Search for: All records

Medium-resolution cryo-electron microscopy (cryo-EM) density maps preserve substantial information about protein secondary-structure organization; however, accurately recovering the topology and connectivity of α-helices and β-strands remains challenging due to noise, structural heterogeneity, and the intrinsic resolution limitations that obscure residue-level detail. Topology determination is a key intermediate step toward building atomic protein models from medium-resolution cryo-EM density maps. It requires identifying the correct correspondence and orientation between secondary-structure elements (SSEs), i.e., α-helices and β-strands, predicted from the amino-acid sequence and those detected in the three dimensional (3D) density map. Despite significant advances in cryo-EM reconstruction and molecular modelling, this correspondence problem remains a challenging task, particularly in the presence of noisy density maps and in large, topologically complex α/β proteins. To address this issue, we propose a fully automated, classification-based framework that infers protein secondary-structure topology directly from medium-resolution cryo-EM density maps. Specifically, we cast topology determination as a supervised classification problem in three-dimensional space, leveraging geometric learning on model-derived Cα coordinate representations to establish SSE correspondences, and a Dynamic Time Warping (DTW)-based procedure to resolve density-stick directionality. Validation on a benchmark of 38 proteins spanning both simulated and experimental cryo-EM maps and covering diverse fold classes (α, β, and α/β) demonstrates strong and consistent performance. Among the evaluated predictors, the Voronoi (1-NN) classifier achieves the highest average correspondence quality, with a mean F1-score of 96.82% across the full benchmark. The framework also scales to large, topologically dense targets containing up to 65 secondary-structure elements while preserving very fast correspondence inference (<3 ms), offering a substantial improvement over prior baselines in both accuracy and computational cost. Overall, the classification-driven strategy provides reliable SSE-to-density matching and, when coupled with DTW-based direction selection, yields stronger topology constraints that directly support model building and refinement from medium-resolution cryo-EM reconstructions, while remaining easy to integrate into existing structural interpretation pipelines. 

Accurate prediction of protein sidechain conformations is a fundamental challenge in structural biology, with diverse applications ranging from protein structure determination to computational drug design. The performance of backbone-dependent rotamer libraries is often limited by discrete binning artifacts and difficulties handling sparse conformational regions. In this work, we present a Bayesian framework for rotamer prediction that addresses these limitations through Dirichlet priors and spatial smoothing. Our approach models rotamer probabilities as continuous functions of backbone dihedral angles, using circular Gaussian convolution, to make the most of statistical strength from neighboring conformations while respecting the periodic nature of angular data. We constructed rotamer libraries through structural clustering of sidechain conformations rather than chi angle binning, ensuring that each rotamer represents a distinct three-dimensional geometry. We evaluated and compared our framework against the state-of-the-art libraries on two independent test sets. Our Dirichlet model achieved chi angle prediction accuracy of 59–60%. Notably, our method produced consistently lower angular errors, an approximate 13% reduction in mean deviation, suggesting that the continuous probability distributions better capture subtle conformational preferences. Further, we explored the incorporation of non-sequential context by including the identity of nearby non-neighboring residues as an example of extensibility of our framework. 

One of the challenges and a significant part of a protein structure’s prediction in three-dimensional space is a side chain prediction/packing. This area of research has a large importance, due to its various applications in protein design. In recent years, many methodologies and techniques have been crafted for side chain prediction such as DLPacker, FASPR, SCWRL4 and OPUS-Rota4. In this research, we address the problem from a different perspective. We employed a machine learning model to predict the side chain packing of protein molecules given only the Cα trace. We analyzed 32,000 protein molecules to extract important geometrical features that can distinguish between different orientations of side chain rotamers. We designed and implemented a Random Forest model to tackle this problem. Given the accuracy of existing state-of-the-art approaches, our model represents an improvement from among other models. The results of our experiment show that Random Forest is highly effective, achieving a total average accuracy of 73.7% for proteins and 73.3% for individual amino acids. 

method for structure determination. Despite the substantial growth in deposited cryo-EM maps driven by advances in microscopy and image processing, accurately constructing models from these maps remains challenging. Extracting secondary structure information from EM maps is valuable for cryo-EM modeling. In this context, we introduce a novel deep learning secondary structure annotation framework specifically designed for intermediate-resolution cryo-EM maps, employing a three-dimensional Inception architecture. Testing it on diverse datasets, including maps with authentic intermediate resolutions, demonstrates its accuracy and robustness in identifying secondary structures in cryo-EM maps. We conducted a comparative analysis of our results against frameworks that exist in the state-of-the-art, and our framework demonstrated superior performance across nearly all secondary structure elements. We employed the F1 accuracy metric, yielding an average F1 score of 0.657 for helix, 0.712 for coil, and 0.596 for sheet predictions. Notably, certain helix and sheet predictions achieved an impressive F1 score of 0.881.