The RF Norm module takes one (Rouskin and Zubradt methods), two (Ding and Siegfried methods), or three (Siegfried method) RC files generated by the RF Count module, and performs normalization to obtain transcriptome-wide per-base reactivities.
Reactivity scores can be computed using 4 methods:

Scoring of RT-stops/nuclease cuts-based methods

[1] Ding et al., 2014 (PMID: 24270811)

Per-base signal is calculated as the ratio of the natural log (ln) of the raw count of RT-stops/nuclease cuts at a given position of a transcript, to the average of the ln of RT-stops/nuclease cuts along the whole transcript:

$$U_i=\frac{\ln (n_{\mathrm{U}i}+p)}{\left(\sum _{j=0}^l{\displaystyle \frac{\ln (n_{\mathrm{U}j}+p)}{l}}\right)}$$

$$T_i=\frac{\ln (n_{\mathrm{T}i}+p)}{\left(\sum _{j=0}^l{\displaystyle \frac{\ln (n_{\mathrm{T}j}+p)}{l}}\right)}$$
where n_Ui and n_Ti are respectively the raw read counts in the untreated (or RNase V1) and treated (DMS, CMCT, SHAPE, or Nuclease S1) samples at position i of the transcript, l is the transcript’s length, and p is a pseudocount added to deal with non-covered regions. U_i and T_i are respectively the normalized number of RT-stops/nuclease cuts at position i in the untreated and treated samples.
Score at position i is then calculated as:

$$S_i=max(0, (T_i-U_i\left)\right)$$
[2] Rouskin et al., 2014 (PMID: 24336214)

The untreated sample is not considered. Per-base RT-stops/nuclease cuts are used as a direct measure of the raw signal.

Scoring of mutational profiling-based methods

[3] Siegfried et al., 2014 (PMID: 25028896)

This method takes into account both an untreated sample, and (optionally) a denatured control sample.
Per-base raw signal is calculated as:

$$S_i=\frac{{\displaystyle \frac{n_{Ti}}{c_{Ti}}}-{\displaystyle \frac{n_{Ui}}{c_{Ui}}}}{{\displaystyle \frac{n_{Di}}{c_{Di}}}}$$

where n_Ti, n_Ui, and n_Di are respectively the mutation counts in the treated, untreated, and denatured samples at position i of the transcript, while c_Ti, c_Ui, and c_Di are respectively the reads covering position i of the transcript in the treated, untreated, and denatured samples.
If no denatured control sample is provided, raw reactivities are simply calculated as:

$$S_i=\frac{n_{Ti}}{c_{Ti}}-\frac{n_{Ui}}{c_{Ui}}$$

[4] Zubradt et al., 2016 (PMID: 27819661)

The untreated sample is not considered. Per-base raw signal is calculated as:

$$S_i=\frac{n_{Ti}}{c_{Ti}}$$
where n_Ti, and c_Ti are respectively the mutations count and the read coverage at position i of the transcript.

Since version 2.9.1, RF Norm allows generating plots of raw RT-stop counts (methods 1-2) or mutation rates (methods 3-4), hence enabling quick visual QC of probing experiments:

Normalization of raw reactivities

Raw reactivity scores can be normalized using 4 different approaches:

#	Method	Description
1	2-8% Normalization	From the top 10% of raw reactivity values, the top 2% is ignored, then any raw reactivity value along the entire transcript is divided by the average of the remaining 8%
2	90% Winsorizing	Each raw reactivity value above the 95th percentile is set to the 95th percentile and each raw reactivity value below the 5th percentile is set to the 5th percentile, then the raw reactivity at each position of the transcript is divided by the value of the 95th percentile
3	Box-plot Normalization	Raw reactivity values greater than 1.5x the interquartile range (numerical distance between the 25th and 75th percentiles) above the 75th percentile are removed. After excluding these outliers, the next 10% of remaining raw reactivities are averaged, and all raw reactivities (including outliers) are divided by this value.
4	Mitchell Normalization	The highest value between the mean of the 90th-95th percentile of all raw reactivities, and the 75th percentile of the raw reactivities > 0.001 is used as the normalization factor by which all raw reactivities are divided. This approach introduced in Mitchell et al., 2023 (PMID: 37334863) is supposedly more robust in cases where most of the bases in a transcript are unreactive. Note: this normalization can only be employed for MaP data, therefore it is only compatible with the Siegfried (3) and Zubradt (4) scoring methods.

