AI Paper Atlas

Research map / Probabilistic and scientific modeling / Molecules and proteins

CryoAtlas: A Large Curated Dataset and Unified Benchmark for Cryo-EM Atomic Model Building

Mingrui Li, Minzhang Li, Weichen Qin, Yufan Xie, Zhangzhi Xiong, Sixian Shen, Jiakai Zhang, Yuan Pei, Jingyi Yu

NeurIPS 2026 · Probabilistic and scientific modeling › Molecules and proteins · approach: Generative protein structure design · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces CryoAtlas, a curated dataset and unified benchmark built from EMDB and PDB entries for density-guided cryo-EM atomic model building.

Problem

Deep learning progress on automated atomic model building from cryo-EM maps is limited by the lack of large, high-quality datasets and standardized evaluation benchmarks.

Approach

A rigorously filtered dataset integrating EMDB maps with PDB atomic models (entries up to March 2026), paired with a unified benchmark for density-guided model building.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
train ModelAngelo CNN on CryoAtlas instead of Cryo2StructData (ablation)ModelAngelo (retrained CNN) · CryoAtlas benchmark (618 targets)Ca RMSD (A): 1.108 -> 0.769Improved
train ModelAngelo CNN on CryoAtlas instead of Cryo2StructData (ablation)ModelAngelo (retrained CNN) · CryoAtlas benchmark (618 targets)GT Score: 0.697 -> 0.755Improved
retrain ModelAngelo CNN on CryoAtlas vs released checkpointModelAngelo · CryoAtlas benchmark (618 targets)GT Score: 0.723 -> 0.755Improved
restore discarded entries to CryoAtlas training set (ablation)ModelAngelo CNN training · CNN training, first 5,000 stepsEMA loss: 0.0898 -> 0.1140Hurt
apply CryoAtlas filters to raw Cryo2StructData (ablation)ModelAngelo CNN training · CNN training, first 5,000 stepsEMA loss: 0.1127 -> 0.0865Improved
use released normalized Cryo2StructData maps instead of raw (ablation)ModelAngelo CNN training · CNN training, first 5,000 stepsEMA loss: 0.1127 -> 0.1475Hurt
lower map resolution (stratification)all benchmarked methods · CryoAtlas benchmark (618 targets)GT ScoreHurt
lower map resolution (stratification)all benchmarked methods · CryoAtlas benchmark (618 targets)Geom ScoreNo clear effect
longer sequence length (stratification)AlphaFold 3 · CryoAtlas benchmark (618 targets)GT Score and Map ScoreHurt
longer sequence length (stratification)density-guided methods · CryoAtlas benchmark (618 targets)Map ScoreNo clear effect
higher train-test sequence identity (stratification)density-guided methods · CryoAtlas benchmark (618 targets)GT ScoreNo clear effect
use sequence-only AlphaFold 3 instead of density-guided methodsmodel building · CryoAtlas benchmark (618 targets)Geom Score vs GT Score: Geom 0.598 vs 0.537; GT 0.552 vs 0.787 (CryoAtom2)Mixed

Experimental setup

Datasets: EMDB PDB

Benchmarks: CryoAtlas

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy