Newbie question about microarray analysis

Hello,

I apologize in advance for the extreme vagueness of the following
question -- a friend is asking me if I can help with a project of his
on protein microarrays, and I don't know anything about the field
myself.

Here goes: for someone new to the field, how long would it take to
learn to do some meaningful analysis of protein microarrays with a
software package? I heard TM4 is good ([Only registered users see links. ]). I have a
strong mathematical background but little knowledge of molecular
biology. Is this generally the kind of thing that can be picked up in a
week or so?

Any kind of response would be greatly appreciated.

Thank you,

Rex

Rex Eastbourne wrote:

Ultimately it depends what you want to do.

There is no need for a molecular biological background if you want to
help your friend to analyze his data. If you know how to analyze a
robust data set in general, then you can help him out.

All you need to be aware of is that essentially the data will be in the
form of a large matrix, typically each row representing a
gene/protein/cell, and each column representing different
sample/time/dose. You want to arrange the data by some algorithm either
by row or by column for all the data points, so that some pattern will
reveal itself. If you have a mathematical model, you can see which data
fit the model. If you don't, you can build a model (most people use a
linear model or a fourier) based on the profile generated. Then let your
friend interpret the results. In some ways it is "better" because you
cannot be biased in your analysis.

One package that is commonly used is "R". The advantage is that it is
just a good statistical package but the learning curve is a bit steep.
Or you can simply use any software that allows you to decompose a data
matrix or cluster it (Eisen's group has freeware that does heirarchical
clustering).

Just be aware that any analysis is only as good as the quality of the
data obtained, and there is no standard in quality for any arrays out
there (i.e. the data is very noisy).

Austin

Thank you, Austin! Very helpful.

I've got a better description of what I need to do now. I have a
collection of proteins in a spreadsheet, and need to find out:

-Are some of them related to one another?
-Do they participate in a single transduction pathway in the cell or
between cells?

Does this sound like a hard task for someone with lots of math but no
bio?

Thanks again,

Rex

Rex Eastbourne wrote:

Hi Rex...

If it is a matter of "relationship" (in the sense that a set of proteins
behave i.e. go up or down in the same way over a given treatment), then
any k-means algorithm will do the trick. I'm pretty sure this would be a
no-brainer for you. I'm actually surprised they don't give it a go by
themselves...

Anything else (like your second point...though I'm curious about the
experimental design) you'd have to give more details, say:

col1=proteinID
col2=expt1
col3=expt2
etc...

BTW...If you don't want to do it, then don't. Technically, if the person
is doing a PhD, then he/she should know how to do it themselves anyway
and not just hand it off...

Austin


P.S. what is your math background? I should have probably asked that
first...

Hi Austin,

Thanks again for replying. The k-means algorithm should be a snap. But
how do I convert the proteins, which are in the format
"UPSP_SLDJK_HUMAN_P12182" to vectors that can be handled by the
mathematical algorithm (i.e. what is the "distance" between two
proteins)? Is there already a program that does this? (I understand
there's something on the NCBI's website.)

It seems there is no before-and-after. This is a simple measurement of
proteins in people with a disease and without a disease. I do not know
which is which; I simply have a list of proteins.

Mathematical background: graph theory, combinatorics, probability,
theory of computation, linear algebra, multivariable calculus,
differential equations. Also a lot of programming.

I can't tell you how much I appreciate your help. My friend is swamped
and trying to make a deadline. I'm trying to give him a hand but am not
up to snuff it seems

Best,
Rex

Austin P. So (Hae-Jin) wrote:

Rex Eastbourne wrote:

So, if I understand the format of the data:

1. "UPSP_SLDJK_HUMAN_P12182" is just a name...say it is a row id.
2. with that name (i.e. in each row), you will have a series of data
points, each data point corresponding the amount of protein found in
patient X (technically you don't have to know if they have the disease
or not).
3. each column (i.e. patient data) will therefore be a
(multidimensional) data vector, with each protein being an "axis".

patient1 patient2 patient3 patient4
protein1 1 50 49 3
protein2 2 35 30 1
protein3 30 20 20 31

In this way you can apply (hierarchical) k-means clustering on the
column "vectors".

Note that you may not get anything either since ultimately your analysis
is only as good as your data...

Austin

Hi Austin,

I just have a plain list of 200 proteins, without data from the
experiment. I need to cluster the proteins by their inherent
characteristics (function, ancestry). I used the protein database on
the NCBI website to get the sequences. Now, I want to take all these
200 sequences and get some measure of how similar each is to each
other. I figure this would require some specific software that would
allow me to enter all the proteins and see how they're related. I found
ProtoNet, but it seems you can only enter one protein and explore its
specific cluster. Are there any other tools for this I might not be
aware of?

I'm sorry to keep asking you questions like this -- just referring me
to a website that explains this would be greatly appreciated.

Thank you,

Rex


Austin P. So (Hae Jin) wrote:

Rex Eastbourne wrote:


Oh...then just ignore everything I've said...you seem to be asking a
number of different unrelated questions though...

This is not an easy project to do for you then, if you are going to go
outside of the standard tools at NCBI.

A simple comparison can be made through the BLAST server:
[Only registered users see links. ]

A good resource for constructing phylogenic trees (which is what you
seem to want to do here):
[Only registered users see links. ]

Good luck

austin

Thanks a lot Austin!

Austin P. So (Hae-Jin) wrote:

Rex Eastbourne wrote:


The standard way to do this is to use ClustalX, which does an alignment
of amino acid sequences (Needleman/Wunsch algorithm) of every protein
with every other and thus calculates a similarity matrix. With this
matrix a phylogenetic tree is calculated and printed in a text format.
Programs like TreeView can read this and display the trees graphically.
All these programs are freely available on the net.

Note however that this has nothing to do with protein array data. In a
protein array experiment you measure the expression of proteins in cells
or tissues depending on experimental factors (e.g. the presence or
absence of a disease) and then find groups of proteins which react in a
similar way (e.g. expression goes up).

"Similarity" thus has totally different meanings in the two fields.