8/9/01

Lecture #16 - Computational Biology

GenBank: if new sequence found, it's automatically put in database, set up in the USA by the National Library of Medicine
EMBL database: set up by European Molecular Biology Laboratory
DDBJ
the above three exchange sequences daily
PIR and SwissProt: protein databases, in all 6 reading frames
GenProt- translated sequences from GenBank
http://ncbi.nlm.nih.gov is National Center for biotechnology information; contains links to databases mentioned above
for other kinds of databases, see handout

algorithms- finite sequence of well-defined actions whose purpose is to accomplish a given task
example: take DNA and find ORF - have 6 possible ways
need to know the mechanism of the algorithm, otherwise you may misinterpret your results simple algorithm: sequence assembly
cut up the DNA into fragments, sequence them and then try to re-assemble the fragments into a big piece but if you have repeats, you'll have a problem
could have sequence errors or mismatch. How much mismatch should you allow?
sliding window algorithm - view the property of the sequence as a function of position along the sequence
example: see slide #4 - get a picture of how the base composition varies along the DNA

look at side chain of all amino acids, look at relative solubility (how hydrophilic/ hydrophobic)
different types of hydropathy:
1. Hopp-Woods- window of 6
- look up the value of each amino acid in a table, and average the values over the window, and plot those numbers
- above 0 = hydrophilic (outside face of protein)
- below 0 = hydrophobic (inside face of protein, interact with hydrophobic parts of other proteins)
- used to find out the property of a segment of the protein, not just one amino acid
2. Kyte-Dolittle
- for the entire protein and its structure (not looking at just the side chains of the a.a.)
3. antigenic index
if you superimpose the above three types, it's a little different, so where all three agree is the most accurate
window size is usually figured out by trial and error and finding one that is the best choice

1. CF prediction graph: Chou and Fasman found an algorithm for determining protein folding; find a.a. stretches that always form alpha helices, or beta sheets
- plot the probability that it would form an alpha helix or beta sheet
- gives you a 60% accurate plot of what the protein structure looks like
2. GOR Structure Prediction (squiggles) &endash; employ a slightly different mechanism to find the alpha and beta sheets
- sliding window of 17, look at all 17 values (weighted)
- plotted as a squiggle

example: train it to understand a page of text and speak it; output = phonym (specific sound of English language)
have 3 levels (1-3-7) - 1 output fed by 3 layers which is in turn fed by 7 detectors
first time of read through, computer goes through page of text and makes random sounds
then adjust values and reread the page, and generates new output
network learns spaces between words, vowels then consonants
same mechanism for neural net for protein- you train the neural net with a protein sequence so that it will predict the protein structure
it's about 70-75% accurate

there's different preference for codons depending on the organism
codon usage table- use it for codon preference analysis
slide #14 &endash;
curve shows use of codons in ORF (how the use of codons matches the known codon preference in the Drosophila codon table)
above the line = 95% likelihood that the codon is used
rare codon = used less than 10%
see few rare codons in the middle (good, because it will be translated better); an example of biased codon usage

comparison of sequences of myoglobin from pig and that of cow
sliding window of 10 on top sequence, and scan along side sequence
straight line = clear pattern, relationship
leghemoglobin dot matrix: don't see a diagonal across, but gap in diagonal means intron (it's not in the cDNA)&endash; so the plot shows the locations of introns and exons
ribosomal DNA dot matrix #1 = allow up to 2 mismatches, used to find repeated sequences
ribosomal DNA dot matrix #2 = allow up to 5 mismatches, window of 15, see repeated sequences in background (splotches)
Dot Matrix (Protein&endash; Identity Table): alpha and beta globin; gaps = different codons specifying the same a.a.?
PAM250 table:
- Lys + Phe = -5 (because very diff a.a.)
- Lys + Arg = 3 (both have + charge, similar)
- Lys + Tyr = -4
- Ser + Ser = 2 (low number because Ser is very common, so you can substitue Ser for many a.a.)
- Trp + Trp = 17 (very high number because it's a rare a.a., it's a very important match)
Dot Matrix (Protein&endash; PAM250 Table): plot values using the PAM250 table, get a nice diagonal)
can compare your sequence with that of a database (e.g. BLAST). Gives you a value (it's a scoring system) so you can know how similar the sequences are

alignments and gaps- devising a scoring system is difficult
need gap penalties
problem with end gaps- if the gap's length is important, you need to give a high penalty

GenBank database has grown very fast
NCBI website has BLAST search
finding genes on the website - feed your DNA sequence into the algorithm and gives you possible genes and predicts where those sequences are; different algorithms give you different results; if you link that with Swissprot, you can get the protein