v0.0.3.0001

tinydenseR 0.0.3.0001

Framework overview

tinydenseR is a clustering-independent framework for sample-level modeling of
single-cell data. It constructs a landmark-by-sample fuzzy density matrix from
UMAP-derived cell–landmark connection strengths and supports four downstream
analysis modes within a single analytical scaffold:

1. Differential cell-state density analysis — landmark-level linear
   modeling of density shifts across conditions
2. pePC (partial-effect Principal Component projections) — supervised
   quantitative sample embedding that isolates variation attributable to a
   specified effect
3. plsD (graph-diffused, density contrast-aligned partial least squares
   Decomposition) — multivariate feature interpretation identifying
   expression programs aligned with density contrasts
4. Pseudobulk differential expression — manifold-informed, connection-
   strength-weighted pseudobulk aggregation with design and marker modes

S4 TDRObj class

The data container is a formal S4 class with 12 slots, providing a
structured representation for landmark-based single-cell analysis objects.
A backward-compatible $ accessor preserves legacy syntax
(obj$pbDE, obj$markerDE, etc.) while all internal code uses @-access.

Multi-backend support

All analysis and plot functions dispatch via S3 generics with methods for
multiple input formats:

Backend	Format
Seurat	v4, v5 (on-disk or in-memory)
SingleCellExperiment	On-disk or in-memory
H5AnnData	On-disk .h5ad files (via anndataR / rhdf5 + BPCells)
BPCells	IterableMatrix (on-disk)
dgCMatrix	Sparse matrix (in-memory)
DelayedMatrix	Converted to BPCells for on-disk efficiency
flowSet / cytoset	Flow, mass, and spectral cytometry (flowWorkspace)

• RunTDR(): Unified entry point that sets up a TDRObj from any
  supported container.
• GetTDR() / SetTDR(): Extract / embed the TDRObj within a
  container.
• S3 methods follow a three-step pattern: extract → compute on TDRObj →
  re-embed, so all functions work identically on any backend.

Manifold-informed pseudobulk DE

get.pbDE() supports two analysis modes:

Mode	Purpose
design	Pseudobulk DE across experimental conditions
marker	Marker identification comparing cell populations within samples

Pseudobulk aggregation uses connection-strength-weighted sums over landmark
neighborhoods rather than hard cluster or cell-type membership, yielding a
manifold-informed pseudobulk representation while retaining samples as the
unit of inference.

plsD: graph-diffused partial least squares decomposition

get.plsD() identifies graph-smoothed expression programs maximally aligned
with a fitted density contrast. Key properties:

• Sparse-implicit NIPALS PLS1 implementation (no dense intermediates)
• Concordance-filtered and raw loadings
• Post-hoc sign alignment with density contrast
• Designed for interpretation and hypothesis generation, not formal inference

celltyping: multi-solution support

celltyping() accepts two input modes:

Mode	Input type	Semantics
A	Named list	Cluster-to-label mapping
B	Named character vector	Per-cell labels for the full dataset

Multiple named celltyping solutions can coexist. Switch active solutions with
set_active_celltyping() and list available solutions with
list_celltyping_solutions().

Reclustering and multi-resolution workflows

recluster() re-runs Leiden community detection with new parameters and
refreshes all clustering-dependent downstream slots. Multiple clustering
solutions can be stored and switched via set_active_clustering().

New features

• get.embedding(): Sample-level embeddings (PCA, diffusion-map
  trajectory, and pePC) from the landmark-by-sample density matrix.
• get.cellmap(): Accessor for per-cell mapping data by sample and slot.
• sim_trajectory_tdr: Bundled simulated scRNA-seq dataset with
  condition-dependent differential abundance for examples.
• simulate_DA_data() / simulate_DE_data(): Generate synthetic
  cytometry datasets for benchmarking.

Vignettes

• Clinical scRNA-seq analysis: End-to-end workflow with on-disk data.
• Clinical cytometry analysis: Flow cytometry with longitudinal design.
• Multi-backend workflows: Demonstrates all supported input formats.
• Package architecture: Object structure and analysis pipeline overview.

---

tinydenseR 0.0.2.0001

On-disk caching for get.map()

get.map() supports on-disk caching (.cache.on.disk = TRUE, default),
serializing large per-sample data to disk to reduce peak RAM usage for
large datasets. All downstream functions read from cache transparently.

Cache management: tdr_cache_info(), tdr_cache_validate(),
tdr_cache_cleanup().

tinydenseR 0.0.1.0013 — Legacy API (preprint‑compatible)

Summary
This first public release captures the legacy API of tinydenseR used in the preprint so that results remain reproducible and citable. The package provides a landmark‑based, sample‑centric workflow for single‑cell analysis with built‑in visualization and support for scRNA‑seq and cytometry data.

Why this release

• Ensures a stable tag that matches the preprint for benchmarking and reproduction.
• Provides a clear baseline before introducing the upcoming, redesigned API.
• Establishes semantic versioning for future development.

Highlights

• Sample‑centric modeling: treats samples (not cells) as biological replicates for robust statistics.
• Scalable & memory‑efficient: suitable for atlas‑scale datasets.
• Technology‑agnostic: works with scRNA‑seq, flow, and mass cytometry.
• Built‑in plots & vignettes to explore landmark geometry, differential abundance/expression.

API status

• This tag preserves the legacy API used by the preprint.
• Breaking changes are planned for the next release (new API).
• No deprecations were introduced in 0.0.1.0013; the legacy functions remain as‑is.

Compatibility

• Requires R ≥ 4.1.
• Tested on Linux and macOS with typical single‑cell pipelines.

Installation

# Install the legacy API exactly as used in the preprint:
if (!require("devtools")) install.packages("devtools")
devtools::install_github("Novartis/tinydenseR@0.0.1.0013")

Documentation & examples
See the repository README and vignettes for quick‑start examples, background, and detailed workflow description. Novartis/tinydenseR GitHub

Known limitations

• Expect breaking changes in naming/arguments in the next release to streamline the workflow.

Citation
If you use tinydenseR, please cite the preprint:
“Sample‑level modeling of single‑cell data at scale with tinydenseR” — bioRxiv, DOI 10.1101/2025.11.26.690752. View on bioRxiv

Acknowledgments
Thanks to the contributors and reviewers within Novartis Biomedical Research and the single‑cell community for feedback leading up to this release.

Full Changelog: https://github.com/Novartis/tinydenseR/commits/0.0.1.0013