Ari Vasvaria

AlphaFold on flexible proteins

I compared AlphaFold's predictions with lab measurements for 140 flexible proteins. It gets the rigid sections right and misplaces them relative to each other.

Python / AlphaFold DB / NMR ensembles

AlphaFold predicts a protein's 3D shape from its sequence and gives one answer. Many proteins do not have one shape. They have rigid sections joined by flexible links and move between many arrangements.

For 140 of these proteins, experiments (NMR) have recorded a set of shapes each one takes. I compared AlphaFold's prediction with those shapes in two ways: lining up the whole protein at once, and lining up each rigid section separately.

The bright chain is one prediction. The faint chains are the same rigid sections with the joints turned.
3.86 Åtypical error, whole protein aligned
1.60 Åtypical error, each section aligned separately
140proteins with measured shapes
0.50.511225510102020no differenceNFATC1: 4.52 Å whole, 4.29 Å by sectionLif: 5.9 Å whole, 2.52 Å by sectionHmgb1: 3.22 Å whole, 2.2 Å by sectionCTLA4: 4.19 Å whole, 2.14 Å by sectionSOS1: 3.73 Å whole, 3.03 Å by sectionBTK: 3.11 Å whole, 0.94 Å by sectionNos1: 8.91 Å whole, 4.71 Å by sectionIL4: 2.9 Å whole, 1.86 Å by sectionBlk: 7.43 Å whole, 6.32 Å by sectionStx1a: 2.14 Å whole, 1.92 Å by sectionPLA2G1B: 1.8 Å whole, 1.45 Å by sectionenvZ: 8.3 Å whole, 2.68 Å by sectionMSP1: 3.86 Å whole, 1.09 Å by sectionC5: 5.77 Å whole, 3.02 Å by section1COU: 9.36 Å whole, 2.99 Å by sectionPTPN13: 2.53 Å whole, 1.94 Å by sectionSGS1: 1.66 Å whole, 0.89 Å by sectionRAD53: 3.3 Å whole, 1.73 Å by sectionrub: 4.07 Å whole, 2.43 Å by sectionFOXO4: 2.06 Å whole, 1.49 Å by sectionCNOT4: 13.88 Å whole, 1.81 Å by sectionSin3b: 3.03 Å whole, 0.61 Å by sectionPIN4: 3.79 Å whole, 3.36 Å by sectionada: 8.01 Å whole, 2.51 Å by sectionBAX: 3.86 Å whole, 1.79 Å by sectionLDLR: 8.95 Å whole, 1.02 Å by sectionCrebbp: 1.34 Å whole, 0.67 Å by sectionFRQ1: 5.16 Å whole, 3.42 Å by sectionarcB: 3.02 Å whole, 1.17 Å by sectionMcl1: 1.74 Å whole, 0.9 Å by sectionPard3: 1.65 Å whole, 0.92 Å by sectionTARBP2: 1.35 Å whole, 0.88 Å by sectionKDM5D: 10.54 Å whole, 0.91 Å by sectionGTF2IRD1: 6.33 Å whole, 1.26 Å by sectionPTPRF: 1.04 Å whole, 0.89 Å by sectionPLA2G1B: 2.52 Å whole, 2.37 Å by sectionTRIP4: 2.04 Å whole, 0.87 Å by sectionABLIM2: 2.38 Å whole, 1.72 Å by sectionCbfb: 2.62 Å whole, 2.35 Å by sectionPrkn: 1.01 Å whole, 0.7 Å by sectionPLAA: 6.93 Å whole, 1.07 Å by sectionPlcg1: 11.12 Å whole, 4.74 Å by sectionpdp1: 3.44 Å whole, 0.96 Å by sectionSH3YL1: 2.27 Å whole, 0.56 Å by sectionSOST: 37.02 Å whole, 2.32 Å by sectionXRCC1: 1.96 Å whole, 1.65 Å by sectionPHF24: 4.75 Å whole, 2.76 Å by sectionSh2b2: 6.62 Å whole, 1.31 Å by sectionCYLD: 2 Å whole, 1.49 Å by sectionZfyve19: 4.89 Å whole, 0.84 Å by sectionSbf1: 1.84 Å whole, 1.42 Å by section2LJV: 2.21 Å whole, 1.13 Å by sectionglpG: 18.02 Å whole, 0.97 Å by sectionU2af2: 1.59 Å whole, 1.17 Å by sectionM50: 6.73 Å whole, 4.61 Å by sectionSRSF2: 11.2 Å whole, 1.94 Å by sectionNudc: 11.29 Å whole, 5.76 Å by sectionAPLF: 11.87 Å whole, 0.87 Å by sectionSCRIB: 1.81 Å whole, 1.69 Å by sectionPlcg2: 3.13 Å whole, 2.4 Å by sectionCLIP3: 1.76 Å whole, 1.24 Å by sectionC7: 