A tube of blood is the most constrained recording device in biology: one measurement channel, many independent sources. Plasma carries microbial cell-free DNA (mcfDNA) — microbial DNA fragments shed into the bloodstream from the gut, lungs, mouth, and skin — so every plasma sample is a monaural mixture. This mixture is clinically rich: a study of mechanically ventilated ICU patients found lung-to-blood microbial translocation in 41% of severe pneumonia patients, where standard blood cultures detect infection in fewer than 10%, with plasma mcfDNA burdens of 4,015 molecules/μL in microbiologically confirmed pneumonia versus 210 in clinically diagnosed cases and 0 in uninfected controls (Microbial lung-to-blood translocation associates with systemic inflammation in severe pneumonia: evidence from paired plasma and lower respiratory tract metagenomics). Yet current assays read this mixture crudely: genus-level matching at a >0.1% abundance threshold, with no way to say which body site a detected microbe came from, no uncertainty on any single call, and — because metagenomic reads can re-identify patients — no way to pool the multi-hospital cohorts that rare-taxon diagnostics require.
We propose to treat plasma mcfDNA as a single-channel source separation problem and solve it with the tools of entrywise spectral inference. The framing comes from audio separation, where a single microphone signal must be split into sources using only their structural signatures; the theory comes from two adjacent literatures that have never been combined. Entrywise eigenvector analysis bounds the error of individual coordinates of estimated spectral directions — the kind of control a per-species, per-patient diagnostic call needs, and which global average-error bounds cannot supply (Spectral Methods for Data Science: A Statistical Perspective). Eigen-gap sensitivity theory governs when a spectral direction is stable at all: the gap between consecutive eigenvalues is the currency that determines how much noise — including noise deliberately injected for differential privacy, a guarantee that outputs reveal nothing about any single person's record — a subspace can tolerate (Private Low-Rank Approximation for Covariance Matrices via Dyson Brownian Motion; Spectral Perturbation Bounds for Low-Rank Approximation with Applications to Privacy). Our claim is that these three threads meet at exactly the right place: weak, closely spaced compartment signatures in compositional, >90%-sparse count matrices, diagnosed one species and one patient at a time.
Core Claim
Model the plasma data matrix as $X = AS + E$, where $X \in \mathbb{R}^{d \times n}$ stacks $d$ species abundances across $n$ patients, the columns of $A \in \mathbb{R}^{d \times C}$ are the signatures of $C$ body-site compartments (lung, gut, oral, skin), the rows of $S$ are per-compartment activity levels, and $E$ collects sampling noise, batch artifacts, and — when cohorts are pooled across institutions — calibrated privacy noise. This is structurally the same object that nonnegative matrix factorization splits in single-channel audio, where a nonnegative magnitude spectrogram is decomposed into spectral bases and activations (A Two-step NMF Based Algorithm for Single Channel Speech Separation); and it inherits the same permutation ambiguity — nothing in $X$ alone labels which factor is "lung" — that permutation invariant training resolves in speech by minimising loss over all output permutations (Single-Channel Multi-talker Speech Recognition with Permutation Invariant Training).
The claim has two parts. First, identifiability: the number of compartments recoverable from the single channel is governed by the spectral gap $\delta_C$ separating the $C$-th signal direction of $AS S^\top A^\top$ from the noise bulk, and existing separability heuristics have no theory of when they fail — the audio analogue of this failure is documented directly, where sources with overlapping low-frequency structure defeat matrix-factorization masks (A Two-step NMF Based Algorithm for Single Channel Speech Separation). Second, inference: because a diagnostic is a per-entry decision — "is this species in this patient lung-derived?" — the relevant error metric is the worst row error $\|\hat{U} - UW\|_{2,\infty} = \max_i \|(\hat{U} - UW)_{i\cdot}\|_2$, where $\hat{U}$ is the estimated top-$C$ subspace of $X$, $U$ the population subspace, and $W$ an alignment matrix. Our theoretical target, stated as a conjecture to be proven, is an entrywise bound of the form
$$\|\hat{U} - U W\|_{2,\infty} \;\lesssim\; \frac{\sigma \sqrt{\log d}}{\delta_C} \cdot r(C),$$
where $\sigma$ is the per-entry noise scale, $\delta_C$ the effective gap, and $r(C)$ a residual term reflecting the row-correlated, non-independent noise that metagenomic count matrices actually exhibit. The key qualification is that the gap in the denominator is random under injected noise, and its collapse probability can be pushed from $O(s^2)$ to $O(s^3)$ by using complex-valued rather than real-valued noise, because eigenvalues of complex Hermitian random matrices repel each other more strongly (Private Covariance Approximation and Eigenvalue-Gap Bounds for Complex Gaussian Perturbations). That repulsion, combined with leave-one-out decoupling, is what we expect to keep the entrywise bound informative precisely in the weak-signal regime — signal strengths scaling like $n\rho_n$ — where the field's entrywise Berry–Esseen theorems already show per-coordinate normality survives (Entrywise limit theorems for eigenvectors of signal-plus-noise matrix models with weak signals).
