.

Figure 6. Rank (lower = better) distribution of 20 Pasteurella multocida confirmed interacting pairs. The violin plot for each program shows the data density for different rank values and the horizontal line inside each box indicates the median rank of confirmed interacting pairs.

Figure 7. Percentage of Escherichia coli confirmed interacting sRNA-mRNA pairs (recall) as a function of percentage top predicted interacting pairs.

Figure 8. Percentage of Synechocystis confirmed interacting sRNA-mRNA pairs (recall) as a function of percentage top predicted interacting pairs.

Figure 9. Percentage of Pasteurella multocida confirmed interacting sRNA-mRNA pairs (recall) as a function of percentage top predicted interacting pairs.

Figure 10. ROC curve for sRNARFTarget and IntaRNA on E. coli and Salmonella data. The plot shows the sensitivity (also called recall or true positive rate) as a function of the false-positive rate (FPR). The dash line indicates random classifier performance.

Figure 11. Rank (lower = better) distribution of 119 E. coli and Salmonella confirmed interacting pairs. The violin plot for each program shows the data density for different rank values and the horizontal line inside each box indicates the median rank of confirmed interacting pairs.

Table 7. Execution time for sRNARFTarget and IntaRNA on benchmarking data. Both programs were run on an Intel Core i7 (2.2 GHz) with 4 cores and 16 GB of RAM computer

Bacterium	No. of sRNAs/ mRNAs	Execution time (HH:MM:SS)
		sRNARFTarget	IntaRNA
P. multocida	1/1804	0:00:31	1:43:16
Synechocystis	2/3179	0:01:18	2:33:02
Salmonella	7/4450	0:05:47	6:18:16
E. coli	22/4240	0:15:56	38:52:43
Average	8/3418	0:05:53	12:21:49

Table 8. CopraRNA web server job execution time on selected sRNA for each bacterium on the benchmark data

CopraRNA web server
Bacteria	sRNA	No. of homologs	Execution time (HH:MM:SS)
E. coli	arcZ	8	08:00:00
P. multocida	gcvB	4	08:19:00
Synechocystis	isrR1	19	17:49:00

Table 9. Sporulation-associated genes in sRNARFTarget top 10% predicted RCd1 targets. Smaller ranks indicate higher confidence of sRNARFTarget in the corresponding target prediction

Gene	Symbol	Annotation	Rank
CD630_19350	spoVS	Stage V sporulation protein S	19
CD630_22460	cspC	Subtilisin-like serine germination related protease	71
CD630_36700		Putative cysteine desulfurase family protein	111
CD630_12140	spo0A	Stage 0 sporulation protein A	114
CD630_08560	oppD	ABC-type transport system, ATP-binding component	125
CD630_26700		ABC-type transport system, ATP-binding protein	138
CD630_15110	cotB	Spore coat protein	163
CD630_13860		Putative allophanate hydrolase subunit 1	187
CD630_24920		Two-component sensor histidine kinase, sporulation-associated spo0A	205
CD630_35690		Sporulation-specific protease	270
CD630_11950	spoIIIAD	Stage III sporulation protein AD	282
CD630_05980	cotCB	Spore-coat protein, manganese catalase	378

Pain A, Ott A, Amine H, et al. An assessment of bacterial small RNA target prediction programs. RNA Biol. 2015;12(5):509–513.

 PubMed Web of Science ®Google Scholar

Han K, Tjaden B, Lory S. GRIL-seq provides a method for identifying direct targets of bacterial small regulatory RNA by in vivo proximity ligation. Nat Microbiol. 2016;2(3):16239.

 PubMed Web of Science ®Google Scholar

Zhang YF, Han KS, Chandler CE, et al. Probing the sRNA regulatory landscape of P. aeruginosa: post-transcriptional control of determinants of pathogenicity and antibiotic susceptibility. Mol Microbiol. 2017;106(6):919–937.

 PubMed Web of Science ®Google Scholar

Pita T, Feliciano J, Leitão J. Small noncoding regulatory RNAs from Pseudomonas aeruginosa and Burkholderia cepacia complex. Int J Mol Sci. 2018 Nov;19(12):3759.

 PubMed Web of Science ®Google Scholar

Ramos CG, Da Costa PJP, Döring G, et al. The novel cis-encoded small RNA h2cR is a negative regulator of hfq2 in Burkholderia cenocepacia. PLoS One. 2012;7(10):e47896.

 PubMed Web of Science ®Google Scholar

Gulliver EL, Wright A, Lucas DD, et al. Determination of the small RNA GcvB regulon in the gram-negative bacterial pathogen Pasteurella multocida and identification of the GcvB seed binding region. RNA. 2018;24(5):704–720.

 PubMed Web of Science ®Google Scholar

Fröhlich KS, Haneke K, Papenfort K, et al. The target spectrum of SdsR small RNA in Salmonella. Nucleic Acids Res. 2016;44(21):10406–10422.

 PubMed Web of Science ®Google Scholar

Ryan D, Mukherjee M, Suar M. The expanding targetome of small RNAs in. Salmonella Typhimurium Biochimie. 2017;137:69–77.

 Google Scholar

Mai J, Rao C, Watt J, et al. Mycobacterium tuberculosis 6C sRNA binds multiple mRNA targets via C-rich loops independent of RNA chaperones. Nucleic Acids Res. 2019;47(8):4292–4307.

 PubMed Web of Science ®Google Scholar

Georg J, Kostova G, Vuorijoki L, et al. Acclimation of oxygenic photosynthesis to iron starvation is controlled by the sRNA IsaR1. Curr Biol. 2017;27(10):1425–1436.e7.

 PubMed Web of Science ®Google Scholar

Georg J, Dienst D, Schürgers N, et al. The small regulatory RNA SyR1/PsrR1 controls photosynthetic functions in cyanobacteria. Plant Cell. 2014;26(9):3661–3679.

 PubMed Web of Science ®Google Scholar