Normalization - Science topic

More notes from my experience...

The sequelPrep plate wells have a coating up to the volume of ~50uL. To get a more concentrated sample,

1. Avoid touching the plate walls while pipetting, as this may scrape off the coating where the DNA binds.

2. Make sure the PCR product contains the fewest primer dimers possible since it is not essentially a size selection process (generally, bead cleanup is used for size selection). Although the starting material is more than 250 ng, this includes the primers and free nucleotides that add to the concentration. More amplicons in = more amplicons out in this step.

3. If the total volume of PCR product + binding buffer is 50 uL, DNA is bound to the wells up to 50 uL. When we add only 20 uL of the elution buffer, we probably do not get the rest of the DNA bound to the region above 20 uL while eluting. To improve concentration, we could gently vortex the plates.

Just think how you'd do it for, say...a western blot. You'd have all your densitometry for your protein of interest, all your densitometry for your loading control, and you'd...divide the former by the latter to get your per-sample protein expression. Then you'd look at your (now normalized) protein levels and see if anything interesting is happening.

You would never work out the SEM of your loading control and the SEM of your GOI and then...do stuff with those SEMs: that's madness. You normalize each sample first, THEN do your comparisons.

qPCR is no different: you take exactly the same approach and do exactly the same thing, with the exception that qPCR data is in log-space, so divisions become subtractions.

So: let's say you have five samples of treated cells, and five control samples.

Isolate RNA, spec it, QC it, make cDNA.

Now, run qPCR, using these ten samples, in triplicate (so three wells per gene, per sample) for two (or three) validated reference genes and your GOI.

This will give you replicate Cq values for everything.

First thing: manual QC.

Look at the replicate Cq values for each sample, and see if they agree. This is basically "did you pipette the same way three times, and also did anything weird happen": you might see something like 22.2, 21.9, 22.1 as replicate Cq values for GAPDH, and this is great. The final GAPDH Cq for that sample is the arithmetic average of those three Cqs.

If instead you have 22.2, 21.9, 26.8, then you can pretty safely assume well #3 just went weird, and discard that data. Use the mean of the two remaining (non-weird) Cqs.

So eventually you should have, for each sample, mean Cq values for your GOI, and mean Cq values for each reference.

Make your two reference gene Cqs into a "normalisation factor": the average of the two.

Then subtract this from your GOI Cq for that sample. This is the dCt value for that sample, for that GOI: that sample is now normalized.

Do this for all your samples. Now all your data is normalized, and is in the form of dCt values.

Note that dCt is somewhat counter-intuitive, as low numbers are high expression, and vice versa. You can just multiply everything by -1, because the statistical comparisons are exactly the same. "-dCt" is a perfectly valid metric.

dCt values are approximately normally distributed, and entirely statistically-tractable. Use these for your stats. From the example above, you will have, for your GOI, five treated dCt values, and five control dCt values: this can be a straight T-test. If paired data, paired T-test.

There are other ways to do this that are mathematically identical but less "OMG LOG SPACE SCARY", which I typically use because log space scares me, but this basic method works and is easy to work through.

So, TL:DR, ignore SEMs, they're not useful here. Go back to the raw data.

If you have data you're willing to share, I could workshop up a basic spreadsheet for you.

1. It is not uncommon to get skewness even after normalization of the data. Some degree of skewness is inherent to RNA-seq data due to the natural variation in gene expression.

2. It is normal to get some common differentially expressed genes between two different conditions. This might occur due to perturbation in common biological processes or pathways.

3. You do not require any sophisticated formula to calculate fold change. The simplest way is:

a. For each gene, you calculate the average expression value across all replicates in each condition (conditions 1 and 2).

b. Divide the mean expression in condition 2 by the mean expression in condition 1.

Hi El Ro ,

So if I'm understanding your question correctly, you want to normalise every blot to HSP90 as a loading control, then use those normalised values across blots for comparison?

You would need to normalise all bands to loading control, then average normalised values across your control and experimental conditions

So average 2 controls, average 3 disease 1, average 3 disease 2

Then you can compare fold change from the average of your controls vs disease 1 vs disease 2

Does that make sense?

Best wishes,

Steven

It is preferable to avoid downsampling (randomly reducing the number of reads in some samples).

