🧬 biology

Donor-allocation effects vary with reference-cell budget in MuSiC PBMC simulations

This study demonstrates that while the accuracy penalty of unequal donor allocation in MuSiC PBMC deconvolution decreases as the reference-cell budget increases, significant variability in individual configurations and persistent discrepancies with flow cytometry in real samples indicate that the trend's transferability to measured bulk data remains unestablished.

Original authors: Yunhao Jiang

Published 2026-09-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yunhao Jiang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

To understand the human body at a microscopic level, scientists often look at a sample of blood and try to count how many different kinds of cells are present. A single drop of blood contains a complex mixture of immune cells, such as B cells, T cells, and monocytes, each playing a unique role in defending the body. When researchers analyze the genetic material from a whole drop of blood, they get a blended signal, like hearing a choir sing a single chord rather than hearing individual voices. To figure out which singers are present and how many of each, they use a mathematical method called deconvolution. This process relies on a reference library: a collection of genetic profiles from individual cells, ideally gathered from many different people, to teach the computer what each cell type sounds like on its own.

The quality of this reference library matters immensely. If the library is built from cells taken from just a few people, or if those people are represented unevenly, the final count of cells in the blood sample might be wrong. A critical question arises when building these libraries: does it matter if one person contributes many cells to the reference while another contributes very few, as long as the total number of cells remains the same? This question sits at the heart of a recent study by Yunhao Jiang, which investigated how the distribution of cells from different donors affects the accuracy of these cell counts. The research suggests that while having more total cells generally helps, the way those cells are shared among donors still leaves a mark on the results, even when the total number is quite large.

The study focused on a specific tool called MuSiC, a popular method used to estimate cell proportions from bulk genetic data. The researchers wanted to test a specific scenario: if they kept the total number of reference cells fixed, would it hurt the accuracy if they took most of those cells from one donor and very few from the others? To find out, they created thousands of synthetic mixtures using data from peripheral blood mononuclear cells, a common type of immune cell found in blood. They simulated a situation where three donors provided the reference cells for six different cell types. In some simulations, the three donors contributed equally, like a fair split of twenty cells each. In others, the split was heavily skewed, with one donor providing fifty cells while the other two provided only five each. They ran these tests at different total budgets, starting with a small group of sixty cells per cell type and increasing the number up to three hundred.

The results revealed a clear trend that depended on the size of the reference group. When the total number of reference cells was small, skewing the contribution toward one donor made the estimates significantly less accurate. The error increased noticeably when one donor dominated the reference pool. However, as the researchers increased the total number of cells in the reference library, this penalty began to fade. When the budget was raised to three hundred cells per type, the average difference in accuracy between the fair split and the skewed split became very small, almost disappearing. This suggests that having a larger pool of data can help smooth out the problems caused by an uneven distribution of donors.

Yet, the story is not as simple as "more data fixes everything." Even at the largest budget tested, where the average error was nearly zero, the researchers found that the skewed allocation still caused problems in many specific cases. In nearly half of the individual test configurations, the uneven distribution still produced higher errors than the balanced one. The average result looked good only because some cases improved while others got worse, canceling each other out. This means that for any single experiment, relying on an uneven donor mix could still lead to inaccurate results, even if the overall trend suggests it is safe. The researchers also tested whether simply throwing away extra cells from a dominant donor to make the groups equal would help. They found that discarding cells to force a balance actually made the results worse, because it reduced the total amount of information available to the computer.

To see if these findings held up in the real world, the team applied the same methods to actual blood samples from healthy adults, comparing the computer's estimates against physical counts taken by a machine that sorts cells one by one. In this real-world test, the method struggled to identify major cell types correctly, regardless of how the donors were allocated or how many cells were used. The computer failed to detect B cells and T cells in almost every sample, suggesting that the issues in real blood samples are much deeper than just how donors are grouped. The study concludes that while increasing the total number of reference cells reduces the average penalty of uneven donor allocation, it does not eliminate the risk for individual cases. Furthermore, the problems seen in simulations did not translate to a solution for the real-world failures observed in actual blood samples. The work highlights that when building reference libraries, scientists must report not just the total number of cells, but exactly how many came from each person, because the balance of contributions still matters.

