Skip to content

Troubleshooting & Common Pitfalls

A quick-lookup page for "something seems off" moments — organized by symptom, not by tool. Each row links to the page with the full explanation; this page is for finding the right row fast, not for reading start to finish.

Reading a prediction

Symptom Likely cause Where to read more
A region looks confidently folded (high pLDDT) but you know it's disordered Conditionally-folded IDR — AlphaFold can predict the bound or modified conformation with high confidence even from a naked sequence pLDDT is not a clean disorder detector
A clean-looking helix has low pLDDT Probably a real signal, not noise — likely a transient/conditional helix, or (in AlphaFold3 specifically) a hallucinated one Confidence Metrics
Two domains each look great individually, but you're not sure about their arrangement pLDDT can't tell you — that's what PAE is for Why do I need PAE if I already have pLDDT?
Two chains are touching in the 3D view — are they really interacting? Not necessarily. AlphaFold/ColabFold always places chains near each other in a multimer job, even with no real interaction Chance or determination?
ipTM varies a lot across the 5 models for the same complex Not a bug — can be a real signal of interface ambiguity, or one outlier among four agreeing models Real-world example
You mutated a residue in silico and pLDDT barely changed (or moved the "wrong" way) Expected — AF2 is largely insensitive to point mutations; pLDDT often doesn't track real stability effects What Actually Influences a Prediction?
A PTM you specified doesn't seem to change anything You're using AlphaFold2/ColabFold — they have no PTM representation at all. Only AlphaFold3 supports PTMs as explicit input What Actually Influences a Prediction?

AlphaMissense

Symptom Likely cause Where to read more
AlphaMissense fetch returns nothing for your protein It's not human, or not the canonical sequence — scores only exist natively for human canonical UniProt sequences Using AlphaMissense on a non-human protein
Residue numbering seems off vs. what you expected from AlphaFold DB You may be looking at an isoform-specific entry, or a canonical one when you needed an isoform (or vice versa) Check which sequence you're actually looking at
A score seems too extreme / doesn't match known clinical classification AlphaMissense over-predicts pathogenicity in some benchmarks and is supporting evidence only, never standalone Known limitations

ChopChopMF

Symptom Likely cause Fix
Crop/Delete removed too much and can't be undone No duplicate made first Always Duplicate Structure before a destructive edit — see Trimming a model
PAE Contacts tool won't run / gives odd results More than one model open Close/hide other models — only one is supported
Foldseek hits look noisy or miss the obvious homolog Flexible/disordered regions still in the model, or cropped down to a single helix/sheet Crop a complete domain first — see Finding structural homologs
Foldseek finds nothing for a poorly characterized organism Searched PDB only Switch the target database to AlphaFold DB (afdb50)
ΔG plot dominated by tiny values Neutral band toggle off Leave Neutral band (±ε) checked

Running ColabFold / HPC

Symptom Likely cause Where to read more
A run is spending most of its time before inference even starts Expected — MSA search (not the network) is the classic AF2 bottleneck, historically ~80%+ of wall-clock time AF2 vs. ColabFold: same brain, faster search
AlphaFold3 container won't run on an older cluster CUDA version mismatch — AF3 needs CUDA 12.3+, older cluster drivers may only support 12.2 Running ColabFold (local/HPC) — containerize with Apptainer
AlphaFold Server rejects a job or a ligand/ion Server enforces a 5,000-token cap and only supports a specific ion list (MG, ZN, CL, CA, NA, MN, K, FE, CU, CO) AlphaFold3 & AlphaFold Server practical limits

Still stuck?

Check Resources & Further Reading for primary sources, or — for the tools themselves rather than this guide's explanations — report it directly where they're maintained: ChopChopMF issues · ChimeraX-FigureStyle issues.