Normalization techniques like DESeq2, edgeR, and limma are preferred for handling differences in sequencing depth across samples, preserving the maximum amount of information. Downsampling, on the other hand, can be used to equalize sequencing depth across all samples but may reduce statistical power, especially in samples with lower read depths. Downsampling may be considered if computational capacity is limited, data size is significant, or extreme cases have vast differences in read depths. In most cases, normalization methods are preferable.

In my opinion, the best normalization technique varies depending on the applied data and the equation. Sometimes it can be Min-Max, sometimes vector, sometimes total normalization, etc. Because the power of a technique's ability to transform data fairly may decrease or increase depending on the circumstances.

Dear Suraj, First of all, thanks for your reply.

In your opinion, "Min-Max scaling and Z-Score normalization are often used in AI/ML applications because they are simple, easy to implement, and often effective in many scenarios". But Max, Vector, Sum, or Logarithmic normalization techniques are also simple and easy to implement. These methods can also be effective in different scenarios.

Secondly, again according to you "Also, many machine learning algorithms are not very sensitive to the particular normalization technique used and these methods usually work well as a starting point". Yes, several types of ML algorithms may not need normalization. But usually, normalization is used in many other methods and seen as a necessity. For example, for a house price prediction, there are many features (criteria) in different units and these are normalized for many ML algorithms.

Again in your view "There are other scaling techniques such as Min-Max scaling and Robust Scaling which are less sensitive to outliers than Z-Score normalization". In fact, apart from these two, there are at least 15-20 other normalization techniques.

I agree with what you said yes "different datasets and algorithms can benefit from different normalization strategies". But it is generally not used and Min-Max and Z-Score are generally used. This is a choice open to criticism. In my view, you have to take into account that some types of normalization may yield different results, especially in cases of decision-making among alternatives for ML.

Hey. Did you get any answers?

The surface charge of a cell can be measured using zeta potential measurements and electrophoresis. Zeta potential measurements are used to measure the cell membrane surface charge of fixed cells in solution. Surface charge density is used to describe the charge distribution on the surface

If this is being done by QPCR you can use an average 16S rRNA gene copy number for all known archaeal genomes and divide your total archaea QPCR 16S count by this number. This average 16S operon/archaea genome is routinely updated as new genomes are sequenced can be found at the following database. https://rrndb.umms.med.umich.edu/ Current average = 1.7 operons/genome

Roberto,

Here is the article partly covering this issue.

yes of course, because our target is to produce strong, rigid materials and avoid carbon as much as possible.

I would calculate the ratio, as you did, of the mutant to the wild-type. Mutants with less of whatever you are measuring would have values less than 1 (possibly 0), and mutants with more would have values greater than 1.

If the numbers vary over a wide range, then you can plot them on a logarithmic scale to spread them out so all the data points can be seen. I would use a log₁₀ scale, personally, but I don't see anything wrong with using a log₂ scale. Since you can't show zero on a log scale, you could omit that data point and simply state in the figure legend or text that no point is shown for that mutant because its value is zero.

It just makes values less than 0 to 0, and greater than 10000 to 1, while anything in between will be from 0 to 1. as the value is divided by 10000.

The way of writing formula is similar to conditional but without using if statement.

For example: (x>=0)*0+(x<0)*1 will evaluate to 0 for positive number and 1 for negatives. the expression (x>0) will either be 1 or 0, based on the value of x.

The exposure time in vermiculite depends on many factors: the material, the dimensions of the parts, their temperature, the number of parts in one box, etc. Usually the exposure is several hours.

You need to interpret OR using Exponential of estimates.

Raquel Andrade, there are two possible solutions of your task. Simplest is to put your LEDs close to focal point of a lens, which image them onto your plane of illumination. More homogeneous distribution can be obtained with two lens system: first lens image of LEDs on the the plane of the second lens, which image surface of this lens into plane of illumination. Due to UV light your lenses have to be silica made.

Good luck!

Actually, the log transformation in the figure I attached was done according to the formula: log((#1+1)/(#2+1)). Just later, I thought that I added "1" to my values to be able to carry out log transformation (not to eliminate zero values). So I considered that maybe, it would be more correct to add "1" to adjusted values just before the transformation.

