MD-SOP: Sequence-based prediction of protein backbone order parameters across computational label sources
Year: 2026
Type: preprint
Abstract: Experimental data for protein backbone N–H order parameters () are limited, which restricts model training and testing. MD-SOP was developed to predict these parameters from protein sequences using labels calculated from molecular dynamics (MD) simulations in ATLAS and pseudo-labels generated by S-OPPE. ESM-C embeddings and hidden Markov model profiles were used as inputs to a bidirectional long short-term memory network. MD-only, S-OPPE-only, and mixed training were compared using two regression heads. The two datasets were split jointly by sequence similarity. Each model configuration was trained with five random seeds on the same split. Models trained and selected using one label source could predict residue-level trends in the other source, although systematic biases remained. Mixed training gave root-mean-square errors (RMSEs) between those of the two single-source models on both computational test sets. On ten experimental proteins excluded from training and model selection, the MD-only model with a single-position regression head had the lowest mean RMSE (0.109). The corresponding S-OPPE-only model had the highest mean Pearson correlation coefficient (0.860). These findings support the use of computational labels for training sequence-based order-parameter predictors. Tests across label sources provide indirect evidence of generalization when experimental data are limited. However, they do not establish that either label source is equivalent to experimental measurements. Further tests on new data splits and more independent experimental proteins are needed.
Source: ChemRxiv
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: green
Awards: G2710976002, G6091391392, G7964239359
OpenAlex ID: W7236108891
