Fig. 4: Structure prediction with in-cell photo-L crosslinking MS data of the E. coli membrane fraction.
From: Protein structure prediction with in-cell photo-crosslinking mass spectrometry and deep learning
a, Comparison of TM score with annotated number of links (marker sizes) and percentage of nonsatisfied (>10 Å) crosslinks (color gradient) in the AlphaFold2 prediction. Performance improvement is bigger for targets with a higher percentage of nonsatisfied crosslinks in the base prediction (darker circles). Each target is predicted ten times with different MSA subsamples at Neff = 10. AlphaLink outperforms AlphaFold2 on average. b, Comparison of TM score with annotated mean distance of nonsatisfied crosslinks in the base AlphaFold2 prediction (color gradient). Prediction quality improves with stronger crosslink violations (darker circles). c, We show the calibration of the pTM. On predictions that are at least 80% covered by the crystal structure, the correlation is 0.75. The true TM score is generally underestimated, meaning that the pTM score of AlphaLink is a conservative estimate. The shaded area corresponds to the 95% confidence interval. Line shows the linear fit. d, Prediction of the ATP synthase subunit AtpB by AlphaFold2 and AlphaLink using in-cell photo-L crosslinks at Neff = 10. e, Prediction of the outer membrane lipopolysaccharide assembly protein. f, Prediction of the ferrienterobactin receptor. In all three cases, the in-cell crosslinking data helps AlphaLink position different regions of the protein relative to each other, yielding a performance improvement over AlphaFold2. The crystal structure of the target protein is shown in gray, overlaid with the AlphaLink prediction.
