Skip to content

Presenting & Sharing AlphaFold Results

Once you trust a prediction (see Confidence Metrics), the next failure mode is presenting it in a way that misleads readers or can't be reproduced. This page summarizes practical reporting norms drawn from the ASBC AlphaFold best-practices guide.

Beginner: the minimum every figure/methods section needs

  • Label it as a model. State explicitly in the figure or legend that a structure is an AlphaFold/ColabFold prediction, not an experimentally determined structure.
  • Show confidence alongside the structure, not just the cartoon: a pLDDT-colored model and a PAE plot, each with its color key.
  • Report pTM / ipTM (the latter for complexes) in the figure or a supplementary table.
  • Name the exact tool and version: AlphaFold2, AlphaFold3, ColabFold, LocalColabFold, or OpenFold are not interchangeable — say which one, and which version.

Advanced: what to include in Methods

  • Model type, number of recycles, early-stopping tolerance, whether Amber relaxation was applied, and random seed(s) used.
  • Which sequence databases were queried (not applicable for AlphaFold3, which uses its own internal pipeline).
  • Enough detail that someone else could reproduce the run — link a GitHub repo or container definition if you modified anything from defaults.
  • Cite the primary literature for whichever tool you used: Jumper et al. 2021 for AlphaFold2, Evans et al. for AlphaFold-Multimer, Mirdita et al. 2022 for ColabFold, Ahdritz et al. for OpenFold.
Expert: full data-sharing checklist
  • Input sequences as FASTA or a supplementary table, with subunits clearly identified for multimeric predictions.
  • All five ranked .pdb/.cif models plus their matching .json PAE files, deposited via ModelArchive, GitHub, or Figshare — not just the top-ranked model.
  • For AlphaFold3: the .json input file(s), since these are what preserve post-translational modifications and ligand specifications — without them the run isn't reproducible.
  • Any custom/non-standard scripts used to prepare inputs or postprocess outputs.
  • .a3m MSA files where feasible — large size can be a reason to omit them, but only if the methods section describes exactly how to regenerate them.
  • Show all five models superposed at least once (even if only in supplementary material) to demonstrate — or expose — consistency across AlphaFold's independent predictions, rather than cherry-picking the best-looking one.

Why this matters: confident-looking ≠ correct

Expert: the evidence for skepticism

This isn't a theoretical concern. Terwilliger et al. (2023/2024, Nature Methods) compared AlphaFold predictions against experimental maps and found that even confidently-predicted regions can diverge from the real structure — at the global scale (domain orientation, overall distortion) and at the local scale (backbone and side-chain conformation). Their conclusion, worth internalizing: AlphaFold predictions are valuable hypotheses that accelerate but do not replace experimental structure determination. Practical implication for anyone publishing: treat a prediction as strong evidence to design the next experiment around, not as a substitute for that experiment — particularly for any interaction or ligand/cofactor binding site that wasn't part of the original prediction input.

For a broader treatment of AF2's practical strengths, failure modes, and how to responsibly improve a raw prediction before using it downstream (e.g. adding missing cofactors/ligands via AlphaFill, or disulfide bridges via MODELLER), see Sun et al., "AlphaFold2 for Protein Structure Prediction: Best Practices and Critical Analyses" (2024).

Common mistakes to avoid

  • Presenting a model without ever calling it a "prediction" or "computational model."
  • Showing a cartoon with no confidence coloring, PAE plot, or pTM/ipTM anywhere in the paper.
  • Reporting only the single top model, with no indication of agreement/disagreement across the other four.
  • Vague methods ("we ran AlphaFold") with no version, parameters, or database information — this alone can make a result irreproducible.
  • Forgetting citations for the specific pipeline used (ColabFold citations are commonly dropped even when ColabFold, not the original AlphaFold2 pipeline, was actually used).

MSA depth as a sanity check

Before even looking at the 3D structure, check the sequence coverage / MSA depth plot that ColabFold outputs alongside the model. As a rule of thumb, roughly 200+ sequences in the alignment at a given position is associated with confident predictions there; regions with much shallower coverage (well below ~100 sequences) are a leading indicator of an unreliable region, often before pLDDT even reflects it clearly. Recommended reading order for a fresh result: coverage plot → pLDDT plot → PAE plot → 3D structure — in that order, so you form an expectation of reliability before the cartoon can bias you.

See also: Confidence Metrics for what pLDDT/PAE/ipTM actually measure, and ChopChopMF Workflows for generating these views interactively in ChimeraX.

Correct isn't the whole job — it should also be legible

Everything above this point is about getting the science right. But a figure also has to actually get looked at closely, and a cluttered or inconsistently-styled one works against that — presentation quality isn't cosmetic, it's part of communicating the result.

Advanced: found a look you like? Save it as a template

Once you've dialed in lighting, cartoon style, coloring (including AlphaFold pLDDT), and export settings for one structure in ChimeraX, there's no need to redo all of that by hand for every figure in a paper. FigureStyle saves that exact setup as a named, reusable template — apply it to any other structure with one click, and every figure in the paper gets identical, reproducible styling instead of drifting slightly each time you re-tune the display.

A FigureStyle template applied in ChimeraX: pLDDT-colored, cel-shaded cartoon with outlines on a white background, plus the confidence color key baked into the render.

See the full plugin card for FigureStyle's coloring modes and export options, or its docs and GitHub.

Continue to Visualization Tools →