Why Cell Composition Matters
A bulk chromatin assay measures material collected from many cells together. If two samples contain different proportions of cell types, their aggregate accessibility or histone-mark signal can differ even when the chromatin state of every cell type is unchanged. Conversely, a real change within a cell type can be obscured or reversed by a shift in the mixture. A bulk difference is therefore not automatically evidence of a cell-intrinsic regulatory change.
This issue is distinct from sequencing and normalization biases. Even when libraries pass quality control and signals are normalized appropriately, a sample with more of a cell type that naturally has strong signal at a locus may show higher bulk signal there. The first task is to define a plausible mixture model, then ask whether the observed signal and sample composition support a composition-based explanation.
A Mixture Model for Bulk Signal
This approximation is most directly plausible when cell-type contributions are measured on a comparable scale and the bulk signal is proportional to the contributions from the cell types. For raw read counts, the contributions add as counts, but their expected amounts can also depend on library size, assay efficiency, and cell-specific recovery. For normalized enrichment or accessibility scores, the weighted-average interpretation is an approximation, not a universal identity. The relevant \(\pi_{ik}\) should represent the material that contributes to the assay; a fraction of cells may not equal the fraction of nuclei, DNA, or recoverable chromatin.
Consider two cell types, A and B, and two conditions, control and case. Write \(p_g\) for the proportion of type A in group \(g\), and \(a_g\) and \(b_g\) for the corresponding cell-type-specific signals. The group mean under the mixture model is \(M_g=p_ga_g+(1-p_g)b_g\). A difference in \(M_g\) can arise because \(p_g\) changes, because \(a_g\) or \(b_g\) changes, or because both happen.
Separating Composition and Within-Cell Changes
One useful decomposition compares the case group with the control group while using control cell-type signals as a reference for the composition component:
To verify the identity, start with \(M_1-M_0=p_1a_1+(1-p_1)b_1-p_0a_0-(1-p_0)b_0\). Add and subtract \(p_1a_0+(1-p_1)b_0\). The resulting first part is \(p_1a_0+(1-p_1)b_0-p_0a_0-(1-p_0)b_0=(p_1-p_0)(a_0-b_0)\); the remaining part is \(p_1(a_1-a_0)+(1-p_1)(b_1-b_0)\). Thus the observed difference equals a composition term plus changes in cell-type-specific signals, weighted by the case composition. Other valid decompositions use a different reference group; the allocation between terms can depend on that choice, but the total difference does not.
The decomposition is a way to organize interpretation, not a way to identify the terms from a bulk difference alone. Estimating cell-specific signals requires additional evidence, such as sorted-cell assays, reliable cell-type markers, or a design with informative variation in composition. Sample metadata can reveal that a composition explanation is plausible, but metadata alone do not prove it.
Worked Example: A Bulk Difference With No Cell-Intrinsic Change
Suppose a hypothetical locus has signal 10 normalized units in cell type A and 2 normalized units in cell type B in both groups. Type A makes up 80% of control material and 30% of case material. Under the two-type mixture model, calculate the group signals and their difference.
For control, \(M_0=0.80(10)+0.20(2)=8.00+0.40=8.40\) normalized units. For case, \(M_1=0.30(10)+0.70(2)=3.00+1.40=4.40\) normalized units. The case-minus-control difference is \(4.40-8.40=-4.00\) normalized units.
Neither cell-type-specific signal changed. The entire difference follows from the lower proportion of type A, which has the stronger signal at this locus. This calculation shows how a bulk comparison can look like a decrease without a within-cell-type decrease; it does not establish that composition caused any particular real bulk result.
Worked Example: A Composition Shift Reverses the Bulk Comparison
In another hypothetical example, control signals are 9 units in type A and 3 units in type B, while case signals are 10 and 4 units, respectively. Type A comprises 70% of control material and 40% of case material. Calculate the bulk difference, then compare the groups at a common type A proportion of 55%.
The bulk control signal is \(0.70(9)+0.30(3)=6.30+0.90=7.20\) units. The bulk case signal is \(0.40(10)+0.60(4)=4.00+2.40=6.40\) units, so case minus control is \(-0.80\) units. At a common proportion \(p=0.55\), the standardized control signal is \(0.55(9)+0.45(3)=4.95+1.35=6.30\) units. The standardized case signal is \(0.55(10)+0.45(4)=5.50+1.80=7.30\) units. The standardized difference is \(7.30-6.30=1.00\) unit.
The bulk comparison is lower in cases, whereas the comparison at the common mixture is higher. Here both cell types have a one-unit increase, but the case group has less of the higher-signal type A. Standardization answers a specified comparison at a common composition; it does not make the observed bulk difference disappear or establish that the standardized contrast is the only biologically relevant effect.
Using Metadata and Adjusted Comparisons
Potential composition information includes measured cell counts, tissue pathology, sorting results, known sample-processing details, and validated marker-based estimates. Examine whether proportions differ by condition and whether those differences are large enough, given cell-type signal contrasts, to plausibly affect the locus-level result. Cell-specific evidence is especially useful: if the bulk signal changes but sorted-cell signals do not, a mixture explanation gains support. If cell-specific signals change as well, composition may be only part of the explanation.
