Size predicts function — a short RNA can only base-pair with a target; a kb-scale RNA can fold into a scaffold.

The non-coding RNA slide cites nobelprize.org/prizes/medicine/2024 — the microRNA discovery. That citation is the lecturer's own framing of why ncRNAs matter.

By number of molecules, rRNA dominates — which is why library preparation exists at all.

Splice-aware aligners are needed. Mature mRNA has had its introns removed, so a read spanning an exon–exon junction is contiguous in the transcript but thousands of bases apart in the genome — the aligner must split the read into blocks. Named tool STAR (also HISAT2, BOWTIE2, Kallisto); assembly is reference-guided (TopHat + Cufflinks) or de novo (Trinity).

Suppose one sample gives these counts for three genes. Nothing has been normalised yet:

Gene A looks 2.4× "bigger" than B, but that is misleading twice over:

So there are two corrections to make. The only question is which one you apply first — and that single choice is the entire difference between RPKM/FPKM and TPM.

Reads Per Kilobase per Million mapped reads. Two steps, in this order:

1. Divide the count by the gene's length in kb → this removes the length advantage.
2. Divide by (total mapped reads ÷ 1,000,000) → this removes the depth advantage.

The ranking changed. By raw reads A beat B by 2.4×; by RPKM B beats A. Raw counts had it backwards, and that reversal is the whole point of the metric.

FPKM is the same formula for paired-end data, where both mates come from one physical fragment. Counting fragments instead of reads matters because a fragment is the real molecule; counting its two reads separately would inflate long fragments. On single-end data RPKM and FPKM are numerically identical.

Transcripts Per Million. Same two corrections, opposite order:

1. Divide each count by its length in kb.
2. Add up those length-corrected values, then express each as a share of that total × 1,000,000.

Note what TPM did not change: B still beats A. Within one sample the two metrics rank genes the same way, because both make the same two corrections. They differ only in what the numbers mean.

Fold change is just a ratio: expression in condition 2 divided by condition 1. The “log2” is only there to make the number symmetric.

On a plain ratio scale, "up 8-fold" is 8 and "down 8-fold" is 0.125 — not symmetric, and impossible to average sensibly. So log2 turns multiplication into addition: a change of +3 is the same size whether it went up or down, and doubling twice is +1 +1 = +2. Every fold change you read on a volcano plot is this scale.

Take one sample and plot the expression value of every gene. The raw values are wildly skewed: a small number of genes sit at very high values while the mass of genes sits near the bottom, so a linear-axis histogram is one tall spike with a long tail and is hard to read or reason about.

Now take the logarithm of each value and plot that. The same data becomes roughly symmetric and bell-shaped. That is what “log-normal” means: the values themselves follow a normal distribution once you take their log — a normal shape that has been exponentiated.

Read results off a volcano plot: log2 fold change (x) against −log10 p-value (y). Standard tool DESeq2.

Thousands of genes are tested at once, so raw p-values produce false positives and a multiple-testing correction is required. Strictest to most permissive: Bonferroni → Bonferroni step-down → Westfall and Young permutation → Benjamini and Hochberg FDR → None.

• GSEA tests whether a set of genes is coordinately shifted rather than testing genes one at a time: genes are ranked by correlation with the phenotype, a random walk along that ranking gives an enrichment score ES(S), the maximum deviation from zero is the score, and the leading edge subset is the part driving it. Gene sets come from MSigDB.
• Clustering / dimensionality reduction — PCA first (PC1 = most variance, then PC2, PC3), then non-linear methods. Worked example: medulloblastoma sub-grouping.
• Co-expression groups correlated genes (e.g. r = 0.7) into modules ME1–ME6.
• Regulatory networks link transcription factors (CEBP, STAT3, BHLHE40, FOSL2, RUNX1) to targets — the slide carries a literal “Target?” arrow, because inferring the target is the open problem.
• Deconvolution — CIBERSORT/CIBERSORTx infers cell-type composition from bulk. This is the bridge to single cell: if composition is the answer, bulk is not enough.
• Alternative splicing — the same data reveals splicing patterns, which can create a peptide–MHC neoantigen the genome alone would never predict.
• Integrative analysis — WES + RNA-Seq (FPKM ≥ 1) → NetMHCpan → candidate neoantigen: genomics supplies the mutation, transcriptomics proves it is expressed.

By therapeutic interest the cell is the small and long non-coding classes. lncRNA mechanisms shown: HOTAIR (PRC2/LSD1 → H3K27me), BACE1-AS, HULC (sponges miR-372), NEAT1 (paraspeckles), NRON (NFAT) — with a commercial pipeline slide from NextRNA. microRNA drugs and circular RNA are developed as products, and a 26-row table lists approved RNA medicines (see Detail Pass).

