Amino Acid Codon Table
8.1 Expression is a count with a design #
8.1.1 Arrays and sequencing count different things #
A microarray measures hybridisation to probes chosen in advance. Intensity has background, a limited range, and cross-hybridisation, and a transcript with no probe is invisible. RNA-seq counts fragments that map to a gene or a transcript. The number is an integer, an unexpected transcript can appear, and a distinguishing base can separate alleles or paralogues.
Poly(A) selection and ribosomal depletion do not keep the same molecules, and a stranded protocol keeps the strand. A read in a shared exon supports the gene and is ambiguous among isoforms. Gene-level counts are the stabler summary. Transcript-level counts answer a splicing question and are less certain. Say which table you have.
8.1.2 Why raw counts cannot be compared #
Counts scale with depth, with length, and with whatever else is in the library. The same molecular share yields about twice the counts in a library sequenced twice as deeply. A size factor removes that depth effect.
Length matters inside one sample, because a longer transcript yields more fragments. In two million mapped reads, 200 reads from a 1 kb gene and 200 from a 4 kb gene are 100 and 25 reads per kilobase per million. Across samples the same gene has the same length, so length does not justify a differential call. If one transcript dominates a fixed number of reads, every other share falls even when molecules per cell did not. Size factors are therefore taken from the typical gene, often a median of ratios, not from a total a few genes can capture. TPM rescales those length-adjusted rates so that they sum to one million. It is a proportion. Neither TPM nor RPKM is molecules per cell unless an external standard was included.
8.1.3 A test, and then a correction for having tested everything #
Differential expression asks whether conditions differ by more than a model allows, after size factors. Under a Poisson model the variance equals the mean. Biological replicates are more scattered, so a negative binomial sets the variance to the mean plus a dispersion term that grows with the square of the mean. With few replicates, that dispersion is shared across genes of similar abundance.
Twenty thousand tests at 0.05 produce about a thousand false positives if nothing is differential. BenjaminiโHochberg controls the false-discovery rate, the expected fraction of the rejections that are false. Sort the p-values and keep up to the largest i with pโแตขโ โค (i/m) q. For p = 0.001, 0.012, 0.030, 0.200, 0.800 and q = 0.05, the thresholds are 0.01, 0.02, 0.03, 0.04, and 0.05. The first three pass. That does not make each call 95% certain. Bonferroni, testing at q/m, suits one gene you mean to assert. A list with no correction is not a result. If treatment is perfectly confounded with the day of preparation, the test reports the day, and no false-discovery rule removes it.
8.1.4 What an mRNA change is not #
A change in mRNA is not a change in protein. Translation and turnover sit between them, and the proteome is a different measurement.
It is not a change in flux. Pathway control is shared among enzymes; that argument is Biochemistry, and the transcript that moves need not be the controlling step. It is not, by itself, the cause of a phenotype. A tiny effect with a tiny p-value is a precisely estimated small change. Importance is a separate sentence.
8.2 Variation #
8.2.1 A variant is a disagreement that survives the error model #
A call is a disagreement with a reference after alignment. The reference allele is only the allele written there. One disagreeing read is an error until shown otherwise. At about one error per thousand bases, a genome supplies a vast number of one-off mismatches. Depth is how they are outvoted.
Useful evidence is many independent reads, not duplicates; high base quality, not only at the ends; both strands when the chemistry is symmetric; high mapping quality; and an allele fraction near one half or one in diploid germline DNA. Depth four with one alternate is still compatible with error. Depth thirty, on both strands, is a heterozygous candidate. These are illustrations, not cutoffs. `PASS` means this model accepted the call. A low-fraction somatic allele can be real in a tumour and would fail a germline filter. Reference bias under-counts reads that match the reference less easily, especially around indels. Read the pileup.
8.2.2 Where short reads stop #
A structural variant that is large next to a short read is not a column of substitutions. The indirect signs are pairs at the wrong distance or orientation, reads split across a junction, and a change in depth for copy number. All three fail inside repeats. An insertion absent from the reference has nowhere to align, so its bases never enter the pileup.
Paralogues create false variants the same way. Reads from a badly represented copy land on the copy they can find, and the differences look well supported. A call inside a duplication is a claim about a place the aligner may be unable to name. Absence from a file of single-nucleotide calls is not absence of structural variants.
8.2.3 Association is not causation #
A genome-wide association study tests correlation of genotype with phenotype. The result is an association. Linkage disequilibrium means the typed base is often a tag for an untyped neighbour, so the resolution is the block until an experiment separates the candidates.
Population structure creates correlations in bulk. If cases and controls differ in ancestry, variants that track ancestry associate with the disease. Principal components, or a mixed model of relatedness, are the usual corrections. They cannot rescue a design in which continent and phenotype are the same split. The conventional threshold near 5 ร 10โปโธ for common human SNPs is a Bonferroni-style bar for on the order of a million independent tests (0.05/10โถ). It is a convention for that design, not a biological constant, and not the right bar for rare variants or for one gene declared in advance. A hit is not a mechanism. Do not write “causes” for an association.
8.3 Key terms, summary, and self-check #
Key terms. Microarray; RNA-seq; RPKM; TPM; size factor; negative binomial; false-discovery rate; BenjaminiโHochberg; batch; allele fraction; reference bias; structural variant; linkage disequilibrium; population structure; GWAS.
Summary. Arrays hybridise to prechosen probes. RNA-seq counts fragments from a library that decides which RNAs can appear. Raw counts scale with depth, with length inside a sample, and with composition; shares are not molecules per cell without a standard. Differential expression needs an overdispersed model and a false-discovery correction, and a batch aligned with treatment is called as biology. An mRNA change is not a protein, a flux, or a cause. A variant call is a disagreement that survives quality, depth, strand, and mapping. Short reads see large structural variants only indirectly. A genome-wide association is a correlation confounded by linkage and by population structure.
Self-check.
1. Two genes have equal raw counts and one is four times as long. Which comparison needs a length correction: those two genes in one library, or the short gene against itself in a deeper library?
2. In the five-test example with q = 0.05, how many hypotheses are rejected, and why is each not “95% certain”?
3. Why will a false-discovery correction not remove a Monday-versus-Tuesday batch that matches treatment?
4. Support for a variant is only on one strand, at the last base of each read. Why is depth not enough?
5. A SNP passes 5 ร 10โปโธ, and cases and controls were recruited on different continents with no ancestry correction. Give two reasons it may not cause the disease.
