PaperlayerView exact revision on arXiv

Source-linked AI summary

CryoLVM: Self-supervised Learning from Cryo-EM Density Maps with Large Vision Models

Weining Fu, Kai Shu, Kui Xu, Qiangfeng Cliff Zhang

arXiv:2602.02620v2q-bio.QMcs.AIcs.LG

TL;DR

CryoLVM addresses the limited scalability and generalizability of task-specific cryo-EM methods by learning transferable representations from experimental density maps. It combines JEPA with an SCUNet backbone and a histogram-based alignment loss, then adapts the model across three downstream tasks, where it consistently outperforms established baselines across multiple metrics.

  • Problem

    Task-specific cryo-EM methods have limited scalability and generalizability, while foundation-model robustness on genuinely noisy, low-resolution experimental maps remains insufficiently validated.

  • Method

    CryoLVM uses JEPA with an SCUNet backbone to learn self-supervised structural representations from experimental cryo-EM density maps and adds a histogram-based distribution alignment loss.

  • Results

    CryoLVM consistently outperforms established baselines across three downstream tasks and multiple evaluation metrics, while its composite loss accelerates convergence and enhances downstream performance.

  • Takeaways & Limitations

    CryoLVM provides a versatile foundation-model approach for cryo-EM density-map processing, including sharpening, super-resolution, and missing-wedge restoration.

  • Takeaways & Limitations

    CryoFM’s robustness and performance on genuinely noisy, low-resolution experimental maps remain unvalidated, and missing-wedge experiments assume an experimental tilt limit commonly around 60°.

Abstract

from arXiv · show

Cryo-electron microscopy (cryo-EM) has revolutionized structural biology by enabling near-atomic-level visualization of biomolecular assemblies. However, the exponential growth in cryo-EM data throughput and complexity, coupled with diverse downstream analytical tasks, necessitates unified computational frameworks that transcend current task-specific deep learning approaches with limited scalability and generalizability. We present CryoLVM, a foundation model that learns rich structural representations from experimental density maps with resolved structures by leveraging the Joint-Embedding Predictive Architecture (JEPA) integrated with SCUNet-based backbone, which can be rapidly adapted to various downstream tasks. We further introduce a novel histogram-based distribution alignment loss that accelerates convergence and enhances fine-tuning performance. We demonstrate CryoLVM's effectiveness across three critical cryo-EM tasks: density map sharpening, density map super-resolution, and missing wedge restoration. Our method consistently outperforms state-of-the-art baselines across multiple density map quality metrics, confirming its potential as a versatile model for a wide spectrum of cryo-EM applications.

1 INTRODUCTION

CryoLVM addresses the need for scalable, transferable cryo-EM representations by applying JEPA with an SCUNet backbone to experimental density maps. It is evaluated across three downstream tasks and reports improvements over established methods on most metrics.

  • Motivation: Cryo-EM data growth and intrinsic imaging pitfalls motivate unified computational approaches for structural analysis.The introduction links increasing density-map deposition and diverse analytical needs with limitations in accurate structure determination.
  • Related approaches: Existing cryo-EM machine-learning methods are predominantly designed for specific stages or tasks in the structural-determination pipeline.Examples include reconstruction models such as cryoDRGN and 3DFlex and postprocessing models such as DeepEMhancer.
  • Research gap: CryoFM’s training and evaluation on curated high-quality or synthetic-noise maps leave robustness on genuinely noisy, low-resolution experimental maps unvalidated.This limitation defines an unresolved gap for cryo-EM foundation models.
  • Contribution: CryoLVM applies JEPA with an SCUNet backbone to learn self-supervised structural representations from cryo-EM density maps for adaptation across downstream tasks.Its pretraining uses masked 3D patches, context and target encoders, and prediction in representation space.
  • Contribution: CryoLVM introduces a histogram-based distribution alignment loss intended to accelerate convergence and improve fine-tuning performance.The contribution is presented alongside the model architecture as a core methodological innovation.
  • Evaluation: Experiments across three downstream tasks show CryoLVM outperforming DeepEMhancer, EMReady, EM-GAN, and IsoNet on most evaluation metrics.The reported evaluation covers density-map sharpening, super-resolution, and missing wedge restoration.

2 RELATED WORK

Related work spans cryo-EM density-map processing methods for sharpening, super-resolution, and missing-wedge restoration. These tasks address resolution loss, degraded low-resolution map interpretation, and artifacts caused by restricted angular sampling.

  • Scope: The related-work review focuses on density-map sharpening, super-resolution, and missing-wedge restoration as the paper’s three downstream applications.These applications are presented as the main areas of computational cryo-EM density-map processing covered in the section.
  • Density map sharpening: Density-map sharpening methods compensate for resolution-dependent amplitude falloff that weakens high-frequency contrast and interpretability.Prior approaches include global B-factor correction and spatially adaptive local methods.
  • Density map super-resolution: Super-resolution methods target low-resolution maps because model building and protein identification degrade as resolution worsens.The cited examples report degradation beyond 4 Å for ModelAngelo and beyond 6–8 Å for CryoDomain.
  • Missing wedge restoration: Missing-wedge artifacts arise from restricted cryo-ET tilt acquisition, typically around ±60°, producing anisotropic resolution and beam-direction distortions.Incomplete angular sampling creates wedge-shaped gaps in Fourier space and compromises reconstruction fidelity.
  • Missing wedge restoration: Prior missing-wedge restoration datasets rotate subtomograms across 20 orientations and impose additional wedge artifacts to create degraded-to-less-degraded training pairs.This setup is used to map corrupted inputs toward less-degraded targets.

