Semi-Blind Tumor Subclone Reconstruction: Treating Germline SNPs as Pilots

Bulk tumor DNA is a mixture. Every read comes from one of several subclones, which are groups of cancer cells sharing a mutational history, or from normal cells. Most pipelines take this mixture apart in a fixed order. They first segment the genome into copy-number states using the allele balance at inherited (germline) heterozygous sites. They then freeze that segmentation and cluster the allele frequencies of acquired (somatic) mutations into subclones. Any copy-number error made in the first step passes straight into the clustering step. The largest community benchmark of this task found that the choice of algorithm explained 19–35% of performance variance, more than tumor biology did (~15%). It also found that accuracy collapses below roughly 32 reads per chromosome copy, and that ensembles of existing tools did not beat the best single tool (Crowd-sourced benchmarking of single-sample tumor subclonal reconstruction).

Wireless receivers face a problem with the same shape, called pilot contamination. A base station learns each user's channel from short known training signals ("pilots"). When pilots are reused, the base station sees an inseparable sum of several users' channels. No amount of extra antennas removes the resulting error floor (Massive MIMO: How many antennas do we need?). The floor is broken only when the receiver also uses the unknown data symbols as soft extra pilots, estimating channel and data together (Grant-Free Access via Bilinear Inference for Cell-Free MIMO With Low-Coherence Pilots; Data-Aided Channel Estimation Utilizing Gaussian Mixture Models). This note proposes the same move for tumors. Germline heterozygous sites play the role of pilots, and somatic mutations play the role of data. A damped, soft-feedback joint estimator replaces the frozen segment-then-cluster pipeline. Because tumor biology cares about the parameters themselves and not only the "decoded" assignments, the note also tests whether joint approximate inference biases those parameters.

Core Claim

Germline-only copy-number calling has a contamination-like floor. Two or more subclones with different prevalences and different copy states inside one segment can produce the same expected allele balance and read depth. Adding germline SNPs to the segment reduces noise but does not remove this ambiguity, much as adding antennas does not remove pilot contamination. Somatic mutations in that segment carry the missing information. Their allele frequencies cluster at values set by which subclone carries them and at what multiplicity, the number of copies of the chromosome bearing the mutation.

Write $\phi_k$ for the cellular fraction of subclone $k$, $m_{ik}$ for the number of mutant copies of mutation $i$ in subclone $k$ (zero if absent), $c_{ks}$ for the total copy number of segment $s$ in subclone $k$, and $D_i$ for the read depth at $i$. The expected count of mutant reads is

$$\mathbb{E}[a_i] = D_i \,\frac{\sum_k \phi_k\, m_{ik}}{\sum_k \phi_k\, c_{ks}},$$

where $k$ ranges over subclones including the normal component. The numerator is bilinear in the unknowns: prevalences multiply discrete copy assignments. That is the same structure as channels multiplying unknown symbols in a contaminated receiver.

The prediction is concrete. A joint estimator that feeds soft somatic-mutation evidence back into copy-number states should flatten the accuracy cliff below 32 reads per chromosome copy. The gain should be concentrated in segments with subclonal copy-number changes, and it should be near zero in clonal segments.

Research Program

The estimator alternates two steps on a factor graph.

The first step is a copy-number step. It computes beliefs over per-subclone copy states for each segment from germline allele balance and depth, the "pilot" step.

The second step is a mutation step. It computes beliefs over $(k, m_{ik})$ for each somatic mutation given the current copy beliefs and prevalences, the "data" step. Its soft outputs return to the copy-number step as additional evidence.

Prevalences $\phi$ are updated between the two steps. Following BiGaBP (Grant-Free Access via Bilinear Inference for Cell-Free MIMO With Low-Coherence Pilots), the feedback from mutation to copy number is damped by a belief-scaling schedule $\gamma(t)=t/t_{max}$. Early, unreliable mutation decisions therefore cannot lock in a wrong segmentation. Where reads physically link a somatic mutation to a nearby germline site, the program adds a linkage factor in the spirit of joint genotyping-and-phasing (Joint haplotype assembly and genotype calling via sequential Monte Carlo algorithm). That work showed that refusing to freeze genotype calls rescued over 50% of miscalled sites. Combinatorial copy-state spaces are pruned to the top $K$ configurations per segment, as ParticleHap does, to keep cost linear in genome length.

