10.1 Repeatable means someone else obtains the same object #
10.1.1 Versions and dates #
A result is a program, a version, and a set of parameters, run against a named database. “We used BLAST” is not a methods sentence. Record the version, the query, the database release or the download date, the matrix, the gap costs, the filters, the cutoff, and an identifier for the queries. A sampler needs its seed. A Bayesian tree needs the convergence check from Chapter 6.
The bit score of an alignment can be compared next year. The E-value cannot, because the database has grown. A coordinate means nothing on a different assembly until it has been translated. Defaults change between versions and move variant calls with no change in the reads. Repeating yesterday’s commands on this morning’s RefSeq is a new experiment. It may be a better one. It is not a replication.
10.1.2 Contamination of the training set #
A score on sequences the method was trained on measures memory. Homology leakage is the subtle form: one structure is held out, and a neighbour at 90% identity remains in the training set. State the redundancy cutoff. For folds, even 30% identity can carry the answer. Random cross-validation does not prevent this if homologues were free to fall on both sides of the split.
The same circle appears in annotation. A function written into database B by BLAST, then “confirmed” because it is found in B, has not been tested. Features chosen on every row before the split have already seen the labels. Seal the test set first. The query can be contaminated too. Human or vector sequence inside a bacterial assembly hits human or vector perfectly. The alignment is real. The biology is not. Too much identity to the wrong species is the signature.
10.2 Two statistical traps #
10.2.1 Multiple testing is everywhere a list is #
An E-value corrects for the size of one database search. It does not correct for fifty motifs, five models, or a hundred queries, if only the prettiest number is kept.
A family-wise bound such as Bonferroni suits one gene or one site you intend to assert. A false-discovery rate suits a list you will follow up. Skipping the correction is legitimate only for a single hypothesis named before the data were examined. A threshold of 0.05 taken from a one-test course is not a property of a genome.
10.2.2 The base rate #
Rarity defeats an impressive sensitivity and specificity. The arithmetic is an illustration, not a clinical series. Take prevalence 1 in 1,000, sensitivity 99%, and specificity 99%. In 100,000 people, 100 are affected and 99 of them are true positives. Of 99,900 unaffected people, 999 are false positives. The positive predictive value is 99/1,098, about 9%. About nine in ten positive calls are wrong.
At prevalence 1 in 10 the same test yields 9,900 true positives and 900 false positives, so the predictive value is 9,900/10,800, about 92%. The test did not improve. The base rate did. Accuracy on a balanced benchmark does not transfer to a search in which the true class is rare: most bases are not variants, and most genes are not differential. The thousand false expression calls of Chapter 8, under a global null, are the same fact.
10.3 Before you believe a hit #
10.3.1 A checklist #
Name the search: program, version, database, date, matrix, gaps, filters, and how many queries were run. Put the bit score beside the E-value, then read the alignment. Are the identities in a known motif, scattered, or low-complexity? Do gaps cut a core, and are the functional residues in the columns? A hit covering 5% of a protein is a fragment. Quote a length and a denominator.
If the claim is orthology, require a reciprocal best hit and remember co-orthologues. For family membership, prefer a profile HMM. If the subject’s function is electronic, the hit has inherited a hypothesis: follow the evidence code, or keep the hedge. Species, neighbourhood, domain order, and expression should make the function plausible, and contamination should be excluded before a surprise is announced. Similarity supports “worth testing”. It does not support “the pathway works”, “the variant is pathogenic”, or “this is a drug target”. The experiment has to match the sentence.
10.3.2 A BLAST story that is too good #
The following report is a hypothetical, not a published study. A soil contig contains an open reading frame of 87 amino acids. BLASTP against the non-redundant database returns human cytochrome P450 3A4, with E-value printed as 0.0 and identity 100%. The soil community is therefore said to detoxify drugs by the human hepatic route, and the enzyme is called a drug target. The alignment was not saved, because an E-value of zero was taken to prove the claim.
The printed 0.0 is underflow. For a long identical alignment a tiny E-value is ordinary, and the alarming figure is the identity. One hundred percent identity to a human enzyme, recovered from soil, is contamination or a mislabelled record, not a divergent bacterial homologue. Bacterial P450s exist and are typically far from human CYP3A4. A real distant homologue would show moderate identity, a haem motif, and a profile hit. Eighty-seven residues are not that enzyme: the canonical human protein is about 500 residues long. “100% identity” without a length is either identity on a fragment or a scrap of human sequence, and only the alignment distinguishes them. Database, date, matrix, and the other queries are unstated. The next hit might have been vector.