3 METHODOLOGY

CryoLVM combines JEPA with SCUNet for self-supervised representation learning on volumetric cryo-EM maps, then adapts the pretrained encoder to downstream tasks. Its training uses masked-region prediction and a composite reconstruction–distribution-alignment loss.

  • Method overview: CryoLVM is pretrained on unlabeled cryo-EM density maps and later fine-tuned with task-specific models for downstream applications.The methodology covers architecture, pretraining, fine-tuning, inference, and the loss function used across the targeted tasks.
  • Architecture: The framework uses hierarchical swin-conv blocks for multi-scale feature extraction and transformer blocks in the Target Predictor.Patch embeddings are combined with 3D sinusoidal positional encodings, while downstream decoders use upsampling SC blocks.
  • JEPA pretraining: JEPA predicts masked-region representations from visible context instead of reconstructing voxels directly, with context, target, and predictor components.The target encoder uses stop-gradient, and the objective minimizes prediction loss on masked patches.
  • Architecture: SCUNet combines Swin Transformer paths for long-range dependencies with residual convolutions for local structural detail.The paper motivates this hybrid design for modeling atomic-scale features and cross-regional spatial relationships.
  • Loss design: A unified composite loss combines mean-squared reconstruction error with the histogram-based distribution alignment loss LHistKL.The composite objective is used across all three downstream tasks.
  • Loss design: LHistKL aligns predicted and target density distributions by constructing differentiable soft histograms and measuring their Jensen–Shannon divergence.Gaussian kernel weighting forms the histograms, and the resulting distributions are compared through a divergence based on KL divergence.

4 EXPERIMENTS

CryoLVM was evaluated across density map sharpening, super-resolution, and missing wedge restoration, with ablations examining its backbone and loss design. It consistently achieved strong task-specific results against established methods.

  • Density map sharpening: CryoLVM achieved the highest sharpening scores across CCbox (0.894), CCmask (0.821), CCpeaks (0.806), and Q-score (0.444).The evaluation used phenix.map model cc and Chimera.MapQ metrics.
  • Density map super-resolution: CryoLVM achieved the best super-resolution resolutions under phenix.mtriage, including dmodel=2.33 ˚A, FSC-0.143=2.58, and FSC-0.5=4.58.CryoLVM also reached 3.39 ˚A with CryoRes versus 3.47 ˚A and 4.18 ˚A for the baselines.
  • Missing wedge restoration: At FSC=0.143, CryoLVM reduced missing-wedge dmin from 10.448 ˚A to 10.094 ˚A, while at FSC=0.5 it reduced dmin from 12.361 ˚A to 11.447 ˚A.These improvements were reported against the IsoNet baseline.
  • Missing wedge restoration: CryoLVM reconstructed continuous pore-like channels and maintained higher FSC correlation across mid-to-high spatial frequencies than IsoNet.The comparison concerns the EMD-5331 missing-wedge case.
  • Ablation studies: SCUNet-based models outperformed ViT counterparts across all evaluated downstream tasks under identical training hyperparameters.The comparison was conducted as an architectural ablation.
  • Ablation studies: The composite MSE and histogram KL loss accelerated convergence and improved downstream performance relative to MSE alone.The ablation evaluated CryoLVM fine-tuning for density map super-resolution.

5 CONCLUSION

CryoLVM is presented as a foundation model for cryo-EM density maps using JEPA with a SCUNet backbone. The authors report robust performance across three downstream tasks and anticipate broader AI use in cryo-EM.

  • CryoLVM combines JEPA with a SCUNet-based backbone to learn rich structural representations from cryo-EM density maps.
  • A histogram-based distribution alignment loss is introduced to accelerate convergence and enhance fine-tuning performance.
  • CryoLVM consistently outperformed established baselines across density map sharpening, super-resolution, and missing wedge restoration.
  • The experiments included genuinely noisy, low-resolution experimental maps encountered in real-world cryo-EM workflows.

C.1 IMPLEMENTATION DETAILS

CryoLVM uses a hierarchical encoder with a task-specific decoder for downstream applications. The modular architecture transfers pretrained density-map representations while preserving multiscale structural information.

  • The Context Encoder and Target Encoder share an identical hierarchical design, with Target Encoder parameters updated by exponential moving average during pretraining.
  • The modular design separates generalizable encoder representations from task-specific decoder specialization.
  • The downstream decoder uses three up-sampling stages, Swin-Conv Blocks, skip connections, and a final convolution to produce output density.

D.1 PREPROCESSING FOR TRAINING ACCELERATION AND MEMORY REDUCTION

The experiments use curated cryo-EM map datasets and task-specific preprocessing, including quality filtering, resolution-controlled targets, angular masking, and patch-based post-processing.