Normalized reactivities can be further remapped to values ranging from 0 to 1 according to Zarringhalam et al., 2012 (PMID: 23091593). In this approach, values < 0.25 are linearly mapped to [0-0.35[, values ≥ 0.25 and < 0.3 are linearly mapped to [0.35-0.55[, values ≥ 0.3 and < 0.7 are linearly mapped to [0.55-0.85[, and values ≥ 0.7 are linearly mapped to [0.85-1].

Four-base DMS normalization

Unlike SHAPE reagents, which react rather uniformly with all four nucleotides, DMS strongly prefers A and C. It does, however, also modify U-N3 and G-N1 in a structurally informative way, albeit at substantially lower rates (Mitchell et al., 2023; PMID: 37334863). Since the intrinsic reactivity of the four bases differs by more than an order of magnitude, a single normalization factor calculated across the whole transcript is dominated by A/C residues, and the U/G signal is compressed towards zero. The -4d (or --four-base-dms) parameter addresses this by normalizing each base independently (hence implying -nm 4 -ni), so that reactivities at all four nucleotides are brought onto a comparable scale.

When a base is poorly modified, its normalization factor becomes vanishingly small, and dividing by it would turn noise into seemingly structural signal. For this reason, -4d also introduces a minimum normalization factor of 0.002: any base whose factor falls below this threshold is reported as NaN across the entire window, and the total number of masked windows per base is summarized at the end of the run. This threshold can be changed via -mn (or --min-norm-factor), which can also be used independently of -4d in any -nm 4 run.

Custom normalization factors

By default, the normalization factor is recalculated for every transcript (and, when sliding window normalization is enabled, for every window), which makes reactivities derived from transcripts of very different length or sequencing depth not strictly comparable. A factor calculated once, across the whole transcriptome (and, possibly, across multiple experiments), can instead be provided via the -nf (or --norm-factor) parameter, in which case it is applied to every transcript.

Such factors are typically derived with RF NormFactor. When running rf-normfactor with the -rn (or --run-norm) parameter, this will cause rf-norm to be directly invoked on the input samples with the calculated normalization factors.

Normalization factors can be provided in four ways:

Format	Example	Description
<float>	-nf 0.045	A single factor, by which the raw reactivities of all the reactive bases are divided
<float>,<float>	-nf 0.002,0.308	The 5th and 95th percentiles, as required by 90% Winsorizing (requires -nm 2)
<base>:<float>[,...]	-nf A:0.03,C:0.02,G:0.001,T:0.004	One factor per reactive base (requires -ni)
<base>:<float>:<float>[,...]	-nf A:0.001:0.308,C:0.002:0.257	The 5th and 95th percentiles of each reactive base, as required by 90% Winsorizing (requires -ni -nm 2)

The per-base forms are supported by all four normalization methods, and require -ni (or --norm-independent), as the factor of one base cannot be meaningfully applied to another. Conversely, the two single-factor forms cannot be combined with -ni, as a single factor cannot be applied to each base independently. Bases must be among those specified by -rb (or --reactive-bases), and any reactive base left out of a per-base list keeps being normalized independently for each transcript.

Sliding-window normalization

RF Norm supports data normalization in sliding windows. Windows can be both static (default) or dynamic. When a window size is chosen, data normalization is performed by sliding by the chosen offset, a window of that size. While choice of window's type is irrelevant with SHAPE data, it becomes particularly important when dealing with base-specific reagents. Let consider the example below, in which an RNA has been modified by DMS, which only reacts with A and C residues.



In the above example, use of static windows of size 10 nt results in an erroneous overestimation of base reactivities for certain residues (marked in red). This is caused by the fact that A/C residues are unevenly distributed along the transcript, thus causing certain windows to have far less than 50% of A/C bases (contrary to what it would be expected by chance). Instead, use of dynamic windows of size 10 nt avoids this overestimation, as the window's size is dynamically adjusted to always include 10 A/C residues.
The overestimation effect can also be minimized by increasing the size of static windows.

