One might suggest the goals of functional genomics to be to define cellular roles for all proteins encoded by DNA in a genome, including:

• structure

• biochemical function(s)

• cellular location(s)

• participation in cellular processes

• interactions with other proteins, nucleic acids, and small molecules

The middle three items represent the base categories of the Gene Ontology (GO).

Why do we want to automate prediction of protein structure and function?

• With many genomes completely sequenced, we now have the sequences of thousands of proteins that have not yet been isolated and characterized

• Isolation, purification, crystallization, and X-ray diffraction represent an extremely slow way to learn the structures and functions of these proteins

• Being able to predict complete structure and function from primary sequence data would greatly simplify the task of interpreting the genome

• Further, if we can predict structure, we should be able to design structure

In a recent review of the state of the art, David Baker [Science, 2005, 310, 638] makes the point that structure prediction and design are complementary processes, and employ essentially the same methodology.

Predicting Secondary Structure is relatively easy. For example:

• b-strands tend to have one face hydrophobic, one hydrophilic
  • Since side chains alternate up and down, hydrophobic side chains of period 2 tend to indicate b-strand

• Helices tend to have one hemisphere hydrophobic, one hydrophilic
  • Hydrophobic side chains of period 4 tend to indicate helix

For greater accuracy we need:

• A look-up table of the tendencies of amino acids to appear in various kinds of structure

• Some kind of weighting function, like a PAM or BLOSUM matrix, to deal with how to penalize "wrong" amino acids and reward "right" ones

Results tend to be right 70-75% of the time.

For an example of this and other prediction methodologies, I have used the sequence of the amyloid precursor protein (APP), one of the proteolysis products of which is the Alzheimer's amyloid protein:

MLPGLALLLLAAWTARALEVPTDGNAGLLAEPQIAMFCGRLNMHMNVQNGKWDSDPSGTKT
CIDTKEGILQYCQEVYPELQITNVVEANQPVTIQNWCKRGRKQCKTHPHFVIPYRCLVGEFVSD
ALLVPDKCKFLHQERMDVCETHLHWHTVAKETCSEKSTNLHDYGMLLPCGIDKFRGVEFVCC
PLAEESDNVDSADAEEDDSDVWWGGADTDYADGSEDKVVEVAEEEEVAEVEEEEADDDE
DDEDGDEVEEEAEEPYEEATERTTSIATTTTTTTESVEEVVREVCSEQAETGPCRAMISRWYF
DVTEGKCAPFFYGGCGGNRNNFDTEEYCMAVCGSAMSQSLLKTTQEPLARDPVKLPTTAAST
PDAVDKYLETPGDENEHAHFQKAKERLEAKHRERMSQVMREWEEAERQAKNLPKADKKAVI
QHFQEKVESLEQEAANERQQLVETHMARVEAMLNDRRRLALENYITALQAVPPRPRHV
FNMLKKYVRAEQKDRQHTLKHFEHVRMVDPKKAAQIRSQVMTHLRVIYERMNQSLSLLYNVPA
VAEEIQDEVDELLQKEQNYSDDVLANMISEPRISYGNDALMPSLTETKTTVELLPVNGEFSLDDLQ
PWHSFGADSVPANTENEVEPVDARPAADRGLTTRPGSGLTNIKTEEISEVKMDAEFRHDSGYEV
HHQKLVFFAEDVGSNKGAIIGLMVGGVVIATVIVITLVMLKKKQYTSIHHGVVEVDAAVTPEERHLS
KMQQNGYENPTYKFFEQMQN

Submitted to the PsiPred server (link on the main page), this brought the result by email:

PsiPred Prediction

Going after tertiary structure is MUCH more difficult.

The difficulty is related to the overall protein folding problem.

Levinthal's Paradox [J. Chim. Phys., 1968, 85, 44] suggests that given the size of the conformation space available to a protein of any significant size, finding a particular conformation starting from any other conformation could take longer than the age of the Universe.

Consider a polypeptide with 100 amino acid residues, and two equally possible conformations per residue

• The number of possible conformations is then of the order of 10¹³⁰

• Assume an interconversion time for conformations of the order of 10^-12 seconds (which is too short by several orders of magnitude)

• Then, it will take 10¹⁰ years to sample all the conformations, proceeding randomly

Hence, simply searching the conformation space of a protein for the "best" tertiary structure is futile, even with a supercomputer.

Several methodologies have been developed for dealing with this problem.

We will take a brief look at:

• Homology (comparative) modeling

• Ab initio predictions of secondary and tertiary structure

• Designing new proteins

Comparative Modeling consists of four steps:

• Find a known structure (a template) related to the sequence to be modeled

• Align the new sequence with the template

• Build a model

• Assess the quality of the model

Two general methodologies are used for finding the template:

• Sequence comparison (e.g., Psi-BLAST) with proteins of known structure
  • Models fall off substantially in accuracy below about 30% sequence identity

  • The identity must be distributed in blocks throughout the length of the protein