Research Program
The theory program closes a gap that its own authors flag as open. The Dyson Brownian Motion line models privacy noise as a continuous matrix diffusion whose eigenvalues repel like charged particles, and by integrating error along the diffusion trajectory it improves the Frobenius-norm error of the Gaussian privacy mechanism to $O(\sqrt{kd})$ — but these bounds are strictly aggregate, and their extension to coordinate-wise $\ell_\infty$ or $\ell_{2,\infty}$ error is explicitly open (Private Low-Rank Approximation for Covariance Matrices via Dyson Brownian Motion). Similarly, contour bootstrapping bounds the spectral-norm error of low-rank approximation as $\|E\| \cdot \lambda_p / \delta_p$ but says nothing per-row (Spectral Perturbation Bounds for Low-Rank Approximation with Applications to Privacy). We will build the entrywise analogue by combining their diffusion/contour machinery with the leave-one-out technique — construct an auxiliary matrix with the $i$-th patient's noise replaced by its expectation, prove the auxiliary and real eigenvectors are close, and read off per-row limits (Spectral Methods for Data Science: A Statistical Perspective). The genuinely new difficulty, and the reason the combination is not mechanical, is that metagenomic noise violates the independence assumptions leave-one-out relies on: rows are coupled by compositionality (abundances sum to one) and by batch artifacts. We will handle this with the weak-signal entrywise limit machinery of Entrywise limit theorems for eigenvectors of signal-plus-noise matrix models with weak signals, whose $n\rho_n$ scaling is designed for signals that barely clear the noise, and we will further import the two-stage "sparsistent" strategy — select the species support first, then run spectral estimation on the submatrix — whose $\ell_{2,\infty}$ rate $\sqrt{k \log p / n} + s \log p / n$ under subgaussian designs is the closest existing entrywise result (Entrywise Recovery Guarantees for Sparse PCA via Sparsistent Algorithms). That paper's own limitation — its bounds require exact support recovery, which rare translocated taxa will not give — is exactly the regime our Berry–Esseen-based calibration must repair.
The method program turns the theory into a deployable pipeline: a spectral decomposition of the pooled, differentially private covariance of mcfDNA profiles; an eigengap statistic that gates each compartment claim (reporting "not separable" when $\hat{\delta}_C$ falls below the noise floor, rather than emitting a confident wrong answer); compartment labels anchored by a small reference set of taxa with known body-site associations, resolving the permutation ambiguity the way an enrollment prompt anchors target-speaker extraction in audio; and per-species, per-patient confidence intervals derived from the entrywise limit laws. The deployment program is a prototype for cross-institutional pooling: each hospital releases a DP-noised covariance summary rather than raw reads, directly addressing the re-identification hazard that currently prevents pooling.
Baselines
The naive reference point is the clinical practice embodied in the translocation study itself: call a plasma species present when its genus-level abundance exceeds 0.1%, with no compartment attribution and no uncertainty (Microbial lung-to-blood translocation associates with systemic inflammation in severe pneumonia: evidence from paired plasma and lower respiratory tract metagenomics). It is the incumbent because it is what runs today, and its documented operating characteristics (41% translocation detection in ventilated ICU patients; median burdens of 4,015 versus 210 versus 0 molecules/μL across confirmed, clinically diagnosed, and uninfected groups, $p < 0.0001$) are the numbers any replacement must be read against. We will run it on exactly the same plasma samples; no reported result found on Lacuna for compartment-level attribution accuracy in this cohort, so we will establish that number ourselves. We count as beating it if we attribute lung-origin species at higher sensitivity at a fixed false-discovery rate of 10%, at species rather than genus resolution.
