AI Figures for Single-Cell Biology: UMAP Plots, scRNA-seq Workflows, and Trajectory Analyses (2026)

It is Saturday afternoon. A postdoc has the same UMAP open in three windows — the Scanpy notebook, an exported PNG, and Illustrator — and is on cluster #11's fourth color in twenty minutes. The biology in the figure is correct. The science is finished. What is left is the slow, manual work of making it look the way a Nature Methods reviewer expects a single-cell figure to look. Anyone who has been a corresponding author on a scRNA-seq paper knows this Saturday.

This post is about how that Saturday changes when an AI scientific-figure tool sits next to the analysis stack. We focus specifically on single-cell biology figures — UMAP and t-SNE embeddings, dot plots and heatmaps of marker genes, pseudo-time trajectories, RNA-velocity streamlines, cell-cell communication diagrams, and the experimental-workflow schematics that always end up as panel A. We map each figure type to a concrete workflow, name the journal conventions that actually get figures rejected at production, and call out where AI is still the wrong tool.

Key takeaways

Single-cell figures live or die on a few specific conventions — UMAP cluster labels at centroids, viridis-family palettes, 7 pt minimum type, vector output at column width. AI tools that respect those conventions save a Saturday; tools that don't add one.
The figures you keep redrawing are a short list — embeddings, marker dot plots, trajectories, communication diagrams, workflow schematics. Each maps to a different LabFig workflow.
For your real dataset's UMAP geometry, generate the points from Scanpy and let AI build the surrounding figure schema (panels, labels, legends, workflow art). AI is not a substitute for re-running your analysis.
Nature Methods and Cell production teams care about line weights, font, and color choice far more than about novelty. Get those right and the figure is accepted; get those wrong and it bounces from production.
AI helps most where the visual grammar is repetitive across papers — workflow art, communication diagrams, multimodal integration schemas. It helps least where the figure encodes your specific data points.

What single-cell figures actually have to communicate

A single-cell figure has more communicative load per square inch than almost any other figure in biology. In one panel you might need to show: high-dimensional structure compressed into 2D, the identity and proportion of every cell type, marker-gene expression as a third channel, the direction of a developmental trajectory, and — between panels — how this all connects to a model. Five layers of information in a 90 mm column.

Reviewers at Nature Methods, Cell, and Nature Cell Biology have internalized a set of visual conventions that compress this load:

Cluster identity is conveyed by color, anchored by labels near the cluster centroid, not pushed into an external legend if it can be avoided.
Marker expression is shown in a separate panel — usually a dot plot or heatmap — so the UMAP itself stays uncluttered.
Trajectory direction is shown either by a smooth pseudo-time gradient overlaid on the embedding, or by arrows from RNA-velocity, but rarely both at full opacity in the same panel.
Cell-cell communication lives in its own circular or chord diagram, where the cell type ordering and ribbon weighting carry the meaning.

If you have never sat across the desk from a Nature-family production editor, the rule of thumb is this: the figure must be readable at 50% of column width on a phone screen, in print, and grayscale. That is what every convention above is ultimately solving for.

The figure types you keep redrawing

If you audit a typical scRNA-seq paper, the figure inventory is short and predictable. Here is the working set, in roughly the order they appear in most papers.

UMAP, t-SNE, and PHATE embeddings

The embedding panel is the single most reproduced figure type in single-cell biology. The visual unit is a scatter plot of tens-of-thousands of points, each colored by cluster, with cluster labels placed near the centroid and a restrained categorical palette (often 8–20 distinguishable colors). The axes are usually unitless and lightly drawn — many papers strip the tick numerals entirely, leaving just the axis stubs.

A UMAP scatter plot on a white background showing approximately 400 small data points organized into 6 color-coded clusters in a restrained ColorBrewer-style palette of muted teal, sand, slate, plum, ochre, and dusty pink; thin axis tick marks on the lower-left form labeled UMAP1 and UMAP2 axes, with a small compact color-coded cluster legend tucked in the upper-right corner of the panel.

For the data itself, generate the points from your Scanpy or Seurat object — no AI tool will recover your dataset's cluster topology. For the surrounding figure (labels, legends, multi-panel framing), AI is genuinely useful, especially when you have ten of these to produce in matching styles.

Dot plots and heatmaps of marker expression

The companion to every UMAP is the marker-expression panel. Two formats dominate: the dot plot (genes × cell types, dot size = fraction expressing, dot color = mean expression) and the heatmap (cells × genes, often grouped by cluster, with a viridis or blue–white–red scale). Both are dense, both have strict typographic requirements, and both look terrible if a generic image generator tries to invent the data.

Use the AI tool for the figure shell — axis labels, legend formatting, cluster annotation bars — and import the data layer from your analysis. The two cleanly separate.

Pseudo-time trajectory plots

Monocle, Slingshot, and PAGA produce a similar visual: an embedding colored by a continuous pseudo-time gradient, often with the inferred trajectory drawn as a smooth curve or graph on top. Branching trajectories add their own conventions (forks labeled with terminal cell states, often with a small inset of the branching tree).

