Lecture 2.3

RNA quantification

Learning objectives

After this lecture, you should be able to do the following.

Account for how a choice made in the annotation changes the matrix while the reads stay the same.

  1. Explain why a gene missing from the annotation carries a count of zero however much of it the sample transcribed.
  2. Explain why counting exons only and counting the whole gene body give different numbers from one set of reads.
  3. Given a read lying inside an intron and a switch from exon-only counting to gene-body counting, determine how the count for that gene changes.
  4. Given two matrices built from one alignment file that agree everywhere except a handful of genes, determine which differences between the two annotations could produce that pattern.

Account for what a count matrix holds and what quantification is trying to recover from it.

  1. Identify the rows, the columns, and the library-size row of a count matrix, and state what one cell of it holds.
  2. Define the transcriptome as the set of mRNA transcripts a cell could carry.
  3. Contrast what a genome and a transcriptome each report, one what a cell could do and the other what it is doing.
  4. Explain why quantification asks about the number of transcript copies the sample held, a quantity never observed directly, rather than about the reads that were observed.
  5. Describe what happens to a transcript between the sample and the read (fragmentation, amplification, sampling, sequencing error), and state why each step stands between a count and an abundance.
  6. Explain why the transcriptome of one organism differs with environment, developmental stage, and cell type, and what that shows about an annotated gene that the annotation alone cannot.
  7. Given a question about a biological sample, determine whether genome sequencing or RNA sequencing is the one that answers it.

Account for the three rules a counter can apply to an ambiguous read, what each one costs, and how a modern tool splits the read instead.

  1. Name the three rules for a read compatible with two genes: discard it, give each compatible gene half, or split it in the ratio the unambiguous reads suggest.
  2. State that counting tools differ in which of the three rules they apply by default, so one alignment file can give different matrices depending on the tool used.
  3. Describe the loop in which a rough abundance is guessed, each ambiguous read is split in proportion to it, and the abundances are revised until they stop moving, and name it as expectation maximization.
  4. Explain what each of the three rules costs, in reads thrown away or in genes credited with reads they may not have produced.
  5. Given the steps of that loop out of order, put them in the sequence a tool runs them in.
  6. Given a gene family whose members share sequence, predict how the counts of its members change when the rule is switched from discarding ambiguous reads to splitting them in proportion.
  7. Given a result that rests on one member of a duplicated gene family, determine which counting rule is consistent with the reported pattern and what the same data would show under another rule.

Account for why two raw counts from libraries of different size cannot be compared, and determine which sample gives a gene the larger share.

  1. State what \(\text{CPM}_{g,s} = (c_{g,s}/N_s) \times 10^6\) reports, the share of a million mapped reads that belong to one gene, and identify the two quantities it is built from.
  2. State that libraries targeted at one depth come back at different sizes, and name what varies to make that happen (loading, cluster density, pooling).
  3. Explain why a CPM is a share of a library rather than a number of molecules.
  4. Explain why doubling the depth doubles the count and leaves the share alone, so that depth buys precision rather than a different answer.
  5. Explain why two raw counts drawn from libraries of different size cannot be compared, using the case where the raw ratio and the per-million ratio point in opposite directions.
  6. Given two samples with round counts and round library sizes, determine which sample gives the gene the larger share, and whether that agrees with the raw counts.
  7. Given a reported change in one gene between two samples, determine whether it survives normalization, and identify which features of the matrix settle the question.

Account for why transcript length has to be divided out before two genes can be compared, and select the normalization a stated comparison needs.

  1. Distinguish what CPM makes comparable from what TPM makes comparable, columns against one another and rows within a column.
  2. Explain why a longer transcript yields more reads than a shorter one at the same abundance, so that two counts within a sample cannot be compared.
  3. State the two steps behind \(\text{TPM}\), dividing each count by the length of its transcript and then rescaling the column, and explain why the order matters.
  4. Given a comparison to be made, one gene across two samples or two genes within one sample, determine which normalization it requires.
  5. Given two transcripts of different length carrying the same count in one sample, determine which of them TPM ranks higher.
  6. Given an unfamiliar claim drawn from a count matrix, such as the most highly expressed gene named from raw counts or a fold change quoted with no denominator stated, judge whether the evidence supports it and propose what would settle it.

Distinguish the reasons a read can be compatible with more than one feature, and say which of them mapping could have settled.

  1. Explain why a read in an exon that two isoforms share fixes the gene it came from but not the isoform.
  2. Explain why duplication leaves paralogs sharing enough sequence that a 100-nucleotide read from a shared stretch fits more than one gene.
  3. Relate the row range that backward search returned in L1.6 to the counting problem, where one block of positions now raises the question of which gene is credited.
  4. Distinguish a read that aligns equally well to two places in the genome from a read that aligns to one place two features share, and state which of the two is a mapping problem.
  5. Given a described read, classify its ambiguity as multi-mapping, paralog sharing, or isoform sharing.
  6. Given a question posed at gene level or at isoform level, determine whether summing the reads over the whole gene removes the ambiguity or leaves it in place.

Account for how a set of aligned reads becomes one column of a count matrix, and determine the counts a stated annotation produces.

  1. State the definition of a count, \(c_{g,s} = \#\{r \in s : r \subseteq g\}\), and say in words what \(r \subseteq g\) requires of a read.
  2. State what a GTF annotation supplies that the reference sequence does not.
  3. Explain why an alignment gives a chromosome and a coordinate but no gene, so that counting starts where mapping stops.
  4. Explain why two labs counting the same alignment file with different annotations end up with different matrices, given that the annotation comes from a separate project rather than from the experiment.
  5. Explain why counting is interval bookkeeping rather than sequence comparison, and why it needs neither an alignment matrix nor an index.
  6. Given a read that falls inside an intron or spans the gap between two genes, determine whether it can be credited to a gene, and name what has to be decided before it can be.
  7. Given gene intervals and a set of aligned read intervals, determine the count for each gene.
Back to top