4.2 Å whole, 1.36 Å by sectioncheA: 2.38 Å whole, 1.09 Å by sectionRbm28: 4.8 Å whole, 2.89 Å by sectionREV1: 2.27 Å whole, 0.86 Å by sectionNsd1: 1.44 Å whole, 0.76 Å by sectionUSP16: 2.36 Å whole, 0.98 Å by sectionCREBBP: 1.35 Å whole, 0.9 Å by sectionMYBPC1: 16.16 Å whole, 8.27 Å by sectionMtif3: 2.15 Å whole, 1.35 Å by sectionIFNAR2: 4.17 Å whole, 2.04 Å by sectionZHX1: 10.52 Å whole, 1.09 Å by sectionRIMBP3: 3.78 Å whole, 0.61 Å by sectionDlg4: 11.69 Å whole, 1.39 Å by sectionUFD1: 2.48 Å whole, 0.94 Å by sectionBAK1: 3.82 Å whole, 3.1 Å by sectionNAB1: 2.91 Å whole, 2.19 Å by sectionXRCC5: 2.63 Å whole, 1.2 Å by sectionItk: 2.76 Å whole, 1.93 Å by sectionGADD45A: 9.45 Å whole, 1.79 Å by sectionDLG5: 5.22 Å whole, 1.09 Å by sectionHMA4: 8.74 Å whole, 0.66 Å by sectionPKN1: 3.89 Å whole, 0.67 Å by sectionDda3937_01343: 11.11 Å whole, 6.86 Å by sectionNcam1: 1.07 Å whole, 1.06 Å by sectionSLC8A1: 18.26 Å whole, 1.54 Å by sectionPTK2B: 3.25 Å whole, 2.21 Å by sectionHECTD1: 8.15 Å whole, 0.72 Å by sectionMark3: 2.46 Å whole, 2.11 Å by sectionMyo6: 2.67 Å whole, 0.71 Å by sectionalcR: 6.04 Å whole, 3.08 Å by sectionFCER2: 6.27 Å whole, 3.78 Å by sectiontraI: 16.17 Å whole, 4.9 Å by sectionDtx2: 6.72 Å whole, 6.33 Å by sectionOnecut1: 16.65 Å whole, 4.41 Å by sectionPRNP: 4.1 Å whole, 2.5 Å by sectiontomB: 3.18 Å whole, 0.91 Å by sectionkdpD: 10.49 Å whole, 3.75 Å by sectionS100A1: 5.07 Å whole, 3.68 Å by sectionDSY0195: 3.28 Å whole, 3.05 Å by sectionARHGAP4: 5.48 Å whole, 0.65 Å by sectionVAV2: 2.73 Å whole, 0.9 Å by sectionDkk2: 4.86 Å whole, 2.09 Å by sectionsmo: 3.21 Å whole, 3.09 Å by sectionFoxd3: 6.62 Å whole, 3.17 Å by sectionSnta1: 23.93 Å whole, 15.9 Å by sectionIGHMBP2: 1.76 Å whole, 1.14 Å by sectionPSAP: 5.68 Å whole, 1.32 Å by sectionAIRE: 8.38 Å whole, 1.23 Å by sectionBC_2506: 6.85 Å whole, 0.94 Å by sectionTJP1: 9.25 Å whole, 1.2 Å by sectionmvn: 1.45 Å whole, 1.03 Å by sectionRCVRN: 11.76 Å whole, 6.17 Å by sectionNCAM1: 3.96 Å whole, 3.31 Å by sectionSMARCA2: 18.59 Å whole, 9.88 Å by sectionPARP1: 6.33 Å whole, 1.07 Å by sectionPlek: 1.21 Å whole, 0.92 Å by sectionSHQ1: 2.5 Å whole, 2.02 Å by sectionAMA-1: 17.83 Å whole, 11.85 Å by sectionSRSF1: 9.28 Å whole, 0.86 Å by sectionEUBREC_1200: 2.03 Å whole, 1.02 Å by sectionMGG_03307: 10.56 Å whole, 1.1 Å by section2N8K: 1.42 Å whole, 0.86 Å by section1PSY: 10.79 Å whole, 2.52 Å by sectionTSEN15: 2.86 Å whole, 2.04 Å by sectionNHP2: 3.14 Å whole, 1.84 Å by sectionOSBPL8: 2 Å whole, 1.47 Å by sectionSos1: 7.94 Å whole, 4.26 Å by sectionMED25: 3.59 Å whole, 1.68 Å by sectionBCR: 8.45 Å whole, 2.45 Å by sectionSMURF2: 1.49 Å whole, 1.21 Å by sectionSYT13: 2 Å whole, 0.91 Å by sectionCESA7: 7.19 Å whole, 1.1 Å by sectionPARD3: 3.67 Å whole, 1.47 Å by sectionHCH_03072: 2.13 Å whole, 0.75 Å by sectionMPP1: 1.88 Å whole, 1.74 Å by sectionCYB_2465: 5.59 Å whole, 1.6 Å by sectionTXNDC5: 1.2 Å whole, 0.91 Å by sectionHDGF: 7.39 Å whole, 3.43 Å by sectionCnot4: 3.73 Å whole, 2.58 Å by sectionError with the whole protein aligned at once (Å)Error with each section aligned separately (Å)
One dot per protein. Dots below the dashed line are proteins where the sections are more accurate than the whole. Hover a dot to see which protein it is.

