AI Figures for Drug Discovery: Target Pipelines, Binding Sites, and ADMET (2026)

It is 9pm on a Tuesday and a medicinal chemist is staring at a spreadsheet with 200 SAR data points, a project review at 9am, and a half-finished 6×6 SAR matrix in PowerPoint that nobody on the team can read. The compounds are right, the assays are right, and the figure is the only thing standing between this program and a green-light decision. Anyone who has lived this evening knows that the figure tool is the bottleneck, not the chemistry.

This post is for that chemist — and for the biologists, modelers, and DMPK scientists alongside them. We walk through the figure types that drug-discovery papers actually rely on, the visual conventions that journals like Cell, Nature Chemistry, and Nature Reviews Drug Discovery quietly enforce, and how to use AI figure tools like LabFig to draft these figures without losing the chemical accuracy that medicinal chemistry demands. If you have not yet read our orientation post on what AI for scientific figures is, start there — this piece assumes that baseline.

Key takeaways

Drug-discovery papers lean on a small, predictable set of figure archetypes: target schematics, binding-site renderings, virtual screening funnels, SAR grids, ADMET radars, lead-optimization timelines, and PK curves. AI tools should compress the time to first draft on every one of them.
Vector output is non-negotiable. Chemical structures, dashed H-bond lines, and tight stereochemistry only survive at production resolution if the figure is vector-native (SVG, PDF) — not a PNG export.
AI is a layout engine, not a chemistry engine. Treat generated molecules as placeholders. Always paste real structures from ChemDraw, RDKit, or a verified source into the canvas. A hallucinated stereocenter or transposed substituent will get you flagged in review.
The strongest workflows pair AI for schematic context (pipeline, pocket diagram, screening cascade) with deterministic tools (ChemDraw / RDKit) for the molecules themselves. Composite figures, not pure generations.
Journals do not ban AI-assisted figure production for drug-discovery work, but disclosure is increasingly expected at Cell, Nature Chemistry, and Nature Reviews Drug Discovery. Check the journal's instructions before you submit.

What drug-discovery figures must communicate

A drug-discovery figure is not decoration. It is the visual contract between the experimental data and the reader's ability to evaluate a program. Across the literature, six communication jobs recur:

Target biology. Why this protein, this pathway, this disease — and what modality (small molecule, biologic, PROTAC, oligonucleotide) the program is pursuing.
Binding mode. How the ligand sits in the pocket — key hydrogen bonds, hydrophobic contacts, solvent-exposed vectors. Reviewers read this figure first.
Chemical structure. The compounds themselves, drawn to ChemDraw / RDKit conventions — uniform bond lengths, correct stereochemistry, IUPAC-aligned labelling.
Screening funnel. How a library of millions became 50 hits, became 5 leads. The figure has to honestly show attrition.
ADMET trade-offs. Where this series is good (potency, selectivity) and where it is not (clearance, hERG, solubility). Radar plots and heatmaps make this comparable across analogs.
Structure-activity relationships. Which substituent shift moved which property — and by how much.

Each of these has a visual convention. Cell figures lean dense and panelled — a 2×2 or 2×3 grid with shared axes and panel letters (a, b, c). Nature Chemistry favors a more restrained, chemistry-first composition with prominent structures and inset mechanism arrows. Nature Reviews Drug Discovery, as a review journal, prefers cleaner conceptual schematics — pipelines, modality landscapes, regulatory timelines — over raw data panels. Knowing which family of conventions you are targeting before you generate anything cuts revisions in half.

The figure types you keep redrawing

If you have shipped two or three medicinal chemistry papers, you have already redrawn each of these figures at least once. This is the high-frequency set worth investing tooling time in.

Target identification and validation schematics

A clean figure that walks from a disease phenotype, through a pathway, to the specific target protein and the chosen modality. Usually a horizontal flow with three to five steps and labelled arrows. Reviewers want to see why this target in one glance.

Ligand–target binding-site visualizations

Two formats coexist: a 2D ligand interaction diagram (LigPlot-style — the ligand at center, residues around it as labelled boxes, hydrogen bonds as dashed lines, hydrophobic contacts as eyelashes) and a 3D pocket rendering (PyMOL or ChimeraX, surface or cartoon, ligand as sticks). The strongest papers show both, often with a zoomed inset.

Virtual screening workflow and funnel figures

A vertical funnel with library size at the top (often 10^6 – 10^9 compounds), narrowing through pharmacophore filters, docking, MM/GBSA rescoring, ADMET filters, and ending at a small triaged set. Bonus points for showing a chemical-space PCA or t-SNE colored by cluster.