Usage

$ rf-norm [options] -t treated.rc
$ rf-norm [options] -t treated.rc -u untreated.rc
$ rf-norm [options] -t treated.rc -u untreated.rc -d denatured.rc

To list all available parameters, simply type:

$ rf-norm -h

Parameter	Type	Description
-u or --untreated	string	Path to the RC file for the non-treated/denatured (or RNase V1) sample (required by Ding/Siegfried scoring methods)
-t or --treated	string	Path to the RC file for the treated (or Nuclease S1) sample
-d or --denatured	string	Path to the RC file for the denatured sample (optional for Siegfried scoring method)
-i or --index	string	An RCI index file to be used for all input RC files Note #1: If no RCI index is provided, RF Norm will look for files with .rci extension in the same input folder as the RC files, named after the RC files (e.g., Sample.rc will look for Sample.rc.rci). Note #2: If no RCI file is found, it will be created at runtime, and stored in the same folder of the input RC files.
-p or --processors	int	Number of processors (threads) to use (Default: 1)
-o or --output-dir	string	Output directory for writing normalized data in XML format (Default: <treated>_vs_<untreated>_norm/ for Ding method or Siegfried method without denatured sample, <treated>_norm/ for Rouskin/Zubradt methods, <treated>_vs_<untreated>_<denatured>_norm/ for Siegfried method with denatured sample)
-ow or --overwrite		Overwrites the output directory if already exists
-c or --config-file	string	Path to a configuration file with normalization parameters (see Configuration files paragraph) Note #1: If the provided file exists, the loaded configuration will override any command-line specified parameter Note #2: If the provided file doesn’t exist, it will be generated using the specified command-line (or default) parameters
-sm or --scoring-method	int	Method for score calculation (1-4, Default: 1): 1. Ding et al., 2014 2. Rouskin et al., 2014 3. Siegfried et al., 2014 4. Zubradt et al., 2016
-nm or --norm-method	int	Method for signal normalization (1-4, Default: 1): 1. 2-8% Normalization 2. 90% Winsorizing 3. Box-plot Normalization 4. Mitchell Normalization
-r or --raw		Reports raw reactivities (skips data normalization)
-rm or --remap-reactivities		Remaps normalized reactivities to values ranging from 0 to 1 according to Zarringhalam et al., 2012
-rb or --reactive-bases	string	Reactive bases to consider for signal normalization (Default: N [ACGT]) Note: This parameter accepts any IUPAC code, or their combination (e.g. -rb M, or -rb AC). Any other base will be reported as NaN
-ni or --norm-independent		Each one of the reactive bases will be normalized independently (e.g. -rb AC -ni will independently normalize A and C residues)
-dw or --dynamic-window		When enabled, the normalization window is dynamically resized to include at least that number of reactive bases (e.g. -rb AC -nw 50 -dw instructs RF Norm to normalize reactivities in windows containing at least 50 A/C residues)
-nf or --norm-factor	string	When provided, this will be used as the normalization factor for all transcripts (default behavior is to calculate the normalization factor independently for each transcript) Note #1: four formats are accepted, including one factor per reactive base (see Custom normalization factors paragraph above) Note #2: 90% Winsorizing requires 2 normalization factors, respectively corresponding to the 5th and 95th percentiles of the distribution of raw reactivities
-4d or --four-base-dms		Enables four-base DMS analysis [Mitchell et al., 2023] (implies -nm 4 -ni -mn 0.002)
-mc or --mean-coverage	float	Discards any transcript with mean coverage below this threshold (≥0, Default: 0)
-ec or --median-coverage	float	Discards any transcript with median coverage below this threshold (≥0, Default: 0)
-nw or --norm-window	int	Window size (in nt) for signal normalization (≥3, Default: whole transcript [Ding; Siegfried], 50 [Rouskin; Zubradt]) Note: a maximum window size of 30,000 nt is allowed when -dw (or --dynamic-window) is enabled
-wo or --window-offset	int	Offset (in nt) for window sliding during normalization (Default: none [Ding; Siegfried], 50 [Rouskin; Zubradt])
-D or --decimals	int	Number of decimals for reporting reactivities (1-10, Default: 3)
-n or --nan	int	Positions of transcript with read coverage behind this threshold, will be reported as NaN in the reactivity profile (>0, Default: 10)
-g or --img		Enables the generation of plots of raw reactivity data (RT-stops or mutation rates) for each transcript across all samples (requires R)
-R or --R-path	string	Path to R executable (Default: assumes R is in path) Note: also check $RF_RPATH under Environment variables
		Scoring method #1 options (Ding et al., 2014)
-pc or --pseudocount	float	Pseudocount added to reactivities to avoid division by 0 (>0, Default: 1)
-s or --max-score	float	Score threshold for capping raw reactivities (>0, Default: 10)
		Scoring method #3 options (Siegfried et al., 2014)
-mu or --max-untreated-mut	float	Maximum per-base mutation rate in untreated sample (≤1, Default: 0.05 [5%])
		Scoring methods #1 and #3 options (Ding et al., 2014 & Siegfried et al., 2014)
-il or --ignore-lower-than-untreated		Bases having raw reactivity in the treated sample lower than the untreated control, will be ignored (not used during reactivity normalization) and will be reported as NaNs
		Scoring methods #3 and #4 options (mutational profiling)
-mm or --max-mutation-rate	float	Maximum per-base mutation rate (≤1, Default: 1 [100%])
		Normalization method #4 (Mitchell et al., 2023) options
-mn or --min-norm-factor	float	Reactive bases whose normalization factor falls below this value are reported as NaN (>0) Note: no threshold is applied unless this parameter (or -4d) is specified