The strongest existing diagnostic method in the file is the cross-cohort support vector machine trained on seven universal colorectal cancer markers, which achieved an AUC of 0.80 across four international cohorts, 0.88 with age and BMI covariates, and 0.82–0.84 in early-stage disease (Multi-cohort analysis of colorectal cancer metagenome identified altered bacteria across populations and universal bacterial markers). It is the strongest competitor because it is the only method with a demonstrated cross-cohort generalization number on microbial data. We will run it with the same leave-one-cohort-out splits, the same tuning budget and compute, both on raw pooled data and on DP-noised pooled data. Beating it means matching or exceeding its AUC while training only on privacy-noised pooled summaries, the regime where direct pooling degrades — and additionally producing compartment attribution, which a discriminative classifier does not output by construction.
The strongest separation baseline is the two-step NMF of A Two-step NMF Based Algorithm for Single Channel Speech Separation, ported to nonnegative abundance matrices: an initial factorization, bootstrapping of noise bases from the residual, a refined pass, and soft masking. We port it faithfully and give it the same data and tuning budget; no reported result found on Lacuna for NMF applied to mcfDNA mixtures, so we establish its operating point ourselves. Beating it means equal or better attribution AUC plus valid per-call intervals, which NMF does not produce. We will also port a deep-clustering variant, whose audio results — 10.3 dB SDR improvement and WER reduced from 89.1% on unprocessed mixtures to 30.8% (Single-Channel Multi-Speaker Separation Using Deep Clustering) — make it the strongest deep baseline in its home domain; no reported result found on Lacuna for deep clustering on metagenomic data, so we treat it as exploratory rather than as a calibrated competitor.
Finally, two ablations isolate our key ingredients. The calibration ablation runs the identical decomposition with thresholds calibrated by global Frobenius/spectral bounds versus our entrywise bounds; the noise ablation injects real-symmetric versus complex-valued privacy noise at matched $\varepsilon$, testing whether the $O(s^3)$ versus $O(s^2)$ repulsion difference (Private Covariance Approximation and Eigenvalue-Gap Bounds for Complex Gaussian Perturbations) measurably improves interval coverage in the weak-gap regime.
Why not just…?
Why not just train a bigger discriminative classifier on pooled cohorts? This is the most natural alternative, and the cross-cohort SVM shows discriminative learning works when data can be pooled (Multi-cohort analysis of colorectal cancer metagenome identified altered bacteria across populations and universal bacterial markers). It falls short on three counts that our experiments will make concrete. The pooling it requires is precisely what re-identification risk blocks, and the cohort confounding it needed rank-sum meta-analysis to remove is exactly what privacy noise plus entrywise calibration must handle; a classifier answers "disease or not" but not "which compartment shed this microbe," which is the clinically actionable question for localizing infection; and it issues no per-call uncertainty. The decisive comparison is the DP-pooled training regime: if the SVM retains its 0.80 cross-cohort AUC under $(\varepsilon=1)$-DP pooled training, our privacy motivation is weakened and we will say so.
Why not just bootstrap the decomposition for uncertainty? Resampling NMF or PCA factors is the standard cheap answer for confidence intervals. It fails here for a structural reason the entrywise literature has documented: bootstrap intervals inherit the independence assumptions that metagenomic count matrices violate through compositionality, >90% sparsity, and batch coupling — the same assumption gaps flagged as open problems for leave-one-out analysis (Spectral Methods for Data Science: A Statistical Perspective) — and rare translocated species, the ones that matter clinically, are the rows where resampling degenerates. Our spike-in experiment will test this directly: if bootstrap intervals achieve nominal 95% coverage under planted batch correlation at realistic sparsity, the bootstrap wins and our theory investment is unjustified for practice.
