Letting Heritable Expression Vote on Single-Cell Trees

When a tumor or tissue is profiled by single-cell RNA sequencing (measuring which genes each individual cell is expressing), every cell yields two kinds of evidence about its ancestry. The first is a handful of mutations visible in the RNA reads. The second is the expression profile itself, because daughter cells inherit chromatin and protein state from their mother and so resemble her for a while. Current tree-building methods use only the mutations and treat expression as something to paint onto the finished tree. This proposal adds expression as a second, deliberately weak term in a maximum-likelihood tree search, where maximum likelihood means scoring candidate trees by how probable they make the observed data. The expression term is designed so that it counts only the part of expression that is inherited. Cells that look alike merely because they became the same cell type are not counted as relatives.

The central bet is that inherited expression memory decays with time. The model makes the rate of decay explicit and estimates it from data. Expression should then help resolve recent splits, where RNA-derived mutations are too sparse and too often missing to place cells. Deep splits stay governed by the mutations. A model that simply rewards expression similarity would be expected to make trees worse, because convergent cell types would be grouped as relatives.

What exists and what is missing

The mutation-only side is represented by SCITE-RNA (Phylogenetic tree inference from single-cell RNA sequencing data). It scores trees with a beta-binomial model of reference and alternative read counts at single-nucleotide variants. It searches greedily, alternating between a cell-lineage representation and a mutation-tree representation to escape local optima. Only after the tree is fixed does it map expression onto it, for example CTSB upregulation along a metastatic branch in patient MM34. We reuse its likelihood and its search. What we change is that expression enters the score.

The expression-on-a-tree side is TreeVAE (Reconstructing unobserved cellular states from paired single-cell lineage tracing and transcriptomics data). TreeVAE takes a known lineage tree and lets each cell's low-dimensional latent state drift as a Gaussian random walk along the branches. It integrates out the unseen ancestors with linear-time message passing. On 603 metastatic lung cancer cells this cut latent cross-entropy from 21,348 under scVI to 3,051. That result motivates the proposal, since lineage clearly explains expression. It also constrains the design. A random walk treats any two similar cells as close relatives, so it will reward trees that group convergent cell types. TreeVAE never faced this problem because its tree was fixed. Inside a tree search the problem becomes central.

Two further works shape the expression term. LARRY (Lineage tracing on transcriptional landscapes links state to fate during differentiation) showed that a model predicting a progenitor's fate from its own transcriptome is 26% less accurate than one using its sister's fate. Heritable state is therefore real but only partly visible in expression. Its clone-labeled data give us a way to measure how long that memory lasts. SCIΦN (Single-cell mutation calling and phylogenetic tree reconstruction with loss and recurrence) is a warning about the mutation side. It found that mutation loss and recurrence are common in tumors, yet its sampling approach took roughly 30 hours on 255 cells. We keep SCITE-RNA's infinite-sites assumption, under which each mutation arises once and is never lost. We flag it as a known limitation and do not attempt to fix it here.

Core Claim

Let $a_{ij}, r_{ij}$ be the alternative and reference read counts of cell $i$ at variant $j$. Let $z_i$ be the cell's expression embedding from a pretrained scVI model, and let $c_i$ be its cell-type cluster with mean $\mu_{c_i}$. We score a tree topology $T$ with branch lengths $t$ by

$$\ell(T,t) = \sum_j \log P(a_{\cdot j}, r_{\cdot j} \mid T, t, \epsilon) + \log \mathcal{N}\big(z - \mu_{c} \,;\, 0,\ \Sigma_{T,t}(\alpha,\sigma)\big).$$

The first term is SCITE-RNA's beta-binomial likelihood with dropout parameters $\epsilon$. The second term is the exact Gaussian likelihood of the residuals $z_i - \mu_{c_i}$ under an Ornstein–Uhlenbeck process along the tree. That process is a random walk with a pull back toward zero at rate $\alpha$ and noise scale $\sigma$. Its implied covariance between two cells decays as $\exp(-\alpha d)$ in the path distance $d$ between them, which is why expression can only inform shallow splits.

Subtracting the cell-type mean removes similarity explained by shared fate. The pull rate $\alpha$ sets an inherited-memory half-life, $\ln 2/\alpha$. We estimate it jointly with the tree and can calibrate it independently on LARRY sister-cell pairs. As $\alpha$ grows the term vanishes and the method falls back to SCITE-RNA, so the data can tell us expression is useless.

Research Program

The work has three stages: build the model, measure inherited memory, and integrate it into the search.

To build the model, we pretrain scVI on each dataset and fix the embeddings, which keeps the expression term a cheap Gaussian on the tree. Message passing in the style of TreeVAE evaluates that Gaussian exactly in time linear in the number of cells, and a local rearrangement only requires updating messages along the affected path. A full joint refinement of the scVI decoder against the final tree is an optional last step, not part of the search.