Technical Summary: Donor-Allocation Effects in MuSiC PBMC Simulations

Problem Statement
Single-cell RNA sequencing (scRNA-seq) references are increasingly used to deconvolve cell-type proportions from bulk expression data using tools like MuSiC. While the total size of a reference dataset is often reported, the internal distribution of cells among donors is frequently overlooked. A reference containing 60 cells for a specific cell type across three donors could be constructed with an equal allocation (20:20:20) or a highly skewed allocation (e.g., 50:5:5). Although both scenarios utilize the same donors and total cell count, they provide different amounts of information regarding each donor's expression profile. MuSiC averages donor profiles equally during fitting, regardless of the precision of the individual donor estimates. This study investigates whether the accuracy penalty of unequal donor allocation depends on the total reference-cell budget and whether this trend holds across different reference groupings and in measured bulk samples.

Methodology
The study employed a rigorous simulation framework using raw-count synthetic mixtures derived from two peripheral blood mononuclear cell (PBMC) datasets:

Original Cohort: Systemic lupus erythematosus (SLE) donor cells (GSE96583), restricted to six donors with sufficient cells across six cell types (B cells, CD14+ monocytes, CD4/CD8 T cells, FCGR3A+ monocytes, NK cells).
External Cohort: A larger set of 14 "normal" donors from CELLxGENE/GSE174188/GSE137029.

Experimental Design:

Reference Construction: For each held-out target donor, references were built from subsets of remaining donors. The total number of cells per cell type ( $B$ ) was fixed across three reference donors, while the allocation ratio was varied (e.g., 1:1:1, 2:1:1, 4:1:1, and 10:1:1).
Budgets: Simulations were run at total budgets of $B=60$ , $B=120$ , and $B=300$ cells per type.
Targets: Synthetic mixtures were generated with known ground-truth proportions (one equal, 59 drawn from a symmetric Dirichlet-multinomial distribution).
Algorithms: The study compared Weighted MuSiC (which utilizes between-donor variation and gene weights) against Ordinary Nonnegative Least Squares (NNLS) (an internal comparator using the same factorized reference basis but without gene weighting).
Validation: The study included diagnostic follow-ups involving alternative donor groupings, additional reference sampling blocks (to assess Monte Carlo standard errors), and an assessment on 12 measured PBMC RNA-seq samples with independent flow-cytometry proportions.

Key Results

Budget-Dependent Penalty: In the external cohort, the mean absolute error (MAE) difference between unequal (10:1:1) and balanced (1:1:1) allocations decreased as the budget increased.
- At $B=60$ , the weighted MuSiC error penalty was +0.621 percentage points.
- At $B=120$ , the penalty dropped to +0.276.
- At $B=300$ , the penalty became −0.066 (indicating a slight advantage for unequal allocation on average, though not statistically significant).
- The prespecified change ( $I(300) - I(60)$ ) averaged −0.687 points and was negative for 13 out of 14 donors.
Persistence of Individual Effects: Despite the average penalty decreasing to near zero at the largest budget, the effect was not uniformly negligible. At $B=300$ , 84 out of 168 specific reference configurations still exhibited higher error under unequal allocation. The trend persisted across alternative donor groupings and additional random sampling blocks.
Mechanism (Sampling Precision): Diagnostic analyses showed that profile variability and between-donor variance decreased as the minor-donor quota increased (from 5 to 25 cells). This suggests that the reduction in allocation penalty at higher budgets is consistent with improved sampling precision of donor expression profiles, though the study does not isolate this as the sole cause.
Measured Bulk Transfer: When applied to 12 measured PBMC samples, the study failed to recover major lymphocyte lineages (B and T cell estimates were effectively zero). Consequently, the simulation trend regarding donor allocation could not be validated or established in measured bulk data. Input diagnostics (gene exclusion, re-quantification) did not resolve this failure.
Thinning vs. Redistributing: A practical alternative—balancing donors by discarding cells (thinning 50:5:5 to 5:5:5)—increased weighted MAE by 0.547 points, suggesting that discarding data is less effective than redistributing a fixed budget.

Significance and Claims
The paper concludes that donor-allocation effects in MuSiC are budget-dependent. While increasing the total reference-cell budget reduces the average penalty of unequal allocation, significant configuration-level heterogeneity remains. A near-zero average contrast at large budgets does not imply that donor allocation is irrelevant for all specific cases.

The author emphasizes that reference design should be reported not just by total cell count, but by cells per donor and cell type. They caution against balancing references by discarding cells, as this reduces total information and increases error. However, the study modestly states that the observed simulation trend remains unestablished in measured bulk due to the failure to accurately deconvolve the specific PBMC samples tested. The findings are conditional on the tested MuSiC workflow and do not claim to identify a universally optimal allocation or a sufficient cell budget threshold for all deconvolution scenarios.

Technical Summary: Donor-Allocation Effects in MuSiC PBMC Simulations

More like this