Release week 2020-10-08
⭐ This week's notable releases
No novel or confidently-wrong structures this week — AlphaFold kept up.
Zoomed into the red "confidently wrong" box above (pLDDT ≥ 70, TM < 0.5) — the structures AlphaFold got confidently wrong.
How to read this
Every point is one experimental protein structure. The horizontal axis is AlphaFold's own confidence in its prediction (mean pLDDT, 0–100). The vertical axis is how well that blind prediction actually matches the experiment (TM-score, 0–1; above 0.5 means the same fold, above 0.9 near-identical). Marker size grows with the FRAUD score (confidence-weighted error).
Cutoff: the shaded red box is the "confidently wrong" zone — AlphaFold was confident (mean pLDDT above 70) yet the fold is wrong (TM-score below 0.5). A predictor that had truly solved folding would leave that box empty.
Take-home: 0 of 2 structures (0.0%) are confidently wrong; median TM-score is 0.762.
Read moreShow less — how each metric is calculated
TM-score: how similar the two 3D shapes are overall, 0–1 (a random pair scores ~0.17, an identical fold ~1). Length-normalised so large and small proteins compare fairly. TM = (1/L) Σᵢ 1/(1 + (dᵢ/d₀)²) — dᵢ is the gap between the i-th aligned Cα atoms after best-fit superposition, and d₀ = 1.24(L−15)^⅓ − 1.8 sets the distance scale for length L. Ref: Zhang & Skolnick, Proteins 2004 doi:10.1002/prot.20264; computed with TM-align, doi:10.1093/nar/gki524.
pLDDT: AlphaFold's own confidence in each residue, 0–100 (higher = surer). It is the model predicting its own accuracy before ever seeing the experiment. We plot the per-structure mean. pLDDT = (1/N) Σᵢ pLDDTᵢ, where pLDDTᵢ is the network's confidence output for residue i.
FRAUD score: the headline number — how wrong the prediction was, weighted by how confident AlphaFold was, so a big error it was sure about counts most. FRAUD = (1/N) Σᵢ (pLDDTᵢ/100) · min(Δᵢ,15)/15, where Δᵢ is residue i's Cα distance from the experiment (Å) after superposition, capped at 15 Å. Runs 0 (perfect) to 1.
Novelty: how little AlphaFold had to go on — 100 minus the highest sequence identity between this protein and any structure released before the 2018-04-30 training cutoff. High = genuinely unseen (100% = nothing similar was ever in the training set). novelty = 100 − maxₚ identity(s, p) over pre-cutoff PDB chains p, with identity = 100 × (matching aligned residues)/(alignment length), from an MMseqs2 search.
Homology: the point colour. High novelty (above 70%) flags the sequence as novel (amber) — AlphaFold had no close template; otherwise a pre-cutoff homolog existed (blue) it could have learned the fold from. novel ⇔ novelty > 70%.
What the metrics mean
TM-score (0–1): overall fold match — above 0.5 is the same fold, above 0.9 near-identical. Cα-RMSD (Å): average backbone distance after best-fit superposition — lower is better (under 2 Å is excellent). lDDT (0–1): local accuracy measured without superposition — above 0.8 is good. Each bar counts how many structures fall in that range.
Take-home: median TM-score 0.762 — most predictions match the experimental fold well, with a long tail that do not.
Read moreShow less — how each metric is calculated
TM-score: overall shape match, 0–1 (above 0.5 = same fold, above 0.9 near-identical), length-normalised. TM = (1/L) Σᵢ 1/(1 + (dᵢ/d₀)²) with dᵢ the aligned-Cα gap after superposition and d₀ = 1.24(L−15)^⅓ − 1.8. Ref: Zhang & Skolnick, Proteins 2004 doi:10.1002/prot.20264.
Cα-RMSD: the average straight-line distance between matching backbone Cα atoms once the two structures are best-fit superposed (Å; lower is better, under ~2 Å excellent). RMSD = √( (1/N) Σᵢ |Pᵢ − (R·Qᵢ + t)|² ), with Pᵢ/Qᵢ the experimental/model Cα coordinates and R,t the rotation and translation from Kabsch superposition.
lDDT: local accuracy with no superposition — the fraction of short-range inter-residue distances the model reproduces, so it is not fooled by a single wrongly-placed domain. lDDT = (1/N) Σᵢ ¼ Σ_t 1[ |d_exp − d_model| < t ] over thresholds t ∈ {0.5, 1, 2, 4} Å and all residue pairs within 15 Å.
Trend over time
The blue line is the mean TM-score of the structures by their deposit month; the red bars count how many were confidently wrong. Binning by deposit date (rather than by when we processed them) gives an even timeline, tracking whether AlphaFold's accuracy on newly-deposited structures is holding steady, improving, or slipping as the PDB keeps growing. Months with fewer than 5 structures are omitted so each point is a meaningful average. A dashed trend line is drawn only if the change over time is statistically significant.
Read moreShow less — how each metric is calculated
Mean TM-score: the month's average shape-match score. TM̄ = (1/M) Σ TM over the M structures deposited that month.
Trend line: a monotonic-trend test over the monthly means — Mann-Kendall (Kendall's τ) for significance, with a robust Theil-Sen slope for the line. Rank-based, so it tolerates the skewed TM distribution and noisy low-count months. Shown only when p < 0.05; otherwise a note reports that no significant trend was found.
Confidently wrong: the count that were both confident and wrong. count( mean pLDDT > 70 AND TM < 0.5 ).
Matched structures
| PDB | UniProt | Protein | Method | Å | Deposited | Novelty % | pLDDT | TM | lDDT | GDT_TS | RMSD | FRAUD | Flag |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 9A06_A | P51787 | KCNQ1 channel-forming domain | Integrative | — | 2020-02-10 | — | 67.75 | 0.76 | — | — | — | 0.16 | ok |
| 9A06_E | Q6FHJ6 | KCNE1 transmembrane domain | Integrative | — | 2020-02-10 | 0.00 | 84.60 | 0.59 | 0.90 | 61.43 | 2.41 | 0.12 | ok |
Click a column header to sort. FRAUD = confidence-weighted error; a structure is flagged confidently wrong when mean pLDDT > 70 yet TM-score < 0.5.