Clone-Fraction-Weighted Rare Mutation Tests for Cancer Survival

When a tumor is sequenced in bulk, a mutation carried by 90% of its cancer cells and one carried by 5% both end up recorded the same way: "mutated." Survival studies then ask whether patients with the mutated gene progress or die sooner. Two problems compound here. The first is that bulk measurement hides how much of the tumor actually carries the lesion, so the 0/1 label mixes strong and weak exposures. The second is that most mutations in any single gene are rare across patients, which leaves gene-level survival tests underpowered from the start. This proposal asks whether coding each patient's mutation by the estimated fraction of cancer cells that carry it can recover usable statistical power for gene-level rare somatic mutation tests against censored survival times. The fraction is inferred from bulk sequencing, and the test carries forward the uncertainty in that estimate.

The core idea fits in two sentences. Replace the "mutated / not mutated" covariate in a Cox survival test with the fraction of cancer cells carrying the mutation, and down-weight fractions that are poorly estimated. Because the binary label is a diluted version of that exposure, the test should gain power by a predictable amount: the gain is large when carriers differ widely in how clonal their mutations are and vanishes when every mutation is clonal.

Core Claim

A Cox proportional hazards model relates a patient's instantaneous risk of the event to covariates through a multiplicative factor, without specifying how baseline risk changes over time. Within that model we propose a variance-component score test for a gene, of the kind used for rare germline variants (SKAT, which pools many rare variants into one test of whether the gene as a whole matters). The genotype entry for patient $i$ and mutation $j$ is $\hat c_{ij}$, the estimated cancer cell fraction carrying mutation $j$, set to 0 if the mutation is absent. Let $M_i$ be patient $i$'s martingale residual from a null Cox fit on clinical covariates, tumor purity and tumor mutational burden, meaning observed minus expected events. The statistic is

$$Q = \sum_j w_j^2 \Big(\sum_i \hat c_{ij} M_i\Big)^2 ,$$

where $w_j$ combines a frequency weight with a reliability weight $r_j$. Here $r_j$ is the average ratio of the variance of true cell fractions to the variance of their estimates for mutation $j$, computed from the posterior of the variant-allele-frequency clustering.

The point to be precise about is what uncertainty does and does not affect. Under the null, $\hat c$ depends only on genomic data, so the score test keeps its nominal false-positive rate whether or not uncertainty is modeled. Measurement error therefore costs power and biases the hazard ratios. It does not cost validity. The proposal handles it in that way: reliability weighting tunes power, and regression calibration corrects the hazard ratio estimates reported for discovered genes.

The quantitative prediction comes from comparing the two codings. Suppose risk truly scales with the fraction $c$. A binary indicator then recovers only the part of $c$ that it correlates with. The relative efficiency of the binary test against the fraction-weighted test, with true fractions, is $(E[c\mid \text{carrier}])^2/E[c^2\mid\text{carrier}] = 1/(1+\mathrm{CV}^2)$, where CV is the coefficient of variation of cell fractions among carriers. The expected gain in effective sample size is therefore $1+\mathrm{CV}^2$, minus a loss from estimation noise that the reliability weight tries to contain. This tells us in advance which genes and cancer types should benefit. It also tells us that genes dominated by clonal driver mutations should show no gain.

Relation to Prior Work

Several studies motivate the gap. SubNetDL clusters bulk variant allele frequencies with SciClone into one to six subclones and feeds subclone-level network features to a graph attention model. It predicts binary therapy response, with a median AUROC of 0.74 on TCGA and 0.77 on MSK immunotherapy cohorts. It shows that subclonal structure carries signal, but it uses no censoring, no hazard ratios and no gene-level hypothesis test. Conversely, the lenvatinib thyroid cancer study runs Cox regression on whole-tumor mutation indicators. This is exactly the diluted coding we replace.

MEAD shows that treating deconvolved cell-type proportions as known constants distorts downstream inference, and that weighting features by inverse noise variance beats hard feature selection. We take that principle into the survival setting, applied to subclonal fractions instead of transcriptomic cell-type proportions.

The Alzheimer's age-of-onset study shows that modeling time-to-event with a Cox mixed model detects rare variants that a case-control logistic analysis weakens; it found ERN1 rs56201815 with a hazard ratio of 4.22. Its tests were single-variant and germline, and it did not aggregate. We reuse its main lesson, that survival coding over binary coding is a source of power, and add aggregation and clonal weighting.

Two papers supply methods. The Hi-C informed kernel association test extends SKAT and combines tests across several borrowing scales with the Cauchy combination (ACAT). We use ACAT the same way, combining across competing mutation codings. That paper is limited to continuous traits, and we do not use its 3D-genome borrowing.

DeepSurv supplies a flexible Cox predictor, reaching a C-index of 0.867 on METABRIC against 0.816 for linear Cox. We use it only as a prediction-side comparison, because it does not produce calibrated gene-level tests.

One constraint deserves mention. The research file notes that bulk VAF clustering cannot separate subpopulations below roughly 5% from artifacts. Low-fraction calls are therefore exactly where reliability weights must shrink contributions, and we report results with and without them.

Research Program

The first step is building fractions with uncertainty. For each TCGA tumor we convert variant allele frequency, local copy number and purity into a posterior over the cancer cell fraction of each mutation, using SciClone cluster posteriors as in SubNetDL. We keep the posterior mean and variance.

