Protein complex prediction using Rosetta, AlphaFold, and mass spectrometry covalent labeling

Abstract

Covalent labeling (CL) in combination with mass spectrometry can be used as an analytical tool to study and determine structural properties of protein-protein complexes. However, data from these experiments is sparse and does not unambiguously elucidate protein structure. Thus, computational algorithms are needed to deduce structure from the CL data. In this work, we present a hybrid method that combines models of protein complex subunits generated with AlphaFold with differential CL data via a CL-guided protein-protein docking in Rosetta. In a benchmark set, the RMSD (root-mean-square deviation) of the best-scoring models was below 3.6 Å for 5/5 complexes with inclusion of CL data, whereas the same quality was only achieved for 1/5 complexes without CL data. This study suggests that our integrated approach can successfully use data obtained from CL experiments to distinguish between nativelike and non-nativelike models.

Using AlphaFold Predictions in Viral Research

“…To overcome the limitation described above, combining AF2 with experimental methods, e.g., cryo-electron tomography and/or other computer-based tools such as RoseTTAFold, provides more robust results [ 64 , 66 ]. Other authors have suggested combining AlphaFold models of protein complexes with differential covalent labelling mass spectrometry data by applying RosettaDock [ 67 ]. The use of cryo-electron microscopy maps, integrated with AlphaFold, for multi-chain protein complex prediction also encourages the creation of accurate and reliable models [ 68 ].…”

The bad and the good of trends in model building and refinement for sparse-data regions: pernicious forms of overfitting versus good new tools and predictions

“…We hypothesize that barbed-wire regions represent predictions that failed at an early stage, presumably because of unhelpful MSA patterns, and also that barbed wire is the reason that pLDDT < 50 was found to be an excellent predictor of intrinsically disordered regions (IDPs) in CASP14 (Ruff & Pappu, 2021). As is also true for many IDPs, some barbed-wire regions are known to fold when they find the right binding partner, which can often be shown by AlphaFold or RoseTTAFold multimer prediction (Drake et al, 2022). For anyone used to looking at protein 3D structures, barbed-wire segments are obvious in visualizations, especially with outliers turned on, but to enable automation we have developed a set of five specially tuned criteria (packing, , !, CaBLAM and geometry) that can identify and delete the barbed wire from low-pLDDT regions, leaving the near-folded parts that may have usable predictive value, as in Fig.…”

Recent Advances and Challenges in Protein Structure Prediction

“…For example, Steffen Lindert et al proposed a hybrid method that combines models of protein complex subunits generated with AlphaFold with differential covalent labeling (CL) data via a CL-guided protein−protein docking in Rosetta. 150 When SAXS experimental data is available, HDOCK can incorporate the SAXS data into docking and postdocking processes. 145 In the aftermath of the AlphaFold2 deep learning revolution, the modeling of protein complexes has become a prominent research focus nowadays in the field of structural biology.…”