To measure inherited memory, we fit $\alpha$ on LARRY clones before any tree inference. We compare residual expression covariance within clones and between clones at matched cell type and matched time point. If within-clone residual covariance is indistinguishable from between-clone covariance, the premise fails and we report that.

To integrate the term into the search, we add it to SCITE-RNA's space-swapping greedy search. Branch lengths are re-optimized after each accepted move, using SCITE-RNA's discrete mutation placement as a starting point. We also report a per-split diagnostic: how much each split's support changes when the expression term is removed. This shows where in the tree expression mattered.

Baselines

The naive reference point is SCITE-RNA run unchanged on the same variant read counts. This is also the strongest existing tree method for scRNA-seq in the file. It beat DENDRO and SClineager in path difference and variant-allele-frequency error across 5 to 50 clones, though the file gives no numeric values (no reported result found on Lacuna for SCITE-RNA path difference values). The proposed method beats it if it lowers path difference and Robinson–Foulds distance to the reference tree, with the gains concentrated in shallow splits, at comparable runtime. Both methods get the same input, the same random restarts and the same wall-clock budget.

The second baseline is the two-stage pipeline: SCITE-RNA's tree followed by TreeVAE on that fixed tree. It tests whether joint scoring beats post-hoc expression fitting. On expression fit, TreeVAE's own number is the bar, 3,051 latent cross-entropy versus 21,348 for scVI on the 603-cell lung data. We are not claiming to improve that number. We claim a better tree, and we will report whether a better tree also improves TreeVAE's fit when TreeVAE is rerun on it.

The ablation that isolates the key ingredient keeps the joint search but uses a plain Brownian random walk on raw embeddings, with no cell-type subtraction and no pull toward zero. If the proposed method is right, this ablation should do worse than SCITE-RNA alone in simulations with convergent cell types, while the full model does better.

We also include NJ on moment-matched corrected distances from Tree reconstruction guarantees from CRISPR-Cas9 lineage tracing data using Neighbor-Joining as a fast distance-based reference. Our only reported numbers for it are the most-parsimonious tree in 17 of 20 lung adenocarcinoma clones, and RF accuracy superior to Hamming NJ on Cassiopeia simulations, without numeric values.

Why not just…?

Why not just run SCITE-RNA and then fit TreeVAE on its tree? This is the strongest alternative. If the variant signal already pins down shallow splits, joint scoring adds nothing. The masking experiments below settle the question by varying variant density: our claim predicts a gap that grows as variants become sparser and vanishes when they are dense. If the gap is zero at realistic sparsity, the two-stage pipeline wins and we will say so.

Why not just add expression similarity, for example a Brownian term or an expression distance in neighbor-joining? Expression similarity mixes inherited state with convergence to the same cell type. LARRY shows the transcriptome misses much of what lineage carries, and nothing in TreeVAE's fixed-tree setting tested whether expression can discriminate between topologies. The Brownian ablation answers this directly. We expect it to group convergent cells and hurt accuracy, and if it does not, the cell-type subtraction is unnecessary.

Why not just use cell-type clusters as clade constraints, the pre-clustering route of DENDRO? SCITE-RNA already showed that pre-clustering pipelines degrade as clone counts rise. Cell types and clones are also different partitions whenever one clone gives rise to several fates, which is exactly the LARRY setting. We include a constrained-search variant only in the LARRY convergence check, where it should fail by forcing multi-fate clones apart.

Proposed Experiments

The first experiment is simulation. We extend SCITE-RNA-style simulations (up to 500 cells, 500 variants, 5–50 clones) so that expression is generated by cell-type means plus an Ornstein–Uhlenbeck residual on the true tree. We vary three things: the true $\alpha$, the variant density, and a convergence rate, meaning the fraction of distant clades assigned the same cell type. Metrics are Robinson–Foulds distance, path difference, and triplet accuracy stratified by split depth. Recovery of $\alpha$ is checked as well.

The decisive real-data check uses the 603-cell metastatic lung dataset from TreeVAE. There, CRISPR barcodes give a reference tree that is independent of expression. We mask most barcode sites, keeping 5–30%, to mimic sparse RNA-derived variants. We then ask whether adding expression moves the inferred tree toward the full-barcode tree. The full model should beat SCITE-RNA on shallow splits, and the Brownian ablation should not.

The LARRY data serve two purposes: the $\alpha$ calibration, and a convergence test in which clones sharing a fate must not be merged. Finally, the multiple myeloma patients MM16 and MM34 are an application. There we report which splits change and whether the chemotherapy-survivor clone and the CTSB-associated branch remain stable.

The outcome that falsifies the proposal is that, at masking levels typical of RNA-derived variants, the full model fails to beat both SCITE-RNA and the two-stage pipeline on shallow-split accuracy in the lung data.