What Actually Influences a Prediction?¶

AlphaFold2 "predicts structures based on those available in the PDB [and the evolutionary patterns across the MSA], rather than by fundamental driving forces of protein folding" — it has no physical force field and was never trained on labeled examples of which specific point mutations destabilize a fold, only on solved structures and their natural sequence variation across species. A single substitution is a vanishingly small perturbation to the pattern the MSA encodes across dozens or hundreds of homologs, so the network has little reason to move the backbone in response to it — even when that substitution would be catastrophic in the lab.

Post-translational modifications (PTMs) are a related but distinct story:

• AlphaFold2 has no representation for PTMs at all. The input is a plain amino-acid sequence; phosphorylation, glycosylation, acetylation etc. simply don't exist in that input space. Some users approximate a phosphorylation with a phospho-mimetic substitution (e.g. Ser→Asp/Glu), but this is a rough proxy, not a modeled PTM.
• AlphaFold3 supports certain PTMs as explicit, first-class input: phosphorylation of serine/threonine/tyrosine/histidine, and glycan chains, each represented as their own atoms/tokens with a dedicated per-atom confidence frame (built from the atom's nearest neighbors in a reference conformer, rather than a protein backbone frame). This means AF3 can, unlike AF2, actually shift a prediction in response to a specified PTM — see AF2 vs. ColabFold vs. AF3.

The deeper reason this all connects to dynamics: proteins are not the single rigid shape AlphaFold outputs. EBI's own limitations page is direct about this: "Proteins undergo structural changes when they perform their functions... AlphaFold2 does not capture such conformational changes, as it was designed to predict static structures." A real protein is better thought of as an ensemble of conformations with different populations — and a mutation can genuinely shift that population (e.g. making a transient helix more or less likely to be occupied at any given moment) without necessarily "breaking" the dominant fold AlphaFold reports. This is exactly the phenomenon discussed in Confidence Metrics → pLDDT is not a clean disorder detector: a helix with borderline pLDDT is plausibly read as transient, not wrong — but as the table above shows, don't expect pLDDT to move in lockstep with every mutation that could shift that population. Researchers who specifically want to probe alternative conformations from AlphaFold have developed workarounds — subsampling or clustering the input MSA (e.g. AF-Cluster-style approaches) to coax the network toward different plausible states of the same protein — but this is a deliberate trick, not something a default single run gives you.