Thanks again :) Jochen Wilhelm

It actually works (but you have to consider rounding errors):

normalize <- function(x, min, max) (x-min)/(max-min)

denormalize <- function(x, min, max) x*(max-min)+min

x <- rnorm(1000)

r <- range(x)

nx <- normalize(x, r[1], r[2])

dnx <- denormalize(nx, r[1], r[2])

all(x == dnx)

# FALSE ---> rounding errors

all(abs(x - dnx) < 1E-8)

# TRUE ---> identical up to tiny rounding errors

شكرا جزيلا

Joan Jiménez-Balado Thank you Joan

Andrey, thanks for your detailed reply.

As I tried to explain in my question, WT and KO are not paired groups as these samples are coming from different animals.

As a result, in the calculations to make the distribution of differences more symmetric, I don't know which log(MFIs) to subtract. Would it make sense to average the MFI WT and KO used in each separate day and run a paired T-test? If there is evidence the groups are different, can I then do the analysis I propose in the last paragraph of my question?

Hi Jonathan,

Have a look at

Hope it will help you out.

Try this :

Thank you Abbas Thajeel Rhaif Alsahlanee and Aparna Sathya Murthy for addressing the question. It was of great help to me. I figured it out through the documentation of statistical methods in python.

Monika Mrozek I think that based on Johannes Elfner shared, it makes sense NOT to scale the discrete variables.

Hi there Mark, thanks for the reply! That's good to know, I'll go ahead and run the data through normalization.

Hello Milad,

If I understand your query, you're asking whether some adjustment(s) could be made to results such that comparisons of results of one study with those of another could be "apples to apples" (e.g., as if equivalent). The most optimistic answer is a qualified "maybe."

If a large corpus of such studies existed, and you could evaluate the singular and combinatorial effects of each and every relevant nuisance variable with respect to observed outcomes (as sometimes people attempt in a crude way when they regress meta-analysis results on study characteristics), then you might be able to impose statistical adjustments. Otherwise, you'd only be able to impose partial "control" for study differences, and possibly these adjustments would over- or under-correct. This is definitely a case in which "one size will not fit all (cases)."

Good luck with your work.

1) Internal standards (IS) are used to normalize the data for structurlly similar compounds in target or semi-targeted analysis... it is not a good idea to use them to normalize your data in metabolomics, as the structural variability is an important factor influecing the detection.

2) in metabolomics, normalization is performed on QC samples: every 5 samples you should inject a pooled sample (quality control sample = QC) and then you use the data from those injections to normalize your data, using normalization softwares (I suggest NOREVA --> http://idrblab.cn/noreva/)

3) by the way, I suggest to use also IS to check the quality of injection of each sample. If the IS signal is missing or is too intense from any sample, you should remove such sample before proceeding to normalization.

x , y coordinateلازم تكون النقاط الها احداثيات

The only way I can think of to get a negative fluorescence reading is if the intensity of a reference is being subtracted, and the intensity of the reference is higher than the intensity of the sample.

Hi Elyas,

I can't really answer you why this classification is done, however this transformation to categorized data is simply integrated in the function of the package. - so its not a problem - it is part of the analysis itself (using anchor weighting and PCA)

You can find more information about how this works here:

(page 191 and the following - transferData function; and page 78 - find transfer anchor function)

and here:

(page 19 and the following)

hope that helps :)

hello，Bibek。 have you solved this problem? now I have IS and total protein. But I still confused how to normalize thie metabolomics data. should I use IS and protein both or other one of them, could you plesae help me? Thank you.

Hi Pham Quynh Khoa,

I found the following article very useful.

Misra BB. Data normalization strategies in metabolomics: Current challenges, approaches, and tools. European Journal of Mass Spectrometry. 2020;26(3):165-174. doi:10.1177/1469066720918446

Do have a read.

Hope this helps.

jaime

Dear Noor,

the difference of PL and EL consists in the occupation of the excited energy level (here band of conductivity).This is important if you investigate an indirect semiconductor. By PL only a small momentum is transmitted by photons. Therefore, the momentum for occupation of the band of conductivity is generated by interaction with phonons. This process strongly depends on temperature. In contrast, the interaction with electrons allows immediately a high momentum transfer. I know this problem from the investigation of GaP.

Of course it depend on the question you want to discuss. I would select the intensity of the first phonon satellite hv = Eg - hw (hv emission energy, Eg gap energy and hw phonon energy). This emission line largely depends on the sample properties.