For trajectory figures the dominant production cost is not the science — it is composing four sub-panels (embedding, velocity field, gene heatmap, branching tree) into a single legible figure. That composition is exactly where AI for scientific figures saves an afternoon. If you also work with deep-learning-based trajectory inference (scVI, scVelo, CellRank), the same compositional pattern applies — and the post on AI figures for machine learning papers covers the architecture-side conventions.

RNA-velocity streamlines

scVelo's signature streamline plot — thin arrows draped across the embedding — has its own typographic identity. The arrows should be uniform in thickness, the streamline density should not obscure the underlying clusters, and the color of the streamlines should contrast with the cluster palette. Most poor velocity figures fail one of those three.

Cell-cell communication chord diagrams

CellChat and NicheNet outputs typically end up as circular chord diagrams: cell types arranged around a circle, with ribbons connecting senders and receivers, ribbon width encoding interaction strength.

This is one of the figure types where AI helps the most. The chord-diagram layout follows a regular convention, but assembling it in Illustrator from a CellChat CSV export is slow. A prompt or sketch describing the cell-type ordering, the ribbon weights, and the inset gets you 80% of the way there.

scRNA-seq experimental workflow schematics

Almost every single-cell paper has a panel A that walks the reader through the experiment: tissue source → dissociation → library preparation (10x, Smart-seq, Drop-seq) → sequencing → analysis. These workflow schematics are the lowest-information panels in the figure but they are usually what reviewers see first.

Multimodal integration schemas

Multimodal papers — scATAC + scRNA, CITE-seq, spatial transcriptomics + scRNA — need a schema that shows how two or more modalities feed a joint analysis. The visual convention is typically a layered diagram, with each modality on its own track converging onto a shared embedding or integrated representation.

Mapping each figure type to a LabFig workflow

Each figure type above maps cleanly to one of LabFig's input modes. The skill is matching the input you have to the workflow that needs the least manual cleanup. For scRNA-seq and other life-science panels, the dedicated biology figure maker defaults to the viridis-family palettes and stroke weights these journals expect.

Text to Figure for workflow schematics and multimodal integration diagrams. These are conceptual figures where the visual structure follows a fixed convention (tissue → library → sequencing → analysis). A short prompt to the text-to-figure tool describing the steps and the order produces a strong first draft. This is the workflow where AI is most clearly faster than starting from a blank Illustrator artboard.
Sketch to Figure for whiteboarded study designs. When you are still thinking about cohort design — tissue source × condition × time point × readout — sketch it. The tool preserves the block layout and arrow direction and tidies the geometry. This is the workflow that wins when you are pre-submission and the study schematic keeps changing.
Reference to Figure for matching a published paper's visual family. Provide a target paper's figure (yours or one from the journal you are submitting to) as a style anchor and generate new content in the same visual language. For single-cell papers this is how you get a consistent palette and stroke weight across UMAP, dot plot, and chord diagram in the same figure.
PDF to Figure for redrawing a reference workflow. If you are writing a review, or adapting a methods section from an older paper where the source files are lost, lift the workflow figure out of the PDF and rebuild it as an editable version. The same approach works across disciplines — see the analogous use in AI figures for clinical AI papers where cohort-flow CONSORT diagrams play the same role.
Figure Enhancer for inheriting a lab's older scRNA-seq figure. Every lab has the senior author's 2019 figure that has been recolored across four papers since. Run it through an enhancer to normalize the stroke weights, modernize the palette to viridis-family, and replace the typography — without changing the underlying science.

The general principle: let AI handle the figure shell, generate the data layer from your analysis, and compose in the vector canvas. The orientation post on what AI for scientific figures actually is covers the canvas model in more depth.

Nature Methods / Cell conventions worth knowing

Single-cell papers are read closely by production editors. A figure can be scientifically excellent and still bounce at the production stage. The conventions below are the ones we see flagged most often.

Column widths and resolution. Nature single-column is 89 mm, double-column 183 mm. Cell figure guidelines specify 300 dpi at print size for combination figures, 500 dpi for line art. A flat raster at unspecified resolution will not survive production. Vector output (SVG, PDF) sidesteps the whole problem.
Line weights at 300 dpi. The convention is roughly 0.5–1 pt for most strokes, with axis lines on the thinner side. Anything heavier than 1.5 pt at column width looks amateurish in print. AI generators sometimes overshoot here by default — check it before export.
UMAP labelling convention. Place cluster labels near the centroid of the cluster, not in a separate legend, if there is room. This is the visual signature of a single-cell figure done well. A legend appears only when centroid labels would overlap or when too many clusters crowd the space.
Color palette ethics. Use ColorBrewer, viridis, or a comparable perceptually uniform palette. The jet rainbow palette is considered outdated and is explicitly discouraged in modern figure-design guidance because it introduces false structure in the gradient and is unfriendly to grayscale and colorblind readers.
Font sizes and family. Helvetica or Arial, 7 pt minimum at print size. Italicize gene symbols (per HGNC convention) and keep cell-type names roman. Mixing fonts across panels is a fast way to look like a draft instead of a final.