Hit triage and hit-to-lead matrices

A grid where rows are compounds and columns are assays (biochemical IC50, cell IC50, selectivity panel, microsomal stability, permeability, solubility). Color-graded cells let a reviewer eyeball which hits cleared each gate.

ADMET radar and spider plots

Six- to eight-axis radar plots for a small set of analogs, often overlaid. The axes typically include potency, selectivity, hepatic clearance, permeability, solubility, hERG, and PPB. Recent trend: smaller, multi-panel small-multiples of radars rather than one over-plotted radar — easier to read in print.

Structure-activity-relationship (SAR) tables and grids

The classic medicinal chemistry figure: a scaffold drawn once at the top, then a table where each row is one analog with the R-group drawn in-cell and assay values to the right. Or a 2D SAR matrix — R1 on rows, R2 on columns, IC50 color-coded inside. This is the figure where chemical accuracy is most often broken by AI tools, and the figure where it matters most.

Lead optimization timeline / arrow figures

A horizontal series of compound milestones with arrows showing which property was optimized at each step — "improved hERG margin 12×", "added MetOx-blocking F", "ring contraction reduced ClogP by 0.8". This is a story figure, not a data figure.

Pharmacokinetic profile plots

Concentration–time curves on log-y axes, often comparing oral vs. IV dosing in two species. Dose–response curves with Hill fits. The conventions here are old and tight — don't get fancy.

Mapping each figure type to a LabFig workflow

The right input mode does more work than the right prompt. Here is how the figure archetypes above map to LabFig's six input modes.

Text to Figure — Use the text-to-figure tool for high-level pipeline schematics, screening cascades, and target-validation flows. Describe the steps in order ("target identification → pathway validation → modality selection → hit discovery → lead optimization") and let LabFig produce the boxes, arrows, and panel rhythm. This is also where a generic ADMET radar shell is easiest to start.
Sketch to Figure — Best for whiteboarded target-mechanism ideas ("kinase → MAPK cascade → transcription factor → phenotype"). Photograph the whiteboard at the end of the project meeting. The geometry transfers more faithfully than any prompt will.
Reference to Figure — Pull a parent-program publication (your last paper, a competitor's flagship paper, or a Nature Reviews Drug Discovery feature) and use it as a style reference. Especially useful when you are writing the second paper in a series and want visual continuity with the first.
PDF to Figure — Extract a binding-site or SAR figure from a published paper and rebuild it as editable vector for a competitive intelligence deck or an internal teaching slide. Cite the original. Do not republish a regenerated competitor figure in your paper.
Photo to Figure — Convert plate or well photos, lab-bench setups, or a screenshot of a docking session into a clean schematic. Particularly useful for the methods section of an HTS paper where the actual plate photo is too noisy to print.
Figure Enhancer — Take an older program figure (from a five-year-old paper or a stale grant) and refresh the line weights, palette, and typography for a review article or renewal grant, without changing the underlying science.

If your topic spans ML-driven generative chemistry — diffusion models, AlphaFold-derived pockets, RF-style scoring — the workflow patterns in our AI figures for ML papers post translate directly.

Visual conventions that matter

Drug-discovery figures fail review most often not for being wrong, but for being wrong-looking. Three convention layers to internalize:

Chemical structures. Bond lengths uniform, bond angles 120°, atoms only labelled when they are not C or H, stereochemistry as wedges and hashes (not as 3D perspective), and ring systems drawn in their canonical orientation. ChemDraw and RDKit both encode these conventions natively — use them for the molecules. AI tools cannot reliably hit this bar, and the difference is visible to any medicinal chemist within two seconds.

Binding-site color. The longstanding convention — hydrophobic residues yellow or grey, polar / hydrogen-bond donors blue, acceptors red, aromatic purple — is still common but increasingly seen as visually loud. Recent papers (especially in Nature Chemistry and J. Med. Chem.) trend toward more restrained palettes: a single accent color for the ligand, soft greys for the pocket surface, only key interaction residues colored. The restraint reads as confidence.

SAR tables as figures. SAR tables are figures, not tables — even when they look like tables. Treat them with the same care: consistent column widths, structures aligned on the scaffold attachment point, IC50 values right-aligned with consistent significant figures (typically two), nM units stated once in the header.

For the underlying journal expectations, the Cell figure guidelines, the Nature Chemistry author guidelines, and the Nature Reviews Drug Discovery author pages are the primary sources.

Limitations: where AI struggles with drug-discovery figures

This is the most important section of this post. Read it twice if you skim everything else.