With Regard

R. Mitdank

I agree with your point that it is difficult to rationalize the application of one drosophila read-derived correction factor genome wide. Our biggest problem is that Active motif suggests 50ng of spike in chromatin is sufficient to achieve ~2-5% drosophila reads, and we rarely do better than 0.05% drosophila reads for most ChIPs. It seems that I would need to make my own drosophila chromatin, because I cannot afford the amount of spike-in chromatin that I think I would need to get to 2%.

I also agree with your point that for a case like H3K27me3 chip in the presence of EZH2i, there should be very little H3K27me3 available to IP anyways. A case such as this is really the only time we see the drosophila spike in working. This is largely due to how we make the libraries. We balance the ng of DNA going in, even though we know there should be less from the EZH2i sample. Perhaps it is valid to just library prep the same volume of DMSO vs EZH2i rather than the same mass? Then I doubt we would need downstream normalization.

Our real trouble is this : We have a situation where in condition A we see many superenhancer-like sites of H3K27Ac that are lost in condition B. In condition B we also see a massive gain of smaller enhancer and promoter-like sites of H3K27Ac that are not present in A. We know by western that total H3K27Ac is higher in B than in A. When we sequence these ChIPs we have trouble calling the small gains in H3K27Ac in B, and we have trouble calling decreased H3K27Ac at superenhancers in condition A. Drosophila spike in fails miserably here because the H3K27Ac is abundant in both samples, and the antibody is great. The H3K27Ac is locally redistributed and globally increased.

So out of 100M reads, we only get like 200k drosophila reads, which is not statistically reliable. So we try this computational method for normalization and we see exactly what we would expect and what all our other data (ChIP-qPCR, western, etc) suggest we should see.

I have spent quite some time trying to understand this ChIPseqSpikeInFree method. It is true that it falls back on your point that the assumptions you make about the data must be orthogonally validated by other experiments. I think that you must have a hypothesis driven reason to use this computational method or the drosophila spike-in method. For the EZH2i H3K27me3 ChIP, I suppose a western blot is really sufficient to say that H3K@7me3 is depleted, but for H3K27Ac, we need to know where it is depleted and enriched, and we need to account for it being globally increased in abundance.

Our worry is that although this computational normalization idea has been published by bioinformaticists, it has not been vetted by biologists for real world application. I am just worried that even though this seems to work for my experiment, I will get killed by reviewers whether this method is valid or not, because drosophila spike-in is a lot easier to understand and explain.

Hello

Andrei S. Babenka is right and I dont think there is another option!

Spectra normalization is commonly used for reducing the variance of devices (such as laser fluctuation) and environmental factors. It is not designed for reducing noises. You need to perform noise reduction prior to that. You can try smoothing by using digital filters.

Hi Jawad, I've always felt it was ideal to at least stick with the same analysis method for both the submaximal and maximal contractions. So if the submaximal stimulation is a series of spikes at a certain frequency, and peak-to-peak (P-P) is the most appropriate method to get the amplitude, then you'd want the same for the maximal signal. However, P-P does not work well with MVCs, so a maximal M-wave from electrical nerve stimulation would be optimal. If that is not an option, and normalization can only occur with an MVC, then I'd use RMS as your analysis method for both the muscle stimulation and MVC (and not P-P).

Dear Sumaia Mohammed Al-Ghuribi , in addition to the drawn-above answers, I would say there has never been any mistake with the normalization as we already have some datasets with normalized ratings in CF. Just compare all evaluation criterion on the same ratings. BTW, the normalization have a great (either positive or negative) impact on the work of some measures like Jaccard which is effective with numeric ratings [1-5] and non-effective with the normalized values chiefly in high-sparse datasets. Finally, you can just google the normalization topic on ratings in CF and you will surely find a great number of resources on this matter.

Good luck with your work.

Hi Elyas,

To me, you get an ideal difference between your counts (especially low counts). If a gene is not expressed in a sample and lowly expressed in another, log2(counts) would give much higher log-ratios than log2(counts/10000), which would give many more significant genes among these low expressed.

Like RPKM in bulk RNA-seq data that you divide the count per million ( because you usually get millions of counts for bulk RNA). While with UMI single-cell data you get 1000s to 10,000s of counts (as LogNormalize usually uses for single-cell UMI data).