A charitable reading still fails the conclusion. Membership of the P450 superfamily, even if the domain and the catalytic residues agree, is not “the same substrates as a hepatocyte” and is not “a drug target”. The open reading frame as described has no substrate, no expression, and no genomic context. The first jobs were the alignment and a contamination decision. If anything biological remained, the honest stop was “candidate, untested”. The zero was the least informative token in the paragraph.
10.4 Key terms, summary, and self-check #
Key terms. Database date; version; homology leakage; circular validation; family-wise error; false-discovery rate; base rate; positive predictive value; reciprocal best hit; contamination.
Summary. Repeatable work names versions, parameters, builds, and database releases, because E-values and coordinates move. Testing on training sequences, or their close homologues, or on annotations the method wrote, measures memory. One claim needs a family-wise bound; a list needs a false-discovery rate. In the illustration, 99% sensitivity and specificity give a positive predictive value near 9% at prevalence 1/1,000 and near 92% at 1/10. Before believing a hit, name the search, read the alignment, demand reciprocity for orthology, trace the evidence, and match the experiment to the sentence. Reported 100% identity of an 87-residue soil ORF to a human enzyme, with E printed as 0.0, is contamination or a fragment until the alignment shows otherwise.
Self-check.
1. Why can the same alignment have a different E-value next year, and which number still compares its strength?
2. A test protein was held out, but it is 80% identical to one in the training set. Why is the accuracy not independent?
3. In the 1/1,000 illustration, about how many positive calls in 100,000 people are false, and what changes at prevalence 1/10?
4. Reciprocal best hit fails. Which orthology claim is no longer available, and what can a profile hit still support?
5. In the soil story, why is E = 0.0 not the surprise, and which facts suggest contamination rather than a new bacterial enzyme?
—
# Closing: how the chapters fit together
Chapters 1 and 2 fix what a data object is, and how little a record guarantees, before an algorithm is allowed a conclusion. Pairwise alignment is the operation inside search, inside a multiple alignment, in the columns of a tree, and behind annotation and variant calls. A bad column becomes a bad tree and then a confident bootstrap. An E-value says whether a resemblance is surprising in a named database. It does not grant homology, function, or orthology.
Assembly builds the coordinates on which expression and variants are counted, and repeats, long reads, and proximity links decide what those coordinates omit. Counts then need a model, a correction for how many tests were run, and a base rate before a rare class is called. Structure prediction borrows alignments and the experimental archive, and returns a confidence that is not a resolution. The last chapter is the scepticism the others assume: versions, a sealed test, the number of looks, and the alignment read by eye.
Revision checklist #
1. Define homology, similarity, and identity, and rewrite “40% homologous” as a sentence that could be true.
2. Fill `ATGC` against `AGC` at +1/−1/−1, give the traceback, and say why affine gaps exist.
3. Say what PAM and BLOSUM were counted from, and which way each number runs as sequences diverge.
4. Say what happens to E and to the bit score if the database grows tenfold.
5. Name the three profile-HMM states, and say what a logo’s stack height measures.
6. Say what likelihood maximises, what a bootstrap resamples, and why long-branch attraction does not fade as sites are added.
7. Define contig, scaffold, and N50, and state the unseen fraction at coverage c, with one assumption repeats violate.
8. Explain why raw RNA-seq counts are not comparable across libraries, and what a false-discovery rate controls.
9. Give the homology-modelling rule of thumb, and two ways high local confidence can still leave a structural claim unsupported.
10. List the checks on a spectacular BLAST hit: the alignment, reciprocity, the evidence behind the label, and the base rate.
Selected sources #
These notes are an original teaching summary. They are not a transcript of any textbook.
- Mount DW. Bioinformatics: Sequence and Genome Analysis. 2nd edition. Cold Spring Harbor Laboratory Press, 2004.
- Pevsner J. Bioinformatics and Functional Genomics. 3rd edition. Wiley-Blackwell, 2015.
- Xiong J. Essential Bioinformatics. Cambridge University Press, 2006.
- Durbin R, Eddy SR, Krogh A, Mitchison G. Biological Sequence Analysis: Probabilistic Models of Proteins and Nucleic Acids. Cambridge University Press, 1998.
- Needleman SB, Wunsch CD. A general method applicable to the search for similarities in the amino acid sequence of two proteins. Journal of Molecular Biology 48:443–453 (1970).
- Altschul SF, Gish W, Miller W, Myers EW, Lipman DJ. Basic local alignment search tool. Journal of Molecular Biology 215:403–410 (1990).
—
# Part C. Visual Figures & Scientific Diagrams Gallery
Comprehensive high-resolution visual reference repository for Bioinformatics & Computational Biology from the B10logy Visual Atlas.
Blast Sample Output
—
# Part C. Visual Figures & Scientific Diagrams Gallery
Essential high-resolution reference diagrams for Bioinformatics and Computational Biology from the B10logy Visual Atlas.
