Reading Cell Identity Off the Mutant Molecule

When a blood sample is sequenced, DNA from millions of cells is pooled into one test tube. Deep sequencing of a small set of genes, thousands of reads per position, can find a mutation carried by one cell in a thousand. But once the cells are broken open, nobody can tell whether that mutation sat in a monocyte, a neutrophil or a T cell. That matters. A cancer-associated clone confined to inflammatory monocytes is plausibly a different biological event from the same clone in lymphocytes. Current studies nonetheless collapse both into a single "variant allele fraction" (VAF), the share of reads that carry the mutant base. We propose to recover the cell of origin for each mutant DNA molecule directly from the molecule itself. We would read the methylation marks on the same fragment that carries the mutation. Methylation is a chemical tag on DNA whose pattern differs sharply between cell types, and on a single fragment it acts as a label for the cell the fragment came from. The result is a lineage-resolved clone fraction for every person in a cohort, obtained from archived DNA and without sorting cells.

The second half of the proposal uses this measurement to ask a sharper version of the "missing heritability" question. Missing heritability is the gap between how strongly traits run in families and how much of that can be traced to specific inherited variants. Somatic mutations, those acquired after conception, are not inherited, so they cannot fill that gap directly. Inherited variants can, however, influence which cell lineages grow mutant clones. Part of the germline effect on a trait may therefore travel through lineage-specific clonal expansion, and averaging over all cells would hide it. We will estimate how heritable lineage-resolved clone fractions are and how much trait heritability they mediate, compared with the bulk VAF that current studies use.

Core Claim

Targeted sequencing that preserves both base identity and methylation can assign each mutant molecule to a blood lineage. The method needs unique molecular identifiers (UMIs), which are random barcodes attached to each original DNA molecule before amplification, and a panel that pairs each mutation hotspot with nearby lineage-informative CpG sites, the positions where methylation occurs. With these, per-person, per-lineage clone fractions can be estimated to within a few percentage points of values from physically sorted cells. These lineage-resolved fractions should carry germline-heritable signal, and should mediate trait associations, that bulk VAF dilutes.

Let $n_k$ be the number of unique molecules at a locus assigned to lineage $k$, and let $a_k$ be how many of those carry the mutant allele. The heterozygous clone fraction in lineage $k$ is then

$$ c_k = \frac{2\,\sum_{j} a_j\, P(k \mid m_j)}{\sum_{j} P(k \mid m_j)}, $$

where each sum runs over molecules $j$, $m_j$ is the methylation pattern observed on molecule $j$, and $P(k \mid m_j)$ is its posterior probability of coming from lineage $k$. That posterior is computed from a reference of sorted-cell methylation profiles and the person's own lineage proportions. The factor of two converts mutant alleles to cells carrying one mutant copy. The quantity this adds to the literature is the vector $(c_1,\dots,c_K)$ per person, in place of the single bulk VAF.

Where This Sits

Three lines of prior work motivate the gap without closing it. The ultra-deep clonal hematopoiesis study in myocardial infarction patients sequenced 42 genes at a median 14,219× in bulk leukocytes. It found pathogenic clones at VAF ≥ 2% in 14.5% of 1,396 patients, and a TET2 mortality hazard ratio of 5.20 when a lipid index was high. It could not say which leukocytes carried the clones (the STEMI cohort study). The Alzheimer's single-nucleus study did resolve cell type, finding somatic burden elevated 30% in excitatory neurons and 17.5% in oligodendrocytes. It did so only for mutations in expressed transcripts and in 4,014 cells from 36 brains (single-cell mosaicism in AlzD). That scale is far from what heritability estimation needs. Methylation deconvolution of cell-free DNA with a read-level autoencoder estimates lineage proportions across up to 35 cell types (cfDecon), but it never looks at mutations. Our method takes cfDecon's central insight, that single reads carry lineage information, and applies it molecule by molecule to mutant reads. cfDecon applies it only to mixture proportions.

Two lines supply method. UMI-based deep targeted calling with network UMI grouping and read-level error models is the detection layer we reuse; the relevant work is the cfDNA UMI benchmarking study and DREAMS. That same benchmarking study also constrains us. It showed that Mutect2 and VarScan2 produce more false positives as depth rises unless samples are downsampled to matched depth. Because per-lineage molecule counts $n_k$ vary widely across people, depth-matched sensitivity has to be modeled per lineage, not only per sample. On the heritability side, mixed-model variance partitioning of the kind applied to 747 ADNI participants (GCTA brain volumes) provides the estimator. The research file's framing is that somatic mutations belong inside the germline genotype matrix. We reject that as conceptually wrong. Non-transmitted variants would inflate the non-shared environmental component of twin models, not heritability. Clone fractions therefore enter as a heritable intermediate phenotype and as a mediator, not as genotypes.

Research Program

The first component is panel design. For each clonal hematopoiesis hotspot, and for tiled exons of loss-of-function genes such as DNMT3A and TET2, we identify CpG sites within typical fragment length (about 200 bp) that separate myeloid, B, T and NK lineages. We will use public sorted-blood methylation references. We will report honestly what fraction of the mutational target has usable lineage-informative CpGs. If that fraction is low, the idea is limited, and that is itself a finding.

The second component is the chemistry and calling. Methylation conversion (enzymatic, to limit DNA damage) turns unmethylated C into T. This collides with C>T, the dominant somatic mutation class. We resolve the collision with strand-aware duplex UMI consensus. A true C>T on one strand appears as G>A on the complementary strand, which conversion does not touch, whereas a conversion-only T has no partner change. Detection then uses DREAMS-style read-level error modeling retrained on converted libraries.