For instance, if you consider a given gene across 4 cells and each with 10,000 total counts, with 0, 1, 2, and 3 counts respectively. The below you can see the difference between data based on 1M and 10K:

Using log(1+counts/million) gives:

[0, 1, 2, 3] -> [0, 4.6, 5.3, 5.7]

Using log(1+counts/10k) gives:

[0, 1, 2, 3] -> [0, .7, 1.1, 1.4]

In this paper, you can find my answer

1)

Neither data no rvariables are "parametric" or "non-parametric". A distribution model can be parametric, what means that the density, mass or distribution function has parameters that define the shape (or the moments).

There are many skewed distribution families that can be expressed in a parametric form (see e.g. the Handbook of statistical distributions: shorturl.at/cAR47). The log-normal is also a parametric distribution.

2)

The distribution of the variables is not (directly) relevant for the validity of the ANOVA. Concretely, the residuals of the entire model should be approximately normal distributed to get a valid interpretation (that's the conditional distribution). For an ANOVA model this automatically means that the distribution within each group should be approximately normal. But it is usually better to have a look at residual diagnostics plots using the data from all groups together. This generalized to models that include metric predictors (regressions, ANCOVA models, multilevel models).

3)

If using logarithms is meaningful depends on the kind of variables you have and how you want to model effects. If effects are proportional to the expectation (mean), using logarithms is actually recommended. An alternative is a gamma-model with log-link (shorturl.at/bily4, https://stats.stackexchange.com/questions/67547/when-to-use-gamma-glms).

The pre-treatments you use will depend on the sample set: noise, level(s) of analyte, number of samples, ranges, etc. In many cases, it is a case of start small (Absorbance data) and add treatments, as needed. Just remember, every treatment can hide/generate peaks and will have an effect on sensitivity.

I agrree to the aforementioned comments. However, it strongly depends on your treatment which gene is not affected under specific conditions and may be - therefore - used as "housekeeper/control". This can be tested by try and error methods.

See ICA papers... FastICA, infomax,...

Peter G Hains This is what I am thinking as well. The next samples I am sending for sequencing will have the same amount of protein so this won't happen. Indeed I have to work with what we have, and a qualitative analysis will be the best. There is no accurate representation in those 7 proteins, perhaps a representation of the most abundant, it is something to discuss.

I will read the manual thoroughly.

Thank you Peter, you have been a true help.

Hello Kym,

Classical factoring algorithms start with a correlation matrix; Pearson correlations standardize all variables as if you had transformed them to z-scores at the outset (indeed, that is the definitional version of the formula for the Pearson product-moment correlation--the product of the first moments about the mean).

The only technical concerns with variables having wildly different scales are: (a) possible loss of computational accuracy due to needing lots of digits (most software packages are likely to use 15.5 digits of accuracy in the internal representation of real number values); and (b) as Holger Steinmetz implies, the interpretation of unstandardized coefficients if you rescale/transform scores.

Starting with LISREL, the trend for SEM programs was to use covariance (not correlation) matrices as the scale information is considered important (e.g., for unstandardized coefficients), as is location (so, means are important as well). However, if your interpretations are based on standardized coefficients (in my experience, most published studies in my areas are), then correlations are fine. Hence, rescaled values are fine.

In general, if a specific variable manifested a spread in my data set of many orders of magnitude, I'd consider rescaling/transforming almost automatically.

Good luck with your work!

I expecte may be interested in z score that is called standard value.

Z=(observation - avarege) /standard deviation.

Standard scores

The raw scores (data) obtained as a result of tests and measurements do not help to infer the levels of individuals tested compared to the mean or comparison between the people themselves, if we know that individuals may be subjected to a set of tests (test battery) and each of these tests enjoy The unit of measurement is different from the other, some of which are measured by the minute and its parts as an expression of the unit of time, including what is measured in meters and its parts as an expression of the unit of distance, including what is measured in degrees or the number of times, and for the purpose of eliminating the units of measurement so that we can collect laboratory degrees and give a uniform degree In the tests we use the standard score.

For the purpose of extracting the standard score it is necessary to extract the mean and standard deviation for each test, through which the standard score can be extracted.

Hello Andac,

The issue with using standardised/normalised variables is that standardised and unstandardised coefficients have different sampling distributions. Stated another way, if the observed variables do not all have the same (similar) variances, then there can be differences in the modeled coefficients (and fit statistic) using observed versus standardised data.