The second strand comes from a warning in the wireless literature. EM with belief propagation was shown to converge to a "surrogate" channel that compensates for the detector's approximations. The surrogate decoded better than the true channel (Blind Channel Estimation and Joint Symbol Detection With Data-Driven Factor Graphs). In a receiver that is harmless. In a tumor, a surrogate $\phi$ would be a wrong biological claim, and joint genomic models such as SIEVE explicitly aim at true parameters. The program therefore measures bias in recovered prevalences, not just assignment accuracy. It also tests a simple correction: after convergence, the soft assignments are fixed and $\phi$ is refit under the exact mixture likelihood, with no factor-graph approximation.

Baselines

The naive reference is the frozen pipeline: segment and call copy number from germline sites alone, then cluster somatic allele frequencies conditioned on that fixed segmentation. Our implementation of it uses the same observation model, the same candidate copy states, and the same code paths as the proposed method. The only difference is that feedback is switched off. Any gain is therefore attributable to coupling and not to a better likelihood.

The strongest existing competitors are the challenge leaders PhylogicNDT, GISL and CCube (Crowd-sourced benchmarking of single-sample tumor subclonal reconstruction), because they define the bar on the only shared benchmark in the file. They will be run on the same 51 simulated tumors and the same downsampled read sets with their released containers and default settings. The proposed method gets no per-tumor tuning; its few hyperparameters ($K$, $t_{max}$) are fixed on a held-out set of our own simulations. The file reports the leaders' rankings and the variance decomposition but no per-subchallenge scores. I have found no reported result on Lacuna for PhylogicNDT, GISL or CCube subchallenge scores, so we will regenerate their scores ourselves. Beating them means a higher score on the cellular-prevalence and mutation-assignment subchallenges at reads per chromosome copy below 32, without a loss above 32.

An ensemble of the leaders is included because the benchmark reported that ensembles failed to beat the best individual method. If our method also fails to beat that ensemble, that is a clear negative.

Three ablations isolate the key ingredient. Hard feedback (one re-segmentation pass using hard mutation clusters) tests whether softness matters. Removing the $\gamma(t)$ schedule tests whether damping matters. Removing the post-hoc exact refit of $\phi$ tests the surrogate-bias question.

Why not just…?

Why not just sequence deeper? Above about 32 reads per chromosome copy the frozen pipeline may indeed be adequate, and the proposal predicts no gain there. Two things still argue for the method. Low effective depth is common in low-purity or highly aneuploid samples even at nominal high coverage. And the depth sweep in our experiments directly measures how much sequencing the joint estimator saves. If the curves coincide everywhere, the idea is wrong.

Why not just ensemble existing tools? The DREAM study already ran this and found ensembles did not beat the top single method (Crowd-sourced benchmarking of single-sample tumor subclonal reconstruction). That is consistent with every member sharing the same frozen-segmentation error. We still include the ensemble so the claim is tested and not just asserted.

Why not just iterate the pipeline once: segment, cluster, re-segment? This is the real threat. It costs almost nothing and captures much of the feedback idea. The hard-feedback ablation is exactly this. If it matches the soft, damped estimator within noise, the contribution shrinks to "iterate your pipeline," and the note should say so. The wireless evidence that a decoupled hard-decision pipeline (MMV-AMP followed by zero-forcing) floors while soft bilinear inference waterfalls (Grant-Free Access via Bilinear Inference for Cell-Free MIMO With Low-Coherence Pilots) motivates expecting a gap. It does not prove the gap exists in tumors.

Why not just run full Bayesian MCMC over everything? That would be the principled answer if it were tractable, but the joint space over trees and copy states is combinatorial. On a small subset of tumors we will run a long-chain joint sampler as a near-exact reference. This shows whether the pruned estimator loses accuracy, and whether its prevalence bias exceeds the sampler's.

Proposed Experiments

The primary benchmark is the 51 simulated SMC-Het tumors, with ground truth for purity, subclone number, prevalences, assignments and trees (Crowd-sourced benchmarking of single-sample tumor subclonal reconstruction). We downsample reads to span roughly 8 to 64 reads per chromosome copy and score all seven subchallenges. Results are stratified by whether a mutation lies in a subclonal copy-number segment.

A second, controlled simulation plants deliberately degenerate segments. In these, two subclone configurations give identical expected germline allele balance and depth. This lets us test the core claim directly: germline-only error should plateau as germline SNP density grows, while joint-estimator error keeps falling as somatic mutations per segment increase.

The decisive checks are three. The joint estimator must beat the frozen pipeline and the challenge leaders below 32 reads per copy, specifically in subclonal segments. It must beat the one-pass hard-feedback ablation. And with the exact refit, its prevalence bias must not exceed that of the frozen pipeline. Failing the second check reduces the contribution to "iterate your pipeline." Failing the third means the joint estimator trades biological truth for fit, the genomic version of the surrogate-channel effect.