How to read the chart

Error is measured in ångströms (Å), a tenth of a nanometre. An atom is about 1 to 2 Å across, so an error of 1.6 Å means the prediction is off by roughly one atom, and 3.9 Å means it is off by two or three. The horizontal position of a dot is the error when the whole protein is lined up at once. The vertical position is the error when each rigid section is lined up separately. Blue dots improve by more than 1.5 times.

How much each protein improves

97 of the 140 proteins improve by more than 1.5 times, 65 by more than 2 times and 46 by more than 3 times.

under 1.5×
43
1.5× to 2×
32
2× to 3×
19
3× to 10×
38
over 10×
8

The biggest differences

For these proteins the whole-protein error is large and the section error is under about 2 Å. AlphaFold has each piece almost exactly right and has the pieces in the wrong arrangement.

ProteinWhat it isWholeBy sectionTimes smaller
glpGRhomboid protease GlpG18.02 Å0.97 Å18.5×
SOSTSclerostin37.02 Å2.32 Å16.0×
APLFAprataxin and PNK-like factor11.87 Å0.87 Å13.6×
HMA4Putative cadmium/zinc-transporting ATPase HMA48.74 Å0.66 Å13.2×
SLC8A1Sodium/calcium exchanger 118.26 Å1.54 Å11.9×
KDM5DLysine-specific demethylase 5D10.54 Å0.91 Å11.6×
HECTD1E3 ubiquitin-protein ligase HECTD18.15 Å0.72 Å11.4×
SRSF1Serine/arginine-rich splicing factor 19.28 Å0.86 Å10.8×

What it means

The error roughly halves when the sections are lined up separately. So the sections are mostly correct, and most of the error is in how they are placed relative to each other.

AlphaFold also reports how unsure it is about the distance between any two parts of the protein. I checked that against how much those distances vary in the experiments, and the two agree well (correlation 0.72). So its uncertainty already points at the flexible joints.

What I am doing next is bending the predicted shape only at those joints, by an amount set by the uncertainty, and checking whether the bent shapes cover what the experiments saw.

Segments are runs of at least 5 residues with pLDDT of 70 or more. RMSD_global uses one rigid superposition of the AlphaFold model onto each NMR model. RMSD_local superposes each segment independently. Medians are 3.86 Å and 1.60 Å, the median ratio is 1.96, and 46% of proteins have a ratio above 2.

The 0.72 is the median Spearman correlation between PAE and the experimental spread of inter-residue distances. The follow-up rotates everything downstream of each joint about its last Cα, with angular magnitude proportional to PAE across that joint, against a uniform-budget control and the static model.