Efficient matching of predicted protein structures into high to intermediate resolution CryoEM maps with Slice’N’Dice and ABCFold

Abstract

Protein structures are essential for understanding biological function because the three-dimensional arrangement of amino acid residues determines molecular interactions, specificity, and activity at the atomic level. Cryogenic Electron Microscopy (CryoEM) has become a powerful method for structural determination, as high-resolution maps can support de novo model building, whereas intermediate-resolution maps often require map-fitting approaches in which atomic models are guided by experimental CryoEM densities (maps). While AlphaFold (AF)2 and newer diffusion-based predictors (AF3 and similar) can accurately model individual domains, their predictions become untrustworthy when multiple conformations are possible, leading to uncertain inter-domain orientations. As a result, global domain arrangements may be unreliable even when local confidence metrics indicate high structural accuracy. This creates a clear need for methods that address uncertainties in inter-domain orientations and reliably place predicted models into CryoEM maps. This work primarily presents the software Slice’N’Dice (SnD), a pipeline designed to address these issues by first slicing a predicted model into structural units then using map-fitting tools of MOLREP and PowerFit to place them into the CryoEM map. Data was collected to assess the feasibility of the tool prior to development. The data was sourced from CASP15, a community-wide benchmark evaluating state-of-the-art structure prediction methods. For the CASP15 competition, a cross-correlation (CC)-based ranking of CryoEM map-model placements showed a strong correlation with SCASP15 scores, demonstrating that high-quality predicted models have intrinsic value for map fitting. Model confidence, measured by pLDDT, was shown to act as a reliable pre-fitting indicator of success, and slicing full-chain predictions into domain-like units consistently improved placement quality. A binary classifier was developed to automatically assess atomic model placements in CryoEM maps by integrating multiple map-model fitting scores. Through systematic model selection, feature ablation and hyperparameter tuning, the final model achieved high predictive performance and generalisability, making it suitable for automated use within the SnD pipeline. This work also presents ABCFold, a complementary pipeline designed to facilitate local use of next-generation diffusion-based atomic model predictors: AF3, Boltz-1, and Chai-1. Thus, complementing the SnD workflow by providing accessible means to both model generation and map-model fitting.