When sample-level proportions are available, an adjusted comparison can be expressed with a regression model. For a locus \(j\), a simple model is
where \(T_i\) indicates condition, and one cell type is omitted because the proportions sum to one. Under this model, \(\beta_{Tj}\) is the condition contrast at the same modeled composition, assuming the specified functional form is adequate. It is not automatically a causal effect or a cell-intrinsic effect. The model also assumes that samples provide enough independent information to distinguish condition from composition, that relevant cell types are represented, and that residual behavior is suitable for the intended inference.
Worked Example: Interpreting an Adjusted Contrast With Overlap
Suppose six hypothetical samples have measured type A proportions of 0.20, 0.40, and 0.60 in each of two conditions. At one locus, control signals are 3.6, 5.2, and 6.8 normalized units, and case signals are 4.6, 6.2, and 7.8 units at the matching proportions. Consider the model \(Y=\beta_0+\beta_TT+\gamma p+\epsilon\), with \(T=0\) for control and \(T=1\) for case.
Within either condition, increasing \(p\) from 0.20 to 0.40 increases signal by \(5.2-3.6=1.6\) units for control and \(6.2-4.6=1.6\) units for case. Since the proportion rises by 0.20, the fitted slope is \(\gamma=1.6/0.20=8.0\) units per unit of type A proportion. At each matching proportion, the case-minus-control difference is \(4.6-3.6=1.0\), \(6.2-5.2=1.0\), and \(7.8-6.8=1.0\) unit. Thus the model gives \(\beta_T=1.0\) unit, with an intercept of 2.0 units because \(3.6=2.0+8.0(0.20)\).
The conditions have overlapping proportions, and the data are consistent with a constant one-unit adjusted contrast and a linear composition association. These six values alone do not justify a population-level significance claim. The interpretation also depends on the estimated proportions being accurate, the linear and no-interaction model being reasonable, and other relevant differences between samples not explaining the contrast.
Assumptions, Diagnostics, and Limitations
Before interpreting adjustment, check whether the conditions overlap in composition. If all controls have low proportions of a cell type and all cases have high proportions, the adjusted condition contrast may rely on extrapolation: the data contain little or no direct comparison at a common mixture. Also check whether proportions are measured with sufficient accuracy, whether important cell types are missing, and whether samples with similar proportions have markedly different signal patterns. A condition-by-proportion interaction may be needed if the condition contrast varies with composition; in that case, one single adjusted contrast does not describe every mixture.
Cell proportions are compositional: they are nonnegative and sum to one. Including every cell-type proportion alongside an intercept creates exact collinearity. Use a reference type or another appropriate parameterization, and interpret coefficients relative to that specification. In addition, estimated proportions can be noisy. Measurement error may leave residual composition confounding, while adjusting for inaccurate or poorly chosen proxies can make estimates unstable.
There is an important causal distinction. If a condition changes cell composition, composition may be part of the biological pathway by which the condition changes bulk tissue. Adjusting for composition then asks about a contrast at fixed composition, not the total difference that includes composition changes. That adjusted contrast should not be called a direct causal effect without additional causal assumptions and design support. If condition assignment and composition are nearly inseparable, neither a regression coefficient nor a bulk assay alone can reliably distinguish their contributions.
A practical comparison of approaches is:
| Approach | What it can contribute | Main limitation |
|---|---|---|
| Inspect metadata and marker evidence | Identifies plausible shifts and guides follow-up | Association with condition does not prove the source of a locus-level signal |
| Adjust for measured proportions | Estimates a contrast conditional on modeled composition | Needs accurate measurements, overlap, and a suitable model |
| Compare cell-specific assays | Directly tests whether signal differs within assayed cell types | Requires suitable cell-specific material and may not represent every cell type or the original tissue mixture |
These approaches answer related but not identical questions. Cell-specific evidence can sharpen interpretation, while bulk data retain information about the aggregate tissue. Neither a disappearance of a condition coefficient after adjustment nor a significant coefficient before adjustment is, by itself, definitive evidence about mechanism.
Common Pitfall
Correction: A bulk difference can result from changed cell proportions, within-cell changes, or both. Check composition evidence and assay scale; use adjusted or cell-specific comparisons only when their assumptions and data support the intended interpretation.
Check Your Understanding
Answer each question, then compare with the solutions below.
- At a hypothetical locus, cell type A has signal 6 units and type B has signal 2 units in both conditions. Type A proportions are 0.25 in controls and 0.75 in cases. What are the bulk signals and the case-minus-control difference?
- Why can a regression with an intercept and proportions for every cell type be unidentified?
- A condition and cell proportion have no overlap across samples. What is the main limitation of an adjusted condition comparison?
- If adjustment removes a bulk condition difference, does that prove the condition has no biological effect on the tissue? Explain.
Solutions. (1) Control signal is \(0.25(6)+0.75(2)=1.50+1.50=3.00\) units. Case signal is \(0.75(6)+0.25(2)=4.50+0.50=5.00\) units, so the difference is \(5.00-3.00=2.00\) units. (2) The proportions sum to one, so including all of them with an intercept creates exact linear dependence among predictors. Omit a reference proportion or use another suitable parameterization. (3) The adjusted comparison must extrapolate to compositions not shared by the groups, so the data do not provide a well-supported within-composition contrast. (4) No. Adjustment estimates a contrast at fixed modeled composition and may remove a pathway through which condition changes tissue composition. It does not establish that the total biological effect is absent, nor that the adjusted contrast is causal.