Why not just reuse the existing privacy-aware low-rank guarantees? The Dyson Brownian Motion and contour-bootstrapping results already bound the error of noise-injected low-rank estimation (Private Low-Rank Approximation for Covariance Matrices via Dyson Brownian Motion; Spectral Perturbation Bounds for Low-Rank Approximation with Applications to Privacy), so a reviewer could ask why the machinery needs extension at all. The answer is the central argument of entrywise analysis itself: those bounds are Frobenius or spectral norm, and an aggregate bound tolerates one catastrophically wrong coordinate — one false or missed species call — while averaging out (Spectral Methods for Data Science: A Statistical Perspective). Our calibration ablation is the empirical test: if Frobenius-calibrated thresholds produce the same false-call rate as entrywise-calibrated ones on rare taxa, the extension is not needed; we expect them not to, because that is precisely the regime the existing theory does not cover and its authors mark as open.
Why not just collect paired samples instead of separating one channel? This is a genuine threat. The translocation study obtained ground truth by sampling the lower respiratory tract alongside plasma (Microbial lung-to-blood translocation associates with systemic inflammation in severe pneumonia: evidence from paired plasma and lower respiratory tract metagenomics), and invasive sampling sidesteps the separation problem entirely. But invasive sampling is exactly what fragile ICU patients, screening contexts, and longitudinal monitoring cannot bear, and a single-channel method with calibrated per-call confidence can triage who warrants the invasive confirmation. We will settle this quantitatively: on the paired cohort, we measure how much attribution accuracy our plasma-only calls give up relative to the LRT-anchored labels, treating that gap as the honest price of non-invasiveness rather than hiding it.
Proposed Experiments
Three data settings, in increasing order of synthetic control. First, the paired pneumonia cohort (Microbial lung-to-blood translocation associates with systemic inflammation in severe pneumonia: evidence from paired plasma and lower respiratory tract metagenomics): using LRT presence as ground truth, evaluate lung-origin calls per plasma species at a 10% false-discovery rate, plus AUC, against the 0.1% genus-level threshold; report sensitivity against the study's documented 41% translocation prevalence. Second, the four-cohort colorectal cancer dataset of 526 samples (Multi-cohort analysis of colorectal cancer metagenome identified altered bacteria across populations and universal bacterial markers): leave-one-cohort-out training on DP-noised pooled covariances at $\varepsilon \in \{1, 4, 8\}$, metric AUC against the reported SVM values of 0.80 (0.88 with covariates; 0.82–0.84 early stage). Third, fully synthetic planted-mixtures: sweep the true gap $\delta_C$ and noise scale $\sigma$ to map the empirical separability phase transition and check it against the conjectured threshold within a constant factor; measure 95% interval coverage for each species row under real versus complex privacy noise.
The decisive checks are four. The entrywise intervals must hold nominal coverage (within two percentage points) across the synthetic sweep, including the weak-gap region, while Frobenius-calibrated and bootstrap intervals degrade. The phase transition in $\delta_C$ must track the theoretical prediction. Complex noise must buy measurably smaller $\varepsilon$ at equal coverage. And on real data, the pipeline must match or beat the cross-cohort SVM AUC under privacy constraints while producing the compartment attributions no existing baseline provides. A cheap transfer check closes the loop to the separation literature: in the overlapping-spectrum regime where two-step NMF masks are documented to fail (A Two-step NMF Based Algorithm for Single Channel Speech Separation), our eigengap statistic should predict the failure — the same gating signal, validated in its home domain, that will tell a clinician when the blood is speaking too quietly to be unmixed.
References
- Microbial lung-to-blood translocation associates with systemic inflammation in severe pneumonia: evidence from paired plasma and lower respiratory tract metagenomics
- Spectral Methods for Data Science: A Statistical Perspective
- Private Low-Rank Approximation for Covariance Matrices, Dyson Brownian Motion, and Eigenvalue-Gap Bounds for Gaussian Perturbations
- Spectral Perturbation Bounds for Low-Rank Approximation with Applications to Privacy
- A Two-step NMF Based Algorithm for Single Channel Speech Separation
- Single-channel multi-talker speech recognition with permutation invariant training
- Private Covariance Approximation and Eigenvalue-Gap Bounds for Complex Gaussian Perturbations
- Entrywise limit theorems for eigenvectors of signal-plus-noise matrix models with weak signals
- Entrywise Recovery Guarantees for Sparse PCA via Sparsistent Algorithms
- Multi-cohort analysis of colorectal cancer metagenome identified altered bacteria across populations and universal bacterial markers
- Single-Channel Multi-Speaker Separation Using Deep Clustering