The second step is the test itself. It is the score statistic above, with a mixture-of-chi-square null computed from the null Cox information matrix, as in SKAT. A permutation fallback applies when a gene has few carriers. We run three codings per gene: binary, clonal-only (fraction above 0.8), and fraction-weighted with reliability weights. These are combined with ACAT into one gene p-value. The combined test is the proposed method, and its components are reported separately so that the fraction coding's contribution is visible.

The third step is estimation. For significant genes we report hazard ratios per unit cell fraction, corrected by regression calibration using the posterior variances. Because purity and mutational burden affect both fraction estimates and survival, both are null-model covariates. A stratified analysis by purity tertile checks that the signal is not a purity proxy.

Baselines

The naive reference point is per-gene Cox regression on a whole-tumor binary "any mutation" indicator, adjusted for the same covariates. This is the standard analysis used in clinical genomic survival studies such as the lenvatinib study. It defines the false-positive and discovery floor.

The strongest existing competitor for the testing question is a binary-coded variance-component survival test, meaning SKAT applied to Cox martingale residuals with 0/1 mutation entries. It shares our statistic, null model and frequency weights and differs only in coding. Any gain we report against it is therefore attributable to the clonal fraction and not to aggregation. On TCGA we found no reported result on Lacuna for gene-level somatic rare-mutation survival tests, so the comparison numbers will be produced here.

For predictive framing, the closest reported numbers are SubNetDL's median AUROC of 0.74 (TCGA) and 0.77 (MSK) for binary response. These are not survival metrics and are not directly comparable. We will retrain a SubNetDL-style subclone model and DeepSurv with a partial-likelihood loss on identical splits, matched hyperparameter search budgets and identical covariates, and compare held-out C-index.

Two ablations isolate the key ingredients. The first removes reliability weights to test whether uncertainty handling matters. The second runs the fraction coding alone, without ACAT, to test whether the gain survives without hedging across codings.

Beating the binary-coded test means more genes at a matched false discovery rate on real data, with the null calibration preserved under outcome permutation. It also means higher power in semi-synthetic simulation at equal false-positive rates. The gain should track the $1+\mathrm{CV}^2$ prediction across genes.

Why not just…?

Why not just restrict to clonal mutations and keep the binary test? This is the real threat. If survival effects come almost entirely from truncal drivers, discarding subclonal calls removes the dilution without any modeling. We take it seriously by including clonal-only coding both as a standalone baseline and as an ACAT component. The decisive check is gene-level. Genes where carriers' fractions are spread widely (high CV) should favor fraction weighting. Genes with tight, near-clonal fractions should show the two tying. If clonal-only matches us everywhere, the fraction-weighting claim is wrong, and we will say so.

Why not just add an intratumor-heterogeneity score or mutational burden as a covariate? Global heterogeneity scores capture "how mixed is this tumor," not "how much of the tumor carries a lesion in this gene." They cannot localize an effect to a gene. SubNetDL's own comparison shows bulk mutational burden underperforming subclone-aware features ($P = 0.02$ and $P = 6.0\times10^{-3}$ on the MSK cohorts). We include burden and purity as covariates, so any gene signal we report is net of them.

Why not just train a deep survival model on all features? DeepSurv-style models can predict well. The METABRIC result shows gains over linear Cox when nonlinearity matters, and the same paper shows parity on linear data. But they give no calibrated per-gene p-values, and with sparse rare mutations they risk overfitting. We report their held-out C-index as a comparison. We expect the scientific value of this proposal to lie in calibrated discovery and corrected hazard ratios, not in prediction leaderboards.

Proposed Experiments

The first experiment is semi-synthetic power and calibration. We keep real TCGA mutation matrices, purity values and cell fraction posteriors, and simulate censored survival times from Cox models under three regimes: risk scaling with true fraction, risk from clonal mutations only, and threshold effects. We vary censoring between 30% and 70% and effect sizes in a range where binary tests have 20 to 60% power. We measure type I error at $\alpha$ down to $10^{-4}$ under a null with no effect, and power for each coding and for the ACAT combination. We then check whether the observed power gain per gene matches $1+\mathrm{CV}^2$ after the reliability loss. A clear win for fraction weighting under its own regime, with no meaningful loss under the clonal-only regime because ACAT absorbs it, supports the proposal.

The second experiment is a real-data discovery run on the TCGA pan-cancer cohort of 507 treated patients used by SubNetDL, using overall survival and progression-free intervals. We run gene-level tests stratified by cancer type and compare discoveries at matched false discovery rate against the binary-coded survival test. Calibration is checked by outcome permutation within strata. Where the MSK_2018 and MSK_2020 immunotherapy cohorts release survival times alongside panel allele frequencies, they serve as an external replication set for genes passing in TCGA. If survival times are unavailable, replication is limited to held-out TCGA cancer types, and we state that limitation.

The third experiment compares survival coding with response coding. Following the logic of the Alzheimer's onset study, we test whether discovery in the same patients is stronger with the Cox fraction-weighted test than with a logistic fraction-weighted test on binary response. This tests directly whether time-to-event information adds power on top of clonal weighting.

The decisive checks are three. Calibration must hold under permutation. The gene-level power gain must correlate positively with carrier fraction CV, and failure here falsifies the dilution mechanism. Hazard ratios after regression calibration should be larger than plug-in estimates by roughly the inverse reliability, and should be more consistent between TCGA and replication cohorts.