These are the kinds of details that production teams catch. They are also the details that AI tools either get right by default or get aggressively wrong by default — there is rarely a middle ground. Pick a tool that defaults correctly.

Limitations: where AI struggles with single-cell figures

AI for scientific figures is genuinely useful in 2026, but there are sharp edges in the single-cell context worth naming.

It cannot recover your dataset's cluster topology. Your specific UMAP — the exact shape of cluster 7, the bridge of cells between cluster 4 and 11, the rare population that justified the paper — is data, not visual structure. AI cannot invent it correctly. Always generate the points from your Scanpy or Seurat object and let AI build the surrounding figure.
Biologically accurate gene-set selection in heatmaps. If you prompt for "marker genes for CD8 T cells," the output may include plausible but non-canonical genes. Use your differential-expression results, not the model's guess, for any gene list that appears in the final figure.
Field-specific manual annotation is hard to replicate. The way a senior single-cell biologist annotates "this is a transitional state, this is a doublet, this is a subset" is hand-curated. AI will not infer those judgments from a UMAP shape — you supply them.
Reproducibility across regenerations. Regenerating the same prompt twice gives slightly different output. Lock the structure early, then edit in the vector canvas instead of re-prompting. This is the same lesson as in AI figures for drug discovery, where structure-activity panels also reward editing over regeneration.

Treat every AI-generated single-cell figure as a draft. The verification step — comparing the figure to your underlying data — is yours.

A workflow for your next paper

This is the workflow we recommend if you are preparing a single-cell figure for a journal submission, regardless of which AI tool you use.

Start with the figure schema, not the panels. Decide that figure 1 is panels A (workflow), B (UMAP), C (marker dot plot), D (trajectory), E (cell communication). Sketch this on paper or describe it to the AI tool in one short prompt. Lock the schema before generating any single panel.
Generate the UMAP separately and import. Run Scanpy or Seurat, export a high-resolution scatter (PNG or vector), and place it into panel B. AI should never invent your data points.
Use Reference to Figure to align style across panels. Pick one panel as the visual anchor (often the workflow schematic, since it has the most stylistic surface area) and use it as a style reference for the others. This gets you consistent palette and stroke weight across panels A, D, and E.
Export SVG and finalize in Vector Canvas. Pull the assembled figure into the vector canvas to align labels, set typography to 7 pt Helvetica, and verify the palette is viridis-family and not jet. This step catches the production-stage problems before they bounce.
Disclose AI involvement in the methods section. A one-line disclosure in methods or figure caption — "Schematic in Figure 1A was drafted with AI-assisted scientific illustration and edited manually by the authors." — is standard practice in 2026.

The most expensive mistake is generating each panel independently and stitching them together at the end. The style will not match and you will spend longer in Illustrator than you would have without the AI in the first place.

FAQ

Can AI generate a clustered UMAP for my specific dataset? No — and any tool that claims it can is fabricating data. Your UMAP cluster shape is a result of your data and your dimensionality reduction parameters; it is not a stylistic choice. Run Scanpy or Seurat, export the scatter, and let the AI tool build the surrounding figure (labels, legend, multi-panel framing, workflow art around it). This is the correct division of labor.

How do I make my figures match a Nature Methods paper's style? Use a Reference to Figure workflow. Provide a figure from a recent Nature Methods paper as the style anchor and generate new figures against it. You will get matching stroke weights, palette, and typography. Always cite the original paper and use the reference only for style, not content. The Single-cell best practices community also publishes figure-style examples from recent papers that are useful as style anchors.

Are AI-generated figures accepted at Cell or Nature Methods? As of mid-2026, none of the major journals — Nature Methods, Cell, Nature Cell Biology — prohibit AI-assisted figure production. They evaluate the figure on technical specification (resolution, color, typography) and scientific validity, not on the production method. Most journals do expect a one-line disclosure in the methods or acknowledgements section. Read your target journal's instructions to authors before submission.

What's the best way to generate a cell-cell communication chord diagram? Export the interaction matrix from CellChat or NicheNet as CSV, then use Text to Figure with a prompt describing the cell-type ordering and the ribbons you want to emphasize. For the ribbon weights themselves, import the values from the CSV — do not let the model invent them. The chord diagram's geometry is easy for AI; the underlying numbers must come from your analysis.

How should I handle color choices for accessibility and colorblindness? Use perceptually uniform palettes — viridis, plasma, magma for continuous data; ColorBrewer "Set2" or "Dark2" for categorical. Avoid jet, rainbow, and red–green diverging palettes. Test the figure in grayscale and with a colorblindness simulator before submission. ColorBrewer and matplotlib's viridis documentation both flag colorblind-safe palettes explicitly.