Configuration files

RF Norm configuration files are used to provide normalization parameters for the analysis, without the need to manually specify them from the command-line.
Configuration files are composed of a list of key/value pairs, separated by the equal sign (=), or by the colon punctuation mark (:). Keys and values are case-insensitive.
Accepted key/value pairs are:

Parameters	Accepted values	Default value
scoreMethod	"Ding" (or 1); "Rouskin" (or 2); "Siegfried" (or 3); "Zubradt" (or 4)	Ding
normMethod	"2-8%" (or 1); "90% Winsorizing" (or 2); "Box-plot" (or 3); "Mitchell" (or 4)	2-8%
reactiveBases	[ACGTURYSWKMBDHVN] (or "all")	all
normIndependent	TRUE/FALSE; Yes/No; 1/0	FALSE
normWindow	Positive integer ≥ 3	1e9 [Ding; Siegfried] 50 [Rouskin; Zubradt]
windowOffset	Positive integer > 0	1e9 [Ding; Siegfried] 50 [Rouskin; Zubradt]
meanCoverage	Positive float ≥ 0	0
medianCoverage	Positive float ≥ 0	0
remapReactivities	TRUE/FALSE; Yes/No; 1/0	FALSE
	Scoring method #1 options
maxScore	Positive float > 0	10
pseudoCount	Positive float > 0	1
	Scoring method #3 options
maxUntreatedMut	0 ≤ r ≤ 1	0.05
maxMutationRate	0 ≤ r ≤ 1	0.2
	Scoring methods #3 and #4 options
maxMutationRate	0 ≤ r ≤ 1	1

# A sample configuration file

scoreMethod=Ding
normMethod=2-8%
maxScore=10
pseudoCount=1
reactiveBases=N
normIndependent=FALSE
normWindow=1e9
windowOffset=1e9
meanCoverage=1

Output XML files

RF Norm produces a XML file for each transcript being analyzed, with the following structure:

<?xml version="1.0" encoding="UTF-8"?>
<data [attributes]>
    <transcript id=”Transcript ID” length=”Transcript length”>
        <sequence>
            Transcript sequence
        </sequence>
        <reactivity>
            Comma-separated list of reactivity values
        </reactivity>
    </transcript>
</data>

The data tag’s attributes allow keeping track of the analysis performed:

Attribute	Possible values	Description
tool	rf-norm	The tool that generated this XML file
scoring	Ding, Rouskin, Siegfried, or Zubradt	Scoring method
norm	2-8%, Winsorizing 90%, Box-plot, Mitchell, or raw	Normalization method
reactive	[ACGT]	Reactive bases
win	Positive integer ≥ 3	Normalization window's size (in nt)
offset	Positive integer ≥ 1	Offset for normalization window sliding
remap	Boolean	Whether normalized reactivities have been remapped according to Zarringhalam et al., 2012
normIndependent	Boolean	Whether each reactive base has been normalized independently
minNormFactor	Positive > 0	Minimum normalization factor below which a reactive base is masked (optional, only reported when set)
		Scoring method #1 (Ding et al., 2014)
max	Positive float > 0	Score threshold for capping raw reactivities
pseudo	Positive float > 0	Pseudocount added to avoid division by 0 during reactivity calculation
		Scoring method #3 (Siegfried et al., 2014)
maxumut	0 ≤ r ≤ 1	Maximum per-base mutation rate in untreated sample
maxmutrate	0 ≤ r ≤ 1	Maximum per-base mutation rate
		Scoring method #4 (Zubradt et al., 2017)
maxmutrate	0 ≤ r ≤ 1	Maximum per-base mutation rate