Transcriptomics: RNA-seq and differential expression

The genome is largely the same in every cell of an organism, but which genes are transcribed, and how much, differs between tissues, developmental stages, conditions and individual cells. Measuring the transcriptome, the set of RNA molecules present, is therefore one of the most informative experiments in biology. RNA sequencing (RNA-seq) has largely replaced microarrays for this purpose. BIOL-207 introduced the idea of normalizing read counts; this lesson covers the full workflow, the statistics of count data, the testing of thousands of genes at once, and single-cell transcriptomics.

Designing an RNA-seq experiment

An RNA-seq experiment begins long before sequencing. The biological question determines the comparison, for example drought-stressed versus watered plants, or tumour versus normal tissue, and the design must allow that comparison to be separated from everything else. The single most important design decision is the number of biological replicates: independent samples from different individuals or cultures, as opposed to technical replicates, repeated measurements of one sample. Biological variation between individuals is usually much larger than technical noise, and without replicates it cannot be estimated, so no valid test of differential expression is possible. Three replicates per group is a common minimum; studies of subtle effects or variable material such as human tissue need many more.

Confounding is the other great danger. If all control samples are extracted on Monday and all treated samples on Friday, or sequenced on different runs, any difference may reflect the batch rather than the treatment, and no analysis can separate them. Samples should be randomized across processing batches, and batches recorded so they can be modelled. Library preparation choices also matter. Messenger RNA can be enriched by capturing poly(A) tails, which gives clean mRNA profiles but loses non-polyadenylated RNAs and requires intact RNA; ribosomal RNA, typically 80 per cent or more of total RNA, can instead be depleted, which retains noncoding and degraded RNA. Stranded protocols preserve which DNA strand was transcribed, resolving overlapping antisense genes.

From reads to counts

Reads from mature mRNA span exon–exon junctions, so mapping them to the genome needs a splice-aware aligner that can split a read across an intron, written in the CIGAR string with the N operation (lesson 6). STAR and HISAT2 are widely used; they discover novel junctions as well as known ones. The number of reads overlapping each gene's exons is then counted. An alternative is pseudoalignment, introduced by kallisto in 2016 and also used by salmon: instead of computing base-level alignments, the read's k-mers are matched against an index of the transcriptome to determine which transcripts it is compatible with. Because many reads are compatible with several isoforms of a gene, an expectation-maximization algorithm apportions them, estimating the abundance of each transcript. Pseudoalignment is very fast and accurate for quantifying annotated transcripts, while genome alignment is needed to discover new genes, junctions or fusion transcripts.

The result is a count matrix: genes or transcripts in rows, samples in columns, each entry the number of reads assigned. Counts are not directly comparable between samples or between genes. A sample sequenced to 40 million reads has roughly twice the counts of one sequenced to 20 million, and a long gene yields more fragments than a short gene expressed at the same molar level. Normalization addresses both.

Normalization

Counts per million (CPM) divide each count by the sample's total count and multiply by one million, correcting for sequencing depth but not gene length. RPKM and FPKM (reads or fragments per kilobase per million) also divide by gene length, but have the defect that the values in different samples sum to different totals. Transcripts per million (TPM) reverse the order: first divide each count by gene length in kilobases, giving reads per kilobase, then scale so that each sample sums to one million. TPM values can be read as the relative molar proportion of each transcript in the sample, and are comparable within a sample across genes.

Neither CPM nor TPM handles a subtler problem, composition bias. If one condition massively expresses a few genes, those genes take a larger share of the reads, and every other gene's share falls, even though its absolute expression is unchanged. Methods for differential expression therefore estimate size factors that assume most genes do not change. DESeq2's median-of-ratios method computes, for each gene, the geometric mean of its counts across all samples as a pseudo-reference; for each sample, it takes the ratio of each gene's count to that reference, and uses the median ratio as the sample's size factor. The median is unaffected by a minority of strongly changed genes. The TMM method of edgeR uses a trimmed mean of log ratios for the same purpose.

