Modern Bioinformatics Method Decision Guide

Use this method decision guide to match an analysis approach to the question, experimental design, and information actually present in the data. It compares options for cell annotation, multimodal mapping, spatial neighborhoods, isoforms, trajectories, and donor-aware differential expression; these are candidates to evaluate, not a universal ranking or a runnable pipeline.

Format: Conceptual guide. For readers choosing an analysis strategy, this guide walks you through a design-first comparison and a short method-selection record; it supplies no dataset, and you should use your own study plan or a clearly hypothetical scenario.

Use this guide to produce: a short method-selection record naming the biological question, input modality and replicate structure, candidate method, expected computational output, and an independent check or limitation.

Scope: this is a methods overview, not an executable pipeline. Package interfaces and reference models change; confirm versions and input requirements in the linked documentation before analysis.

Start with the study design

Write down the unit of observation and the unit of replication before choosing software. A cell is an observation; in a condition comparison, the biological sample or donor is commonly the independent replicate. Record whether measurements are paired across modalities, whether cells have sample and condition labels, what reference annotations exist, and whether the assay contains the signal the method needs (for example, spliced and unspliced counts for conventional RNA-velocity workflows).

State the decision or estimand. Are you assigning labels, comparing conditions across donors, locating cells in tissue, or describing a possible state transition?
Check the input contract. Confirm modality, raw versus transformed counts, feature identifiers, sample-level metadata, reference compatibility, and any paired bridge data.
Select a candidate and a comparator. Prefer a transparent baseline when it answers the question; compare methods only when their outputs and assumptions are meaningful for the same data.
Plan an independent check. Use held-out labels or marker review for annotation, mapping diagnostics for integration, spatial null/permutation and tissue context for neighborhoods, and replicate-aware design checks for differential expression.
Record uncertainty. A predicted label, embedding, pseudotime, or enrichment statistic is an analysis result, not proof of cell identity, causal regulation, or temporal lineage.

Match the question to the method

Method comparison: minimum inputs, expected output, and key assumption
Method area and examples	Minimum inputs	Expected output	Key assumption or check
Cell annotation: marker review, CellTypist, or scGPT (annotation tutorial)	Query expression/features; for classifiers, a compatible reference or model and feature identifiers.	Candidate cell labels (and, where available, scores or uncertainty).	Reference/model covers the tissue and states; verify with independent markers and review uncertain or out-of-reference cells.
Multimodal mapping: Seurat bridge integration (CITE-seq/WNN tutorial)	Compatible unimodal reference and query plus paired multimodal bridge data and required assays/reductions.	Query mapped into reference context and transferred labels or anchors.	Bridge data connect the modalities; inspect anchors and query-side structure. Transferred labels are predictions.
Spatial neighborhoods: Squidpy (spatial workflow tutorial)	Spatial coordinates, annotated cells, and a declared graph/neighborhood definition.	Graph-neighborhood enrichment statistics.	Results depend on coordinate scale, graph, labels, and permutation null; association is not signaling or causation.
Isoform discovery: FLAMES, IsoQuant, Bambu, or scISOr-Seq (long-read sequencing guide)	Long-read transcript data, genome/reference annotation, and platform/library-protocol details.	Transcript structures and/or isoform abundance, as supported by the selected tool.	Tool, reference, and protocol must match the sequencing platform and single-cell/bulk design; inspect junction support.
Trajectory or velocity: trajectory methods guide or scVelo	Expression matrix and cell metadata; scVelo additionally needs suitable spliced/unspliced counts.	Inferred path/pseudotime or velocity estimates.	Topology, root, preprocessing, and (for velocity) splicing-kinetics assumptions matter; inferred order is not observed time or lineage fate.
Donor-aware differential expression: pseudobulk with DESeq2 or PyDESeq2 (pseudobulk tutorial)	Raw counts aggregated per biological sample and cell population, with condition and replicate/design metadata.	Sample-level condition effects and statistical uncertainty.	Biological samples/donors are replicates and the design must be estimable; cells from one donor are not independent replicates.

Worked decision: comparing conditions across donors

Suppose the question is whether expression differs between treated and control samples within a cell population. First identify independent donors or samples and confirm each has the relevant cell population. Sum the unnormalized gene counts within each sample and population, retaining a sample-level table that links each pseudobulk column to donor, condition, and any design covariates. A differential-expression model then compares those sample-level columns. Do not feed normalized cell-level expression into a count model or treat thousands of cells from a few donors as thousands of replicates.

Before fitting: count-matrix columns and metadata rows must describe the same pseudobulk samples in the same identifiers; counts must be non-negative integer-like raw counts. Check sample overlap and whether the design is estimable. If treatment is confounded with donor or batch, software cannot recover a separate treatment effect from that design.

The PyDESeq2 documentation shows a DeseqDataSet built from count and metadata dataframes with a design formula, followed by a contrast in DeseqStats. Exact parameter names differ by release, so treat this description as an input contract rather than a ready-to-run recipe. Inspect the fitted design, contrast direction, sample counts, and adjusted p-values before interpreting results. A significant association alone does not establish a mechanism.

Practice: write a method-selection record

Use this hypothetical planning prompt; it is not biological evidence or a dataset. A study has treated and control samples from multiple donors, cell-level RNA counts, and sample labels. The question is whether a named cell population differs by condition. Complete the record before selecting software:

Question and unit: define the population and condition contrast; identify donor/sample, not cell, as the replicate.
Inputs to verify: confirm raw count availability, population labels, donor and condition metadata, and sample-level replication. Note any missing or confounded labels.
Candidate/output: if those inputs support it, propose aggregation by sample and population followed by a count-model comparison; name the output as sample-level estimated condition effects and uncertainty, not a cell-level prediction.
Check and stop rule: verify one metadata row per pseudobulk column, matching identifiers, and an estimable design. Stop and resolve the design if condition is inseparable from donor/batch; do not treat additional cells from the same donor as added replicates.

A satisfactory record states what evidence is available and what remains unknown. It does not claim that the hypothetical study has a particular result.

Compare outputs without treating them as interchangeable

These methods answer different questions. Annotation transfers or predicts categories; bridge integration places query measurements in a reference context; spatial enrichment summarizes graph neighborhoods; isoform tools call and quantify transcript structures; trajectory methods arrange cells under assumptions; pseudobulk models test sample-level condition effects. An attractive UMAP, high model confidence, or small p-value is not a substitute for checking how the result was produced.

For labels, review marker evidence and unassigned/uncertain cells.
For mapped modalities, compare transferred annotations with independent query structure and known assay coverage.
For spatial results, report the coordinate/graph choices and permutation setup, and distinguish adjacency from mechanism.
For trajectories and velocity, test whether plausible alternate roots/topologies or preprocessing change the interpretation.
For differential expression, report biological replicate numbers, model formula, contrast, and multiple-testing procedure.