NumberReader
The younger version of this idea reads AlphaFold's confidence as four honest colours. NumberReader reads the real numbers. It loads a live AlphaFold model of human haemoglobin (the alpha subunit), then two proteins whose confidence spreads across all four bands, p53 and c-Myc, and, for each region, asks you to commit a confidence guess before the number appears — then reveals the region's actual per-residue pLDDT, computed from the model itself, and the pairwise PAE. Everything runs locally; nothing about you is collected.
Also: BarBet, bet how a whole chain splits across the four bands, then see the model's bar.
What kind of number is this? pLDDT and PAE are the model's own estimates of how likely it is to be right: numbers from a model, not measurements. A model can sound sure and still be wrong, which is why an experiment is the check. For ages 9–14, Spark & Anvil teaches the same idea with Percent Bot, where sounding sure is not the same as being right.
This is the tool. The numbers are not typed in by us — the game fetches the live model from the AlphaFold Database and computes each region's mean confidence from the model's own per-residue values, so you read the real AlphaFold output. Every region and the confidence bands are listed in full below (works without JavaScript).
Loading the live model… If it does not start, the full region guide below still works.
Before you call it: guessing alone would be right about 1 time in 4.
Most first predictions in science miss. Checking is the point.
How to play. Press Start. A region is highlighted in 3D, in violet on a grey model. Commit how confident you think the model is there — 1 Very high, 2 Confident, 3 Low, 4 Very low — before the number shows. The real mean pLDDT for that region is then read off the model and scores your guess, and the region shows in its own confidence colours with the rest of the chain dimmed. After haemoglobin a round reads the real PAE; then come p53 and c-Myc, and one last question asks the thing that matters most: prediction, or experiment? Press Enter or Next to continue.
Why the numbers you read matter
Haemoglobin α's core scores very high because the globin fold is compact and well determined. Its haem pocket is lined by two histidines, His58 and His87 in the usual mature-chain numbering (His59 and His88 in UniProt P69905, which counts the starting methionine). The model shows only the protein chain: the haem itself and the partner chains are not in it.
What the numbers mean: pLDDT and PAE
AlphaFold scores every residue from 0 to 100 for how sure it is of the local shape — a per-residue score called pLDDT. It is stored right in the model file (the B-factor column), which is where this game reads it. EMBL-EBI groups the score into four bands:
- Very high (pLDDT ≥ 90) — backbone and side chains usually both accurate
- Confident (70–90) — backbone usually right; some side chains may be off
- Low (50–70) — treat with caution — often flexible
- Very low (< 50) — should not be interpreted on its own
PAE (predicted aligned error, in ångströms) is different: it says how sure the model is about where two residues sit relative to each other.
The regions this game reads
Answer key — open it after you play
| Region | Residues | What it is |
|---|---|---|
| Hemoglobin subunit alpha P69905 · Homo sapiens · 142 residues | ||
| The very start (N-terminus) | 1–4 | The first few residues of the chain. Chain ends are the classic place a prediction is a little less sure, because they are floppier and have fewer neighbours to pin them down. |
| The heme-pocket helices (E–F region) | 58–77 | The core helices that cradle the heme group and the iron that carries oxygen — including the distal histidine His59 (His58 in the usual mature-chain numbering); the proximal His88 (His87) sits just beyond this stretch. A tightly-packed, well-determined part of the fold. |
| The folded globin core | 30–120 | The main body of the classic globin fold — eight helices packed together. This is exactly the kind of compact, well-constrained region AlphaFold predicts most confidently. |
| The very end (C-terminus) | 138–142 | The last few residues. Chain ends are often less sure, but not always: the last helix of haemoglobin α runs almost to the end of the chain, and here the model stays very sure right to the end. |
| Cellular tumor antigen p53 P04637 · Homo sapiens · 393 residues · regions picked by script | ||
| Residues 29–65 (near the start) | 29–65 | A script picked this stretch from the model’s own numbers: it is the longest run in this chain where all 37 residues sit in one band, very low. |
| Residues 80–86 (near the start) | 80–86 | A script picked this stretch from the model’s own numbers: it is the longest run in this chain where all 7 residues sit in one band, low. |
| Residues 189–290 (in the middle) | 189–290 | A script picked this stretch from the model’s own numbers: it is the longest run in this chain where all 102 residues sit in one band, very high. |
| Residues 349–355 (near the end) | 349–355 | A script picked this stretch from the model’s own numbers: it is the longest run in this chain where all 7 residues sit in one band, confident. |
| Myc proto-oncogene protein P01106 · Homo sapiens · 454 residues · regions picked by script | ||
| Residues 65–80 (near the start) | 65–80 | A script picked this stretch from the model’s own numbers: it is the longest run in this chain where all 16 residues sit in one band, low. |
| Residues 81–107 (near the start) | 81–107 | A script picked this stretch from the model’s own numbers: it is the longest run in this chain where all 27 residues sit in one band, very low. |
| Residues 203–209 (in the middle) | 203–209 | A script picked this stretch from the model’s own numbers: it is the longest run in this chain where all 7 residues sit in one band, confident. |
| Residues 367–452 (near the end) | 367–452 | A script picked this stretch from the model’s own numbers: it is the longest run in this chain where all 86 residues sit in one band, very high. |
PAE: for a single tightly-folded domain like haemoglobin's alpha subunit, the number is small across the whole chain, so the model is confident about the whole shape at once.
Methods & about
The 3D models are fetched live from the AlphaFold Database (Hemoglobin subunit alpha, UniProt P69905; Cellular tumor antigen p53, UniProt P04637; Myc proto-oncogene protein, UniProt P01106) and coloured by pLDDT. The p53 and c-Myc regions were picked by a script from a vendored copy of each model: for each band, the longest stretch where every residue sits in that band, so the regions never overlap. Each region's confidence figure is computed in your browser from the model's own per-residue pLDDT (the mmCIF B-factor column); the PAE figure is read from AlphaFold's published PAE data. No number here is authored by us, and nothing is stored — your browser talks straight to the AlphaFold Database and computes the rest. Everything AlphaFold shows is a prediction, not an experiment: reading the real confidence numbers is exactly how you judge how far to trust each part of a predicted structure.
Want the qualitative, colour-based version for younger learners? Play VanishPoint. Want to drive any structure yourself? Open the free AlphaFold viewer. Everything we make is free and non-profit; learn more on the home page.