# CPM, TPM and DESeq-style median-of-ratios size factors on an invented count matrix.
import numpy as np
genes = ['geneA', 'geneB', 'geneC', 'geneD', 'geneE', 'geneF']
length_kb = np.array([1.0, 4.0, 2.0, 0.5, 3.0, 1.5])
# columns: control_1, control_2, treated_1, treated_2 (treated_2 sequenced twice as deeply)
counts = np.array([
    [500,  520,   510,  1010],
    [2000, 2100, 2050,  4150],
    [800,   780, 4000,  8100],      # strongly induced by treatment
    [100,   110,  105,   205],
    [1500, 1450, 1480,  3000],
    [600,   640,  610,  1230]], dtype=float)
cpm = counts / counts.sum(axis=0) * 1e6
rpk = counts / length_kb[:, None]
tpm = rpk / rpk.sum(axis=0) * 1e6
log_geo_mean = np.log(counts).mean(axis=1)
size = np.exp(np.median(np.log(counts) - log_geo_mean[:, None], axis=0))
norm = counts / size
np.set_printoptions(suppress=True)
print('library sizes:', counts.sum(axis=0).astype(int))
print('size factors: ', np.round(size, 2))
print(f"{'gene':6s} {'CPM c1':>8s} {'CPM t2':>8s} {'TPM c1':>8s} {'norm c1':>8s} {'norm t1':>8s} {'norm t2':>8s}")
for g, c, t, n in zip(genes, cpm, tpm, norm):
    print(f"{g:6s} {c[0]:8.0f} {c[3]:8.0f} {t[0]:8.0f} {n[0]:8.0f} {n[2]:8.0f} {n[3]:8.0f}")

library sizes: [ 5500  5600  8755 17695]
size factors:  [0.82 0.86 0.84 1.69]
gene     CPM c1   CPM t2   TPM c1  norm c1  norm t1  norm t2
geneA     90909    57078   200000      611      606      598
geneB    363636   234530   200000     2444     2437     2458
geneC    145455   457756   160000      977     4755     4798
geneD     18182    11585    80000      122      125      121
geneE    272727   169539   200000     1833     1759     1777
geneF    109091    69511   160000      733      725      729

Look at gene A, which the invented data were built to have unchanged expression. Its CPM falls in the treated samples, because induced gene C takes a larger share of the reads, which would wrongly suggest down-regulation. After median-of-ratios normalization, gene A's values are nearly equal across all four samples, including the deeply sequenced one, while gene C stands out with roughly fivefold induction. TPM, meanwhile, changes the comparison of genes within a sample: in control 1, long gene B has four times as many reads as gene A, but the same TPM, because it is four times longer, so the two transcripts are present in similar molar amounts.

Statistics of counts

If the same library were sequenced repeatedly, counts for a gene would vary approximately as a Poisson distribution, whose variance equals its mean. Between biological replicates, however, the true expression level itself varies, and the counts are overdispersed: their variance exceeds the mean, increasingly so for highly expressed genes. The negative binomial distribution captures this with a dispersion parameter \(\alpha\):

With only three replicates per group, estimating a separate dispersion for each gene is unreliable, so DESeq2 and edgeR share information across genes: they fit a trend of dispersion against mean expression and shrink each gene's estimate toward it. They then fit a generalized linear model for each gene, with the condition, and any batch or other covariates, as predictors, and test whether the coefficient for condition differs from zero. The limma-voom approach instead transforms counts to log-CPM with precision weights and uses linear models with moderated t statistics. These methods are well tested and generally agree on strong effects; they differ at the margins.