Some of the following may be informative:

Baguley, T. (2009). Standardized or simple effect size: What should be reported? British Journal of Psychology, 100(3), 603-617.

Baguley, T. (2010). When correlations go bad. The Psychologist, 23(2), 122-123.

Greenland, S., Schlessman, J.J., & Criqui, M.H. (1986). The fallacy of employing standardized regression coefficients and correlations as measures of effect. American Journal of Epidemiology, 123, 203–208.

Greenland, S., Maclure, M., Schlessman, J.J., Poole, C., & Morgenstern, H. (1991). Standardized Regression Coefficients: A Further Critique and Review of Some Alternatives. Epidemiology, 2(5). 387-392.

Criqui, M.H. (1991). On the Use of Standardized Regression Coefficients. Epidemiology, 2(5), 393.

Happy researching,

Tim

Any other buffer system which replace the maliec acid buffer ?

To your point 3: this was my understanding of your somewhat fuzzy description.

To your point 5: this is the result of function's non-linearity, nothing else.

I am neither "C" nor "C++" programmer, thus I am unable to follow the calculations shown in your illustration; especially when non-standard data types are in use. Sorry.

Anyway, maybe another hint will be useful. As you probably know, the range obtained during interval calculations certainly contains all possible results (is guaranteed) - but may be oversized. To minimize the chances of such an outcome one should use SUE (Single Usage Expressions). In few words: try to rearrange your expression in such a way that no variable is repeated in its new form. For example, the equivalent resistivity of two resistors connected in parallel is usually written as R = R₁R₂ / (R₁ + R₂), while for interval approach it is much better to use equivalent form of this formula, namely: R = 1 / (1/R₁ + 1/R₂). Or even simpler: "x * x" will usually produce wider interval than "x²". The appropriate transformation of your formulas may be tricky, hard to guess, or even impossible, but try to do your best and eliminate repeating variables as much as you can.

In my opinion, yes. I do not think you can distinguish between live and dead cells when measuring total protein. When you lyse your samples for BCA, all cells would lyse regardless of whether they are alive or dead. It may not matter in the grand scheme of things, as irradiated cells would not proliferate/some might be dead (and disappear from culture by apoptosis) and so your total protein levels should still be lower than non-irradiated cells that continue proliferating. The only way you can distinguish live/dead cells at the protein level is if you are measuring a protein that is expressed exclusively in the live/dead cells. If you want to eliminate dead cells as a variable, use cell viability for normalization purposes instead.

You can use any band based on the chemical groups that do not show changes during any chemical reactions or physical reactions. However, majority of the IR bands are the contributions from multiple groups and, thus, tend to change the intensity by either chemical reactions or physical structural changes. Aliphatic CH stretching bands, in the region of 3000-2800 cm-1, are usually very stable as the potential energy distribution of these bands are nearly 99% pure. In the case of PU, you have relatively well isolated 2972 cm-1 band. However, PIR lacks this band. Yet, there are multiple bands between 2950 and 2850 cm-1 that appear to be relatively stable. Ideally, you should curve resolve these heavily overlapped bands (there are at least three bands, maybe 4) to be more accurate, but by measuring the height (absorbance), you can obtain quasi-quantitative comparison. Use the intensity of one of the band (the one near 2850 cm-1 might be good) and divide the absorbance of the band you want to compare to normalize the intensity then compare the normalized intensity of two different conditions, such as different reaction times.

The attached PDF contains several mistakes.

There also a soft-DTW version similar to a kernel. This gives you values between 0 and 1 like a softmax output. It can be found in the tslearn library (Python).

Thank you

I want to answer this discussion in four series in the first I will focus on the when and why to use denormalization.

As with almost anything, you must be sure why you want to apply denormalization. You need to also be sure that the profit from using it outweighs any harm. There are a few situations when you definitely should think of denormalization:

It’s important to point out that you don’t need to use denormalization if there are no performance issues in the application. But if you notice the system is slowing down – or if you’re aware that this could happen – then you should think about applying this technique. Before going with it, though, consider other options, like query optimization and proper indexing. You can also use denormalization if you’re already in production but it is better to solve issues in the development phase. _{Source: Emil (2016)}

Any thought: Dennis Mazur , Nawroz Abdul-razzak Tahir Jeyris Martínez Gutiérrez Nada QASIM Mohammed Dr R Senthilkumar Lilianna Wojtynek ?

geNorm software (sorry, Excel sheet) will do it for you. If only you can find the original sheet.

