Datasets:

junidude14
/

PBISC-data

paper-releases/pbisc-diffusion/updates/20261012T130738Z-strictpairs-implementation/diffusion/dataset/server_1/model8/plans/paper_figures_strictpairs_20261012/archive/20261012T130738Z-strictpairs-implementation/snapshot/manuscript/sections/methods_paired.tex ADDED Viewed

	@@ -0,0 +1,12 @@

+\subsection*{Paired cohorts and exact-support splits}
+The five paired cohorts are healthy ImMicro\cite{babdor2026immicro}, sepsis\cite{ye2026sepsis}, Imperial ESKD\cite{gisby2022multiomics,stephenson2025temporal}, Giroux COVID-19\cite{giroux2022differential} and Ji ChILI\cite{ji2025chili}. The original identity-passed pair lists, assay masks and cohort contrasts are frozen before the strict split. Ji pairs whole-blood bulk with same-visit PBMC single cells. Imperial's undated COVID-negative pairs match patient and status but cannot be proved to be the same blood draw. Each held-out sample is scored on its cohort's assay-measured genes; Giroux disease contrasts exclude sex-linked scoring genes.
+In ImMicro, five outer folds are fixed. Within each fold, nested training supports have $n=5,10,20,40$ independent donors; the random seed is $20261012+\mathrm{fold}$. The strict five-pair setting uses one sample from each of five donors. No additional paired single-cell samples are used for validation. LoRA receives a fixed 3,000 steps and EMA weights. scSemiProfiler's representative set and highly variable genes are derived from these $n$ samples only. Paired gene calibration also uses only the same $n$ samples. The AIFI reference-calibration alternative uses external atlas single cells and all unlabelled target bulks, without their cell labels or observed target single cells.
+For each disease cohort, the five previously selected donor draws remain fixed. For each selected donor, the earliest identity-passed visit in the matching stratum supplies one pair. Other visits of selected donors are excluded from both training and test. LoRA, paired calibration, scSemiProfiler and reference-based deconvolution receive the same five matched samples. All usable visits from unselected donors form the test side. Every arm uses the same test samples and scoring-gene mask within a draw. Split/support manifests bind sample and donor identities and their hashes to the score receipts.
+\subsection*{Conditioning and generation}
+A measured bulk is CPM-normalised over its measured genes and scaled to $K_{\mathrm{cond}}=1{,}000$ cells at the fixed pretraining reference depth $\rho=3{,}869.603$ counts per cell. A native pseudobulk instead sums the sampled raw cell counts, with nominal $K$ equal to its member count and depth $L/K$; the study never substitutes the measured-bulk constant for that native depth. The number of output cells, $N_{\mathrm{gen}}=1{,}500$, is independent of $K_{\mathrm{cond}}$. Inference uses 50 DDIM steps and the drawn support read-out.
+\subsection*{Separate historical five-fold analysis}
+Earlier five-fold LoRA, full-parameter and rank-sweep scores are retained only as labelled sensitivity analyses. Some old few-pair settings used extra inner-validation pairs, and earlier disease five-donor caches could include multiple visits per selected donor. Neither design is an exact-five-pair estimate. Their source tables and old manuscript paragraphs are preserved in the draft's editorial legacy notes for audit, not pooled with strict estimates.

	@@ -0,0 +1,5 @@

+\subsection*{Population pretraining and five paired donors connect bulk to single cells}
+\method{} learns a cell's expression and detection support in gene space from its sample's pseudobulk (Fig.~\ref{fig:fig1}). The pretraining corpus contains 41,395,184 cells from 22 datasets, 3,199 donors and 6,507 samples; 33,141,380 cells belong to the training split. The assay-measurement mask separates an unmeasured gene from a measured zero. Corpus cell-type, condition and health labels describe the data and are not model inputs.
+For measured bulk, adaptation uses one matched bulk--single-cell pair from each of exactly five donors. The principal path trains rank-16 LoRA for 3,000 fixed steps and then draws 1,500 cells for a test bulk. The nominal input cell count $K_{\mathrm{cond}}=1{,}000$ and fixed reference depth $\rho=3{,}869.603$ counts per cell determine the condition; $N_{\mathrm{gen}}=1{,}500$ is only the requested output cell count. A pretraining pseudobulk instead obtains its $K$ and depth from the sampled cells. The 50-step sampler draws independent starting noise and support per requested cell. It has no cell-type labels or sum-to-bulk constraint. An external AIFI single-cell reference can supply a calibration alternative; its performance remains to be measured under the strict support design.

	@@ -0,0 +1,5 @@

+\subsection*{Pretraining design and paired-support ablations}
+The architecture and gene-panel ablations were rescored on a held-out 42-donor SLE target using a common 1,006-gene axis, 100 distinct real reference cells and 100 distinct real query cells per donor, a fixed Poisson read-out and local $k=10$ neighbourhoods (Fig.~\ref{fig:fig2}). These are descriptive scores from saved epoch-16 outputs; the 42 donors are the biological units. The architecture family used a 1\% training detection floor and the panel family a 0.1\% floor, so their scores are separate experiments. The baseline architecture had coverage 0.7812 and precision 0.7800; the four-times-width denoiser had coverage 0.8283 and precision 0.7136. Within the panel family, the 6,000-gene model had coverage 0.7812 and precision 0.7398; the 15,114-gene model had coverage 0.7429 and precision 0.8331. These trade-offs do not establish one arm as best on both axes.
+Historical healthy LoRA rank and full-parameter comparisons used the earlier five-fold design and remain a separate sensitivity analysis. The new nested $n=5,10,20,40$ exact-support curve and the external AIFI reference-calibration arm await complete scoring on their frozen folds. Their empty plotting slots carry no performance claim.

	@@ -0,0 +1,3 @@


1	+ \subsection*{Healthy-donor generation under exact five-pair support}
2	+
3	+ The planned healthy evaluation uses five held-out ImMicro folds. In each fold, one sample from each of five unique training donors supplies LoRA, paired calibration and scSemiProfiler; the remaining donors supply real bulk at inference and observed cells only for scoring (Fig.~\ref{fig:fig3}). Precision, coverage, composition error, classifier AUC, MMD$^2$ and $E_N$ will be reported together with real-reference UMAP and marker displays after the five strict folds pass the shared scoring and provenance checks. Operational time will include adaptation and generation for the target cohort and exclude the supplied population pretraining run. \todo{Five strict folds and time receipts are pending; no comparison is concluded here.}