Explain omics integration (multi-omics) — Biology notes

WHAT is multi-omics integration?

The central dogma gives the natural ordering of layers:

$\text{DNA} \xrightarrow{\text{transcription}} \text{RNA} \xrightarrow{\text{translation}} \text{Protein} \xrightarrow{\text{catalysis}} \text{Metabolite}$

Each arrow is regulated, so information is lost and reshaped at every step — which is exactly why one layer cannot predict the next perfectly.

WHY do we need it? (the core justification)

Steel-manning the naïve view: "Just sequence the genome — it's the master blueprint." Feels right because DNA causes everything downstream. But the genome is nearly identical in every cell of your body, yet a neuron ≠ a muscle cell. The differences live in which genes are expressed and how proteins are modified — dynamic layers the genome can't show.

HOW is it done? (deriving an integration score from scratch)

We want a single number that says: "do two layers agree that feature $X$ matters?" Let's build it.

Step 1 — Standardise each layer. Layer $A$ (say mRNA) and layer $B$ (protein) have different units. Convert each measurement to a z-score so they're comparable:

$z_i = \frac{x_i - \mu}{\sigma}$

Why this step? Because you can't add "counts of RNA" to "μg of protein" directly — z-scoring makes both unitless "how unusual is this value" numbers. Crucially, proper z-scores have mean 0 and variance 1, which is what keeps the final correlation inside $[-1,1]$ .

Step 2 — Measure agreement across $n$ samples. For a gene measured in both layers, the Pearson correlation is the averaged product of z-scores:

$r_{AB} = \frac{1}{n-1}\sum_{i=1}^{n} z^A_i\, z^B_i$

Why this step? If mRNA and protein rise and fall together across samples, their z-scores share sign, products are positive, and $r \to +1$ : strong cross-omics support.

Step 3 — Find shared axes of variation (why matrices). Stack layers into one big data matrix and find directions ("latent factors") that vary jointly. A Singular Value Decomposition does this:

$M = U \Sigma V^{\mathsf T}$

Why this step? The columns of $V$ tell you which combination of genes/proteins moves together. The largest singular value $\sigma_1$ is the strongest shared signal across omics — this is the mathematical heart of tools like MOFA and iCluster.

Figure — Explain omics integration (multi-omics)

Two strategies of integration

Worked examples

Worked example Example 2 — computing cross-omics correlation

Three patients. Raw mRNA co-varies exactly with protein, so after proper z-scoring (mean 0, variance 1) both layers become the same standardised vector $(+1, 0, -1)$ — check: mean $=0$ , sample variance $=\tfrac{1}{2}[1+0+1]=1$ . ✓ $r = \tfrac{1}{2}\big[(1)(1) + (0)(0) + (-1)(-1)\big] = \tfrac{1}{2}(1 + 0 + 1) = 1.0$ Why this step? $r = +1$ means mRNA and protein co-vary perfectly here → this gene's regulation is mostly transcriptional, so RNA is a good proxy (unlike Example 1). This also demonstrates the ceiling: with valid unit-variance z-scores, $r$ can never exceed 1 — if you ever compute $r>1$ , your "z-scores" weren't real z-scores.

Common mistakes

Forecast-then-Verify

Flashcards

What are the four canonical omics layers in central-dogma order?

Genome (DNA) → Transcriptome (RNA) → Proteome (protein) → Metabolome (metabolites).

Why isn't transcriptomics alone enough to know protein levels?

mRNA–protein correlation is only ~0.4, so ~84% of protein variance comes from translation/degradation/modification not captured by RNA.

Define multi-omics integration.

Computationally combining ≥2 omics layers from the same samples to reveal patterns invisible in any single layer.

Formula for cross-omics agreement between two z-scored layers?

r_{AB}=\frac{1}{n-1}\sum z^A_i z^B_i

, ranging −1 to 1.

Why must standardised vectors have unit variance before computing r?

Otherwise

\frac{1}{n-1}\sum z^A_i z^B_i

can exceed 1, giving an impossible correlation; unit variance keeps r in [−1,1].

Difference between early and late integration?

Early = concatenate raw layers then analyse once; Late = analyse each layer separately then combine results.

What must you do to each layer before concatenating (early integration)?

Standardise/z-score (mean 0, variance 1) so no layer dominates by scale or feature count.

What does the top singular vector of stacked z-scored omics represent?

The strongest axis of variation shared across the layers (a latent factor).

A gene has high mRNA but low protein — biological interpretation?

Post-transcriptional regulation: reduced translation or rapid protein degradation.

Fraction of variance NOT shared when r=0.4?

1-r^2=0.84

(84%).

Why can't the genome alone explain cell-type differences?

The genome is ~identical across cells; differences arise from which genes are expressed and how proteins are modified (dynamic layers).

Recall Feynman: explain to a 12-year-old

Imagine a school play. The script (genome) is the same for every performance. But each night the actors say different lines loud or soft (RNA), some actors get sick and are replaced (proteins), and the audience reacts differently (metabolites). If you only read the script, you'd never know last night's show flopped. Multi-omics is watching all of it — script, actors, and audience — so you finally understand what really happened on stage.

Connections

Central Dogma of Molecular Biology — the ordering of the omics layers
Gene Expression Regulation — why RNA ≠ protein
Principal Component Analysis & Singular Value Decomposition — the math of shared latent factors
Pearson Correlation — the cross-layer agreement metric
Precision Medicine — clinical payoff: multi-omics tumour subtyping
Systems Biology — the parent framework treating the cell as an integrated network

Concept Map

Hinglish (regional understanding)

Intuition Hinglish mein samjho

Dekho, ek cell ko ek factory socho. Uska blueprint hai genome (DNA), work orders hain transcriptome (mRNA), machines hain proteome (proteins), aur final maal hai metabolome (metabolites). Agar tum sirf ek hi layer dekhoge — maan lo sirf RNA — toh poori kahani samajh nahi aayegi, kyunki har step par regulation hota hai aur information badal jaati hai. Isiliye multi-omics integration karte hain: kai layers ko ek saath, same samples se, jodkar dekhte hain.

Sabse important baat: mRNA aur protein ka correlation sirf lagbhag $r=0.4$ hota hai. Matlab $1-r^2 = 0.84$ , yaani 84% protein variation RNA se explain hi nahi hota! Woh chhupa hua hissa translation rate, protein degradation aur modification me hai — jo tumhe alag se measure karna padta hai. Yahi reason hai ki sirf genome sequence karke kaam nahi chalta.

Math simple hai: pehle har layer ko z-score me convert karo — matlab mean 0 aur variance 1. Ye variance = 1 wali baat important hai, kyunki isi se correlation $[-1,1]$ ke andar rehta hai. Agar tumne sirf mean minus kiya lekin standard deviation se divide nahi kiya, toh galti se $r > 1$ aa sakta hai — jo impossible hai! Phir do layers ka product average karke correlation nikaalo. Aur agar sab layers ko ek bade matrix me daalo, toh SVD/PCA se "shared factors" mil jaate hain. Yaad rakho: "Scale karo, phir SVD karo."

Kyun zaroori hai? Cancer me do tumour bilkul same gene-expression dikha sakte hain, lekin methylation aur metabolites alag — integration se pata chalta hai ki ye do alag subtypes hain jinka ilaaj bhi alag. Yehi precision medicine ka base hai. Bas do galti mat karna: correlation ka matlab causation nahi hota, aur samples matched hone chahiye warna sab bekaar.

Test yourself — Systems Biology & Frontiers

Connections