AI figure generators in 2026 are excellent at layout, panel composition, color, and the scaffolding of a figure. They are unreliable on three things that matter intensely in medicinal chemistry:

Accurate chemical structures. Generated molecules will have wrong bond orders, missing wedges, hallucinated stereocenters, incorrect ring fusions, or substituents in the wrong position. A figure where atom 7 of a chiral center is drawn flat instead of as a wedge is not "stylistically off" — it is wrong, and a reviewer will catch it. Never publish an AI-generated molecule. Always paste the structure from ChemDraw or an RDKit export.
Exact protein–ligand geometry. AI can render a beautiful pocket, but the residues it labels and the H-bond distances it draws will not correspond to your actual crystal structure or docking pose. Generate the frame with AI; place the science from PyMOL / ChimeraX renders on top.
IUPAC-accurate labelling. Compound IDs, IUPAC names, atom numbering, and isotope labels need to be typed in deterministically. Don't let the model "fill in" a name field — it will produce something plausible-looking and wrong.

The honest framing: treat AI as the layout engine, not the chemistry engine. You verify chemical accuracy against the source data, every time. The savings are real — getting from blank canvas to publication-ready layout is dramatically faster — but the verification step is non-negotiable. A hallucinated stereocenter in a published figure is a correction-worthy event.

A workflow for your next paper or deck

A concrete five-step recipe that has held up across several real drug-discovery figures:

Structure the data first. Export molecules from ChemDraw or RDKit as SVG. Export assay tables from your ELN as CSV. Don't start in the figure tool — start in the structured source.
Generate the schematic context in LabFig. Pipeline, screening funnel, binding-pocket frame, ADMET radar shell, SAR grid layout. Use Text-to-Figure or Sketch-to-Figure for the structure; resist the urge to have the AI draw molecules.
Place real molecules into the schematic. Open the result in the LabFig vector canvas editor. Drop in the ChemDraw / RDKit SVGs onto the placeholders. Replace any AI-generated structure with the verified one.
Fix labels and values deterministically. Type compound IDs, IUPAC names, and assay values by hand. Do not trust any text the model wrote inside the figure.
Export SVG or PDF. Open in Illustrator only if needed for journal-specific final fixes — typeface, exact column width, color profile. Many drug-discovery figures now ship directly from the LabFig canvas without ever opening Illustrator.

When you move from preclinical figures into clinical-trial figures (CONSORT diagrams, Kaplan–Meier curves, forest plots), the conventions shift again — we cover that transition in AI figures for clinical AI papers. And when the upstream biology is single-cell — increasingly the foundation for novel target discovery — the figure grammar in AI figures for single-cell biology is the right next stop.

FAQ

Can AI generate accurate molecular structures for medicinal chemistry papers? No. Not yet, and not reliably. Generated structures will have bond-order, stereochemistry, or substituent errors that are unacceptable for publication. Always draw molecules in ChemDraw or generate them deterministically from SMILES with RDKit, then place those structures into the AI-generated layout. Treat the AI as the figure framer, not the chemist.

How do I depict binding sites and ligand-protein interactions clearly? Two-figure pattern: a 2D ligand interaction diagram (ligand in the middle, residues as labelled boxes, dashed H-bonds, eyelash hydrophobic contacts) plus a 3D pocket rendering (surface or cartoon) from PyMOL or ChimeraX. AI figure tools are excellent for framing both panels and adding the surrounding context (binding-pose inset, residue labels, scale bar). The actual protein–ligand geometry should come from your crystal structure or validated docking pose, not from generation.

Are AI-generated drug-discovery figures accepted at Cell or Nature Chemical Biology? As of 2026, none of the major drug-discovery-relevant journals — including Cell, Nature Chemistry, Nature Chemical Biology, Nature Reviews Drug Discovery, J. Med. Chem., and ACS Chemical Biology — prohibits AI-assisted figure production. They do expect the figure to meet technical specs (vector output, column-width sizing, typography) and most expect a one-line disclosure that AI tools were used in figure preparation. The figure is judged on scientific accuracy and visual clarity, not on its provenance.

What's the best way to visualize a virtual screening funnel? A vertical funnel with library size at top (e.g., 10^9 compounds), narrowing through filter stages — pharmacophore filter, docking, MM/GBSA, ADMET, manual triage — with the compound count and hit-rate annotated at each stage. Pair it with a chemical-space PCA or UMAP colored by scaffold cluster, and a small grid of representative hit structures. LabFig's Text-to-Figure mode produces this layout cleanly from a prompt that lists the funnel stages and counts.

How should ADMET radar plots be drawn for publication? Six to eight axes (potency, selectivity, microsomal clearance, permeability, solubility, hERG, plasma protein binding, CYP inhibition is a typical set), with values normalized to a common scale (often 0–1, with the "good" direction always outward). Modern best practice is small multiples — one small radar per compound, arranged in a row of four — rather than overlaying six radars on one chart, which becomes unreadable in print. Keep the fill semi-transparent, the axis labels small but legible, and the same axis order across all multiples.