Why single cell? Bulk averages millions of cells, so a rare subpopulation vanishes into the mean. Single cell resolves the mean into its components.

Because capture per cell is tiny, a zero is ambiguous — absence of evidence is not evidence of absence. The three responses, ranked by how much they assume:

1. Ignore zero inflation — assume most zeros are real biology.
2. Reduce dimensions — the most widely used; sidestep the zeros rather than imputing.
3. Impute — e.g. MAGIC (van Dijk et al, Cell 2018), smoothing along a cell-similarity graph. Powerful, but it invents values and can manufacture structure that was not there.

Pipeline: dimensionality reduction → clustering → marker genes → annotation → downstream biology. Reduction: linear (PCA, ICA), non-linear (t-SNE, UMAP), model-based (ZIFA), deep learning (VAE, scvis, DCA), ensemble (SIMLR). PCA is a linear algebraic method and its stated limitation is that PCs are linear combinations, so it fails to find non-linear structure. Clustering is k-means, hierarchical or graph-based (node = cell, edge = similarity), with Louvain and Leiden.

Single cell's central limitation: dissociating the tissue destroys where each cell was. Spatial transcriptomics restores it — imaging-based methods (IMC, MERSCOPE, Xenium, CosMx) read targeted panels in place; sequencing-based methods (Slide-seqV2, Stereo-seq, Visium HD) capture transcriptome-wide. Analysis splits into A–G (Detail Pass); newest work is multi-omic (Nature 2024).

MSigDB collections and their sizes (the GSEA gene-set database):

Sets download as GMT files, keyed by gene symbol or NCBI Entrez ID. Pipeline on the slide: Molecular Profile Data + Gene Set Database → run GSEA → enriched sets.

Approved RNA medicines — the table's own columns are originator, name, site of action, platform modality/MOA and approval year. Representative rows:

The full table has 26 rows and also covers Ionis ASOs, aptamers, mRNA vaccines and Sarepta Duchenne ASOs. Note the pattern: most siRNA products target hepatocytes, because that is where GalNAc-conjugated delivery accumulates.

Single-cell platform families and what each trades away (scored + to +++ on the slide):

No platform is best on all five. Plate-based wins flexibility and depth; droplet wins simplicity and adoption; split/pool wins throughput.

Marker gene to cell type — the Louvain-group table that turns clusters into biology:

Marker-test methods: non-parametric (Wilcoxon rank sum, Student's t-test); scRNA-specific MAST (treats cellular detection rate as a covariate); bulk method DESeq2. Annotation tools are benchmarked across datasets (Pancreas, CellBench, Allen Mouse Brain, PBMC and others) — SVM_rejection, scPred and SVM are shown leading that benchmark.

Spatial platforms split by how they read the tissue:

Spatial analysis categories A–G: A Image Analysis · B Cell Type Identification · C Deconvolution · D Spatial Distribution · E Cell–Cell Interaction · F Spatial Expression Patterns · G Subcellular.

Stereo-seq is worth one line of detail: a DNB patterned chip up to 174.24 cm² with 220 nm centres, giving spatially resolved transcriptomics at cellular resolution. It produced MOSTA (Mouse Organogenesis Spatiotemporal Transcriptomic Atlas) — 8 stages, 53 sections, E9.5 to E16.5 — and the companion tool Spateo for 3D reconstruction and spatiotemporal modelling.

The consequence to state in an answer: raw read counts are not expression. De-duplicating by UMI is what converts reads into molecule counts, which is what quantification should be based on.

Design consequence: scRNA and snRNA are not interchangeable datasets. A dissociation step biases scRNA against fragile cell types, which is precisely why snRNA exists. Comparing the two without accounting for that is a common analysis error.

Two consequences worth carrying: (1) any bias in reverse transcription (secondary structure, GC content, primer efficiency) is inherited by the final counts — it is a library-prep artefact, not biology. (2) Because the message is copied before amplification, the UMI must be attached before PCR (see above) for de-duplication to be valid.

Recall the two enrichment choices from the main notes — poly(A) selection (mature mRNA) versus rRNA depletion (keeps non-coding and degraded RNA) — those decide which RNA molecules reach the reverse transcription step in the first place.

Worked example from the notes: NY-ESO-1, a cancer/testis antigen. Its attractiveness as an immunotherapy target rests on expression being confined to tumour and to immune-privileged testis — and that claim is verified by checking GTEx for absence in normal somatic tissues.

This is the experimental form of the L6 target criterion “tissue distribution: primarily in diseased tissue and not healthy tissue”. GTEx is simply where you go to check it. An antigen expressed in myocardium or liver is an on-target off-tumour toxicity risk regardless of how tumour-specific it looks in a cancer cohort.