If you have enough sample numbers than yes.

Mehmet, PCA package as a complete package contains a normalization process. There is therefore no need to normalize inputs to it.

Thank you Henry Chaplin I already filtered the features by the sinuosity, but there was still noise and I wanted try the dangling endpoints way.. Anyway at the end I filtered using:

1 Trim

2 Unsplit line

3 Sinuosity parameter

4 Merge all the lines with the buffer-Join method

5 Dissolve

6 Length parameter

It's a bit complex but my file (stream network) was really full of noise

We had a similar problem in the context of educational assessment where very often people work with "judgement scales" of both types, i.e. two-sided bounded (like your [1,5]) as well as one-sided bounded (like your [1,inf] or [0, inf]), and they have to merge those distinct scales in some way. Fortunately, there are not so many ways to do such mergings, if you want to keep things simple and choose the right mathematical framework which suggests you the appropriate transformation. Just picking the one or another transformation to fit your data e.g. to a normal (or some other) distribution is IMHO a brute statistical approach which doesn't do justice - plausibly - to the inherent structure of your data. For instance, in our case, it was plausible to assume a sub-additive operation on scores, something of the form a + b - a*b, and derive a structure which turns out to be a semi-module. From here, everything else could be derived - without guesswork. The 1-1 transformation from a bounded to a unbounded interval we used had the relatively simple form: 2-log ( score / (1 - score) ), where 2-log is the binary log function, and scores are numbers between 0 and 1. The resulting numbers are between -inf and +inf, but of course we can adapt it such that the resulting numbers start from 0 onwards. Because the transformation is 1-1, it has an inverse which goes the other way back. Works great! Probably there are statistical techniques for doing tests with your data, but that is none of our concern in our context.

Yes, the RPKM metrics is used to normalize for sequencing depth and transcript length within one sample. You cannot use it to compare expression profiles in different samples. For intra-sample comparisons better use TPM (Transcripts Per Kilobase Million) metrics. Here is a link to a short article on how TPM and RPKM are related (https://www.rna-seqblog.com/rpkm-fpkm-and-tpm-clearly-explained/). And here is a useful guide on how to calculate these metrics using scater from Bioconductor package, see part 7.8.8 "Normalization for gene/transcript length" (https://hemberg-lab.github.io/scRNA.seq.course/cleaning-the-expression-matrix.html).

Convert your raw data for interferon gamma elisa into a 8 t cell receptor -normalized value. The normalization can be done by dividing the IFN-g values by 8 t cell receptor values. Then calculate the fold change between the different controls and treatment groups.

MCF-12F is a floating culture and not supposed to adhere; that's what the "F" stands for in MCF-12F. Sorry so late to reply, just saw this, lol

Hi Phung,

I guess it's impossible to analyze these two types of affymetrix arrays by simple command lines. in fact U133A and U133B share only 168 probes among the 22k sondes in both designs. take a look at this post (https://www.biostars.org/p/283639/). in fact both designs are sold to be used as complementary and can't be compared.

fred

Thank you everyone for your valuable suggestions. I will work on these points.

Are you using microarray data? One common way to check if your data has been normalized correctly is to compare expression distributions with boxplot. If it's normalized correctly the boxes should be even.

(image source :http://telemedicina.med.muni.cz/genomic-proteomic-analysis/index-en.php?pg=E-learning--2-data-analysis--2-2-analysis-of-genomic-data--2-2-1-DNA-microarrays-1--2-2-1-6-normalization)

If you're new, I think this slide might help. (It did for me :) : http://barc.wi.mit.edu/education/bioinfo2007/arrays/slides/arrays_lecture1-color.pdf

Taking log2 is a transformation, not a normalization.

Normalizations make the distributions more similar. They can be applied to untransformed or to transformed data (the choice is not arbitrary; it depends on the kind of normalization what makes sense).

Note that log(R/G) = log(R) - log(G).

In dual-color hybridizations, a frequently used normalization is the lowess-normalization that actually detrends the log(R/G) against the log(R*G) data. It therefore makes sense to calculate log(R/G) first and then normalize.