Testing thousands of genes

A typical experiment tests 15,000 to 20,000 genes. At a significance threshold of 0.05, even if no gene changed, about 1,000 would appear significant by chance. Controlling the family-wise error rate with the Bonferroni correction (dividing the threshold by the number of tests) is too strict for discovery. Instead, the false discovery rate (FDR), the expected proportion of false positives among the genes declared significant, is controlled with the procedure of Yoav Benjamini and Yosef Hochberg (1995): sort the \(m\) p-values, find the largest rank \(i\) with \(p_{(i)} \le (i/m)q\), and declare the first \(i\) genes significant. Equivalently, each gene gets an adjusted p-value, often called a q-value or padj, and genes with padj below 0.05 or 0.1 are reported.

# Simulate negative binomial counts, test each gene on log-scale normalized values, and control the FDR.
import numpy as np
from scipy import stats
rng = np.random.default_rng(410)
n_genes, n_rep = 4000, 4
base = rng.lognormal(mean=5, sigma=1.2, size=n_genes)
true_lfc = np.zeros(n_genes)
changed = rng.choice(n_genes, 300, replace=False)
true_lfc[changed] = rng.choice([-1, 1], 300) * rng.uniform(1, 2.5, 300)
alpha = 0.05 + 1.0 / base
def nb(mu):
    n = 1 / alpha[:, None]
    return rng.negative_binomial(n, n / (n + mu))
ctrl = nb(np.repeat(base[:, None], n_rep, axis=1))
trt = nb(np.repeat((base * 2 ** true_lfc)[:, None], n_rep, axis=1))
counts = np.hstack([ctrl, trt]).astype(float)
keep = counts.mean(axis=1) >= 10                          # filter very low counts
logc = np.log2(counts / counts.sum(axis=0) * 1e6 + 1)
t, p = stats.ttest_ind(logc[:, n_rep:], logc[:, :n_rep], axis=1)
p = np.where(keep, p, 1.0)
order = np.argsort(p)
m = keep.sum()
ranked = p[order] * m / np.arange(1, len(p) + 1)
padj = np.empty_like(p)
padj[order] = np.minimum(1, np.minimum.accumulate(ranked[::-1])[::-1])
sig = padj < 0.05
is_true = true_lfc != 0
print(f"genes tested {m}, truly changed {is_true.sum()}")
print(f"raw p < 0.05: {(p < 0.05).sum()} genes, of which false {((p < 0.05) & ~is_true).sum()}")
print(f"BH padj < 0.05: {sig.sum()} genes, of which false {(sig & ~is_true).sum()} "
      f"(observed FDR {(sig & ~is_true).sum() / max(sig.sum(), 1):.1%}), recovered {(sig & is_true).sum()} of {is_true.sum()}")

genes tested 3934, truly changed 300
raw p < 0.05: 520 genes, of which false 233
BH padj < 0.05: 245 genes, of which false 15 (observed FDR 6.1%), recovered 230 of 300

Without correction, the list of genes with p below 0.05 contains more than two hundred false positives, nearly half the list; with the Benjamini–Hochberg adjustment, the observed false discovery proportion is close to the nominal 5 per cent, at the cost of missing some true changes, mostly those with small fold changes in low-count genes. (The example tests log-scale values with t tests for simplicity; with real data, a negative binomial method such as DESeq2 would be preferred, especially with few replicates.)

Results are commonly displayed as a volcano plot, with the log2 fold change on the horizontal axis and the negative log10 of the p-value on the vertical axis, so that genes with large and significant changes lie in the upper corners. Fold changes estimated from few, noisy counts are inflated, so modern methods shrink log fold changes toward zero when evidence is weak, which makes ranking by effect size more reliable.

Interpreting gene lists

A list of several hundred differentially expressed genes is hard to interpret one gene at a time. Over-representation analysis asks whether genes annotated to a Gene Ontology term or pathway (lesson 8) appear in the list more often than expected by chance, using the hypergeometric distribution or Fisher's exact test, with the background set to all genes that were tested, not the whole genome. Gene set enrichment analysis (GSEA) instead ranks all genes by their change and asks whether a set's members cluster toward the top or bottom of the ranking, which detects coordinated modest changes. Enrichment results depend on annotation quality and on many correlated tests, so they suggest hypotheses rather than prove mechanisms.