The third component is the assignment model. Each molecule receives the posterior $P(k \mid m_j)$, with lineage priors taken from the individual's own non-mutant molecules. We then fit $c_k$ with uncertainty that propagates both molecule-count noise and assignment error. It will also flag molecules whose patterns fit no reference lineage, which is the analogue of the unannotated senescent cluster in the Alzheimer's study.

The fourth component is heritability analysis. We treat $c_k$ as quantitative phenotypes and estimate their SNP-heritability in a genotyped cohort. We then test whether lineage-resolved fractions mediate more of the germline genetic association with a downstream blood or cardiometabolic trait than bulk VAF does.

Baselines

The naive reference is bulk VAF from the same deep targeted UMI data with no lineage information. It is what current studies use, and it is the quantity every downstream analysis must beat.

The strongest competitor for lineage assignment is physical sorting. Cells are sorted by flow cytometry into myeloid, B, T and NK fractions, and each fraction is sequenced with the same panel and caller. Sorting is also our ground truth, so we treat it as an upper bound rather than a rival. The claim to test is that molecule-level assignment on unsorted DNA comes within an absolute error of 0.02 in $c_k$ for clones at bulk VAF ≥ 1%.

The strongest sorting-free computational competitor is a two-step pipeline that estimates lineage proportions with cfDecon and then assigns each clone to lineages by feasibility. A clone at bulk VAF $v$ must reside in lineages whose combined proportion is at least $2v$. The mutant allele fraction is then allocated by least squares across individuals. This baseline gets the same sequencing reads, with methylation read off the same fragments but pooled, and the same reference atlas. Beating it means a lower per-person L1 error on $c_k$ against sorted truth, especially for clones spread across two or more lineages, where feasibility constraints cannot decide.

For detection, the baselines are shearwater-AND and DREAMS-vc run on conventional libraries from the same DNA. The cfDNA benchmark reports shearwater-AND at tumor-informed ROC-AUC 0.984 and DREAMS-vc at tumor-agnostic ROC-AUC 0.808 (benchmarking study). Our converted-library caller must not lose more than a small, pre-registered margin of sensitivity at VAF 0.5–2% relative to DREAMS-vc on conventional libraries. Otherwise the methylation gain costs too much detection.

No reported result was found on Lacuna for per-molecule lineage assignment of somatic variants, so there is no published number to beat on that task. The key ablation removes the per-molecule methylation posterior and replaces it with bulk proportions. It is the proportion-only baseline above in its simplest form, and it isolates whether reading identity off the mutant molecule is what matters.

Why not just…?

Why not just sort the cells? For fresh blood in a small study, sorting is simpler and better, and we use it as ground truth. The problem is that heritability-scale cohorts hold banked DNA, not viable cells. The cohorts in the file, such as the 1,396 STEMI patients, 747 ADNI participants and 4,737 GCKD participants, are all of this kind, and sorting also multiplies library cost by the number of lineages. The method is only worth building if it runs on archived DNA. It also extends to tissues that cannot be sorted into DNA-bearing lineages at scale. If sorted-versus-unsorted agreement fails the 0.02 target, the correct conclusion is that prospective sorting is required, and we would report that.

Why not just deconvolve proportions and assign clones by allele-fraction bounds? This is the real threat. For large clones confined to one abundant lineage, a VAF above half the lymphocyte fraction already rules out lymphocytes. The two-step baseline is included specifically to find the regime in which it suffices. We expect it to fail for small clones and for multi-lineage clones, because many allocations fit the same bulk VAF. The head-to-head on sorted truth, broken down by clone size and lineage spread, settles it.

Why not just genotype mutations from single-cell RNA sequencing? The Alzheimer's study shows that transcript-based genotyping only sees expressed genes. Its scale, 4,014 cells from 36 people (single-cell mosaicism in AlzD), is orders of magnitude below what SNP-heritability of a clone phenotype requires. Many clonal hematopoiesis drivers are truncating variants that trigger transcript decay, which makes them even less visible in RNA.

Why not just put bulk VAF into the heritability model? That is the ablation. If bulk VAF captures as much heritable and mediated signal as lineage-resolved fractions, lineage resolution is not needed for this question. We would conclude that, rather than argue around it.

Proposed Experiments

The first experiment uses controlled mixtures. We spike sorted cells from donors with known clones into defined backgrounds at clone cell fractions from 0.2% to 10%. This yields per-molecule assignment accuracy, error in $c_k$, and calibration of the reported intervals. The second is natural validation. Donors with known clonal hematopoiesis, selected by bulk screening at VAF ≥ 1%, have blood split into a sorted arm and an unsorted converted arm. We compare the methods against sorted truth with all baselines above at matched depth. We also check the depth-confounding failure mode flagged in the cfDNA benchmark by downsampling each lineage's molecule count and testing that estimated $c_k$ does not drift with depth.

The third is the population step. It requires a genotyped cohort with banked blood DNA large enough to be informative. We would prefer to estimate power before committing than to promise an effect. In that cohort, we estimate SNP-heritability of myeloid versus lymphoid clone fractions and of bulk VAF. We then fit mediation models between germline polygenic scores and a cardiometabolic outcome.

The decisive check has three parts. Lineage-resolved fractions must beat bulk VAF in heritable variance or mediated effect by more than their added estimation noise. The per-molecule model must beat the proportion-only ablation on sorted truth. The converted-library caller must keep sensitivity at low VAF. If any of the three fails, the corresponding claim is withdrawn.