# Connecting the dots: deep learning-based automated model building methods in cryo-EM

## TL;DR

This paper reviews how deep learning is being used to automate model building in cryo-EM, helping scientists determine biomolecular structures more efficiently.

## Contribution

The paper introduces a classification of deep learning methods for cryo-EM model building based on structural hierarchy and model derivation.

## Key findings

- Deep learning methods can automate model building at different structural levels in cryo-EM.
- Hybrid methods combine learned features with structural templates for improved modeling.
- Current limitations include dataset diversity and capturing conformational heterogeneity.

## Abstract

The resolution revolution in single particle cryo-electron microscopy (cryo-EM) has dramatically expanded our structural knowledge of large biomolecular complexes. While high-resolution cryo-EM density maps enable atomic model building, lower-resolution maps can still reveal secondary structures, folds, and domains. When combined with integrative modeling approaches, such data can provide meaningful insights into biomolecular structure and function. Constructing accurate models, however, remains challenging: at low resolutions it is difficult to interpret density maps features reliably, and at high resolutions traditional model-building workflows can become a time-consuming bottleneck. Deep learning, which is transforming problem-solving across scientific domains, offers powerful new tools to automate and accelerate this process. In this review, we discuss deep learning-based methods developed to automate model building in cryo-EM density maps, assessing their impact on streamlining structure determination. Recognizing that biomacromolecular structures exhibit hierarchical organization, we classify these methods according to their ability to model primary, secondary, tertiary, and quaternary structures of biomolecules. Deep learning tools for building atomic models in cryo-EM density maps are further grouped as de novo, where the model is predicted directly from features learned from the cryo-EM density, or hybrid, where it is derived by integrating structural templates with these features. We outline current limitations, including the challenge of obtaining sufficiently large and diverse datasets for training networks to model different types of biomolecules in the cryo-EM density maps, and the open challenge of constructing training sets that capture the conformational heterogeneity often observed in the cryo-EM maps. We conclude by highlighting emerging directions for this rapidly advancing field, which promise to make automated, data-driven model building an integral part of structural biology.