Single-cell RNA-seq

Bulk RNA-seq measures the average of millions of cells, hiding differences between cell types and states. Single-cell RNA-seq (scRNA-seq) captures individual cells, most commonly in droplet systems that encapsulate each cell with a bead carrying oligonucleotides that include a cell barcode, a unique molecular identifier (UMI) and a poly(T) capture sequence. After sequencing, reads are assigned to cells by barcode, and UMIs are counted to give molecule counts per gene per cell. Tens of thousands of cells can be profiled in one experiment, but each cell yields only a few thousand molecules, so the data are sparse, with most entries zero.

Analysis follows a standard path. Quality control removes empty droplets, doublets (two cells in one droplet) and damaged cells, often identified by a high fraction of mitochondrial transcripts. Counts are normalized and log-transformed, highly variable genes are selected, and principal component analysis reduces the data to tens of dimensions. Cells are clustered, usually by building a graph of nearest neighbours in principal component space and applying a community detection algorithm such as Leiden, and visualized in two dimensions with UMAP or t-SNE. Clusters are labelled as cell types using marker genes. Two-dimensional embeddings distort distances, so cluster positions and sizes on a UMAP plot should not be over-interpreted. Large efforts such as the Human Cell Atlas are building reference maps of all cell types in the human body.

# Synthetic single-cell counts for three cell types: normalize, reduce with PCA and cluster with k-means.
import numpy as np
rng = np.random.default_rng(3)
n_genes, sizes = 300, (150, 100, 60)
base = rng.gamma(0.6, 1.5, n_genes)
programs = []
for t in range(3):
    prog = base.copy()
    markers = rng.choice(n_genes, 20, replace=False)
    prog[markers] *= rng.uniform(4, 10, 20)
    programs.append(prog)
labels = np.repeat([0, 1, 2], sizes)
depth = rng.lognormal(0, 0.4, len(labels))
X = rng.poisson(np.array([programs[t] for t in labels]) * depth[:, None])
print(f"cells {X.shape[0]}, genes {X.shape[1]}, zero entries {np.mean(X == 0):.0%}")
Xn = np.log1p(X / X.sum(axis=1, keepdims=True) * 1e4)
Xc = Xn - Xn.mean(axis=0)
U, S, Vt = np.linalg.svd(Xc, full_matrices=False)
pcs = U[:, :10] * S[:10]
print('variance explained by PC1-PC3:', np.round(S[:3] ** 2 / (S ** 2).sum(), 3))
best = None
for restart in range(5):
    cent = pcs[rng.choice(len(pcs), 3, replace=False)]
    for _ in range(50):
        assign = np.argmin(((pcs[:, None, :] - cent[None]) ** 2).sum(-1), axis=1)
        cent = np.array([pcs[assign == j].mean(0) if np.any(assign == j) else cent[j] for j in range(3)])
    inertia = ((pcs - cent[assign]) ** 2).sum()
    if best is None or inertia < best[0]:
        best = (inertia, assign)
assign = best[1]
for j in range(3):
    members = labels[assign == j]
    print(f"cluster {j}: {len(members)} cells, composition by true type {np.bincount(members, minlength=3)}")

cells 310, genes 300, zero entries 56%
variance explained by PC1-PC3: [0.065 0.037 0.035]
cluster 0: 60 cells, composition by true type [ 0  0 60]
cluster 1: 100 cells, composition by true type [  0 100   0]
cluster 2: 150 cells, composition by true type [150   0   0]

Despite the sparsity, with most entries zero, the clusters recover the three cell types almost perfectly, because each type's marker genes shift its cells consistently in principal component space. Real data add continuous states, rare cell types, batch effects between samples and doublets, which make clustering and annotation more demanding.