FoldCompare · a mode
StickSquare
"The two models differ by 1.2 Å RMSD." What does that number mean? Lay one model on the other as well as it will go, then measure how far each residue's centre misses its partner. That miss is a stick. Square every stick, pour the squares into one average, and take the square root. That is RMSD: root, mean, square, deviation.
Each pair below is a real one from ModelDiff: the AlphaFold Database model and ESMFold's prediction of the same protein. The fit uses the residues both models are sure about (or the whole chain when there are too few), so only those residues are in the number.
Haemoglobin subunit alpha (UniProt P69905) · 141 residues in the fit · RMSD 0.31 Å
1. One stick per residue
The tallest stick is 2.35 Å (residue 142). The average stick is 0.22 Å.
2. Square every stick
Squaring makes every miss positive and makes the big ones count far more: the tallest stick's square is 5.52 Ų.
3. Pour the squares into one average, then take the root
average of the 141 squares = 0.095 Ų → √0.095 = 0.31 Å
That is exactly the RMSD ModelDiff prints for this pair (0.31 Å). It is bigger than the average stick (0.22 Å) because squaring lets the big misses pull it up.
Lysozyme C (UniProt P61626) · 131 residues in the fit · RMSD 0.31 Å
1. One stick per residue
The tallest stick is 2.69 Å (residue 18). The average stick is 0.20 Å.
2. Square every stick
Squaring makes every miss positive and makes the big ones count far more: the tallest stick's square is 7.24 Ų.
3. Pour the squares into one average, then take the root
average of the 131 squares = 0.098 Ų → √0.098 = 0.31 Å
That is exactly the RMSD ModelDiff prints for this pair (0.31 Å). It is bigger than the average stick (0.20 Å) because squaring lets the big misses pull it up.
Alpha-synuclein (UniProt P37840) · 140 residues in the fit · RMSD 35.30 Å
1. One stick per residue
The tallest stick is 72.04 Å (residue 1). The average stick is 31.39 Å.
2. Square every stick
Squaring makes every miss positive and makes the big ones count far more: the tallest stick's square is 5189.76 Ų.
3. Pour the squares into one average, then take the root
average of the 140 squares = 1245.869 Ų → √1245.869 = 35.30 Å
That is exactly the RMSD ModelDiff prints for this pair (35.30 Å). It is bigger than the average stick (31.39 Å) because squaring lets the big misses pull it up.
GTPase KRas (UniProt P01116) · 175 residues in the fit · RMSD 0.81 Å
1. One stick per residue
The tallest stick is 4.56 Å (residue 61). The average stick is 0.45 Å.
2. Square every stick
Squaring makes every miss positive and makes the big ones count far more: the tallest stick's square is 20.79 Ų.
3. Pour the squares into one average, then take the root
average of the 175 squares = 0.657 Ų → √0.657 = 0.81 Å
That is exactly the RMSD ModelDiff prints for this pair (0.81 Å). It is bigger than the average stick (0.45 Å) because squaring lets the big misses pull it up.
Green fluorescent protein (UniProt P42212) · 33 residues in the fit · RMSD 0.42 Å
1. One stick per residue
The tallest stick is 1.18 Å (residue 50). The average stick is 0.37 Å.
2. Square every stick
Squaring makes every miss positive and makes the big ones count far more: the tallest stick's square is 1.39 Ų.
3. Pour the squares into one average, then take the root
average of the 33 squares = 0.180 Ų → √0.180 = 0.42 Å
That is exactly the RMSD ModelDiff prints for this pair (0.42 Å). It is bigger than the average stick (0.37 Å) because squaring lets the big misses pull it up.
Cellular tumour antigen p53 (UniProt P04637) · 225 residues in the fit · RMSD 9.77 Å
1. One stick per residue
The tallest stick is 35.81 Å (residue 355). The average stick is 7.11 Å.
2. Square every stick
Squaring makes every miss positive and makes the big ones count far more: the tallest stick's square is 1282.36 Ų.
3. Pour the squares into one average, then take the root
average of the 225 squares = 95.513 Ų → √95.513 = 9.77 Å
That is exactly the RMSD ModelDiff prints for this pair (9.77 Å). It is bigger than the average stick (7.11 Å) because squaring lets the big misses pull it up.
Reading an RMSD honestly
- Small (around 1 Å or less) over many residues means the shapes match closely where they were compared.
- One number hides where the misses are. A floppy loop or a moved domain can make it large while most of the chain matches; that is why ModelDiff also fits the pieces separately.
- Always ask which residues were fitted. An RMSD over the confident core says nothing about the rest.