Paper Figure Standards: Captions, Units, and Error Bars—Which Parts Does the Office Assistant Agent Handle?
The comment that appears most often in peer-review reports is: What do the error bars in Figure 3 represent? I revised the figures four times, and only on the last round did I understand: half of paper figure standards is the “data-to-figure” work, which tools can help with; the other half is the “figure-to-conclusion” work, which tools cannot help with.
I. First turn data into a table, then into a figure.
My figures all start from a data table: what quantity is on the x-axis, what quantity is on the y-axis, whether each point is a mean, how the error is calculated, and how many samples there are. If the table is incomplete, the figure will inevitably lack information. I use MuchSci’s Office Assistant Agent to organize this table into an editable Excel file (describe the need in one sentence, and it produces editable Word / Excel / PPT), filling in column names, units, and sample counts clearly in one go; after that, every time I revise a figure I only change this one table, not redraw it. The units in figures and tables should be written correctly in the header from the start, not added temporarily on the figure.
II. Figure captions must be understandable on their own.
The standard is: if you extract the caption on its own, you can tell what is plotted without reading the main text. It should clearly state what the symbols mean, what the error bars represent (standard deviation, standard error, or confidence interval; different journals have different requirements for how to write this), the number of replicates, and the statistical test. If the target is an English-language journal, I first use MuchSci’s Paper Translation to produce a draft of the caption (bidirectional academic translation between Chinese and English, preserving formulas and terminology consistency), then check it sentence by sentence—inconsistency between terminology in the caption and the main text is the thing reviewers are most likely to notice at a glance.
III. “The parts it doesn't handle” are here.
· What the error bars represent: standard error and standard deviation differ by severalfold, and how the caption is written depends on your statistical basis. This is determined by experimental design and statistics; the tool does not know which one you calculated.
· Whether group A and group B can be plotted together: this depends on whether the two groups are from the same batch and the same conditions. You can only judge this yourself.
· Whether the trend can be stated this way: connecting four points into a line and saying it “increases monotonically with increasing temperature” is overinterpretation. The figure can be drawn, but the conclusion cannot be stated that way.
· Which figure should go in the main text and which in the supplementary material: this depends on whether it supports the main argument or supporting evidence.
· Pixel-level specifications: line width, font size, dpi, vector or raster, colorblind-friendly color schemes—these are hard requirements of the journal and must be changed according to the Author Guidelines; tools cannot do this for you.
IV. Figure version management.
A figure goes through “draft sketch → first draft → revision version → final version,” and file names do not tell the truth on their own. I put each figure’s vector source file and exported images together in MuchSci’s File Asset Library (centralized storage for generated documents, images, and text, with preview and add/delete/edit/query support), and write the data basis marked on the figure together with the corresponding data table into a text description; when showing figures to my advisor, I send them via “My Shares” so they don’t take up chat history.
V. Where does the data in the figure come from.
For calculations such as compound descriptors and molecular property prediction, I use MuchSci’s Cheminformatics Agent (its toolchain includes RDKit, Datamol, PyTDC, MedChem, and TorchDrug, and it can perform molecular operations, descriptor calculations, and GNN prediction). During its run, I can see progress and interrupt at any time; at key decision points, a decision card pops up for me to confirm—this matters more to me than “producing a pile of numbers”: I need to know how each number is calculated before I dare to mark it on a figure. The calculated results, along with the original literature used for verification, are archived together in MuchSci’s Personal Knowledge Base (supports PDF, DOCX, PPTX, XLSX, TXT, and MD; personal use by default and not shared).
Three boundaries: a figure cannot beautify a conclusion—erasing outliers and truncating error bars are not formatting issues but academic integrity issues; color schemes and axis ranges affect readers’ judgment—shrinking the vertical axis can make changes look more obvious, and I would not do that; results from computational tools must be reproducible—record parameters and versions together, otherwise six months later even you will not be able to explain how that figure was made.
What this kind of graduate student AI assistant does in the figure-and-table step is: organize data into tables, smooth out the caption translation, and manage files well. How that line on the figure is drawn is still my own decision.