  • Task datasets: Sharpening data were filtered using structural, map-axis, resolution, and sequence-identity criteria to reduce overlap with test models.
  • Task datasets: The super-resolution dataset contains 400 training, 50 validation, and 40 test maps spanning 2.3-6 ˚A resolutions.
  • Task datasets: Missing-wedge inputs used cryo-ET maps below 8 ˚A resolution, angular masks, and a 4 ˚A voxel size for sampling and computational efficiency.
  • Post-processing: Overlapping patch predictions were fused with Gaussian-weighted accumulation and normalization before cropping to the original volume extent.

E EXPERIMENTAL SETTINGS

Downstream experiments use standardized optimization, data-parallel training, cropped density-map volumes, and the SCUNet architecture specified for CryoLVM.

  • Training configuration: Downstream training uses AdamW with a 3 × 10^-5 learning rate, 1 × 10^-3 weight decay, and ReduceLROnPlateau scheduling.Experiments run for 500 epochs with batch size 32, bfloat16 mixed precision, and DDP.
  • Data processing: Input density maps are cropped into 483-voxel volumes, with strides of 24 during training and 48 during validation.
  • Model and hyperparameters: The experimental-settings section includes dedicated specifications for downstream hyperparameters and the SCUNet model architecture.

F EVALUATION METRICS

Evaluation combines resolution, local resolvability, map–model correlation, and backbone-comparison measures to assess cryo-EM density-map quality.

  • Resolution metrics: FSC measures resolution by correlating Fourier coefficients from two independently reconstructed half-maps within frequency shells.Resolution is conventionally defined where FSC falls below thresholds such as 0.143 or 0.5.
  • Resolution metrics: The dmodel metric indicates effective resolution by quantifying agreement between an atomic model and experimental density.
  • Map–model agreement: Phenix reports CCbox, CCmask, CCpeaks, and CCvolume to assess global, masked, peak-focused, and molecular-volume map–model agreement.These metrics probe complementary aspects of consistency between atomic models and density maps.
  • Local resolvability: Q-score measures atomic resolvability by correlating each atom’s local density profile with a reference Gaussian profile.Higher Q-scores indicate better-resolved atomic density.
  • Local resolvability: CryoRes estimates voxel-wise local resolution from a single density map, complementing global metrics with spatially varying map-quality information.
  • Backbone evaluation: SCUNet consistently outperforms ViT across sharpening, super-resolution, and missing-wedge restoration on correlation- and resolution-based metrics.Lower-is-better metrics were negated in the radar summary so all scores use a higher-is-better convention.

G.2 ABLATION STUDY OF COMPOSITE LOSS OF MSE AND HISTKL

Ablation studies examine the composite HistKL loss, backbone choice, pretraining, and overlapping-patch fusion across CryoLVM’s downstream tasks.

  • Composite loss: HistKL combined with MSE consistently outperforms MSE alone across all density-map super-resolution evaluation metrics.
  • Backbone ablation: SCUNet outperforms ViT across sharpening, super-resolution, and missing-wedge restoration on correlation- and resolution-based metrics.
  • Composite loss: 107 epochs versus 279 epochs marks the best validation loss for composite MSE+HistKL versus MSE-only training, a 2.6× reduction in training time.
  • Pretraining ablation: Pretrained-then-finetuned models improve over training from scratch, while JEPA pretraining consistently outperforms MAE across sharpening metrics.
  • Postprocessing ablation: Gaussian-weighted fusion consistently outperforms mean-weighted fusion across all missing-wedge restoration evaluation metrics.The comparison concerns predictions from overlapping sliding-window patches.

G.5 ADDITIONAL RESULTS OF SUPER RESOLUTION TASKS

CryoLVM achieves strong super-resolution results on both global and local resolution measures, including improved local resolution for the EMD-0023 example.

  • Overall performance: CryoLVM outperforms all baselines in both FSC-0.5 and CryoRes metrics for density-map super-resolution.
  • Local resolution: For EMD-0023, the local-resolution range changes from 3.39 Å–4.12 Å in the deposited map to 2.31 Å–3.45 Å in CryoLVM’s prediction.The figure compares local-resolution maps calculated with CryoRes and visualized through Chimera.

G.6 ADDITIONAL RESULTS OF MISSING WEDGE RESTORATION

Additional evaluations show that CryoLVM restores missing angular information while preserving structural consistency and high-frequency detail. A ModelAngelo case study further links CryoLVM sharpening to improved automated atomic model quality.

  • Missing wedge restoration: CryoLVM achieves higher correlation coefficients across CCbox, CCmask, and other complementary criteria for missing wedge restoration.CCbox evaluates global structural coherence across the reconstruction volume.
  • Missing wedge restoration: CryoLVM exhibits significantly better FSC-0.143 and FSC-0.5 resolution distributions, indicating more effective recovery of lost high-frequency information.These metrics assess resolution distributions affected by missing wedge effects.
  • Automated model building: CryoLVM sharpening reduced ModelAngelo RMSD to the reference structure from 0.68 Å to 0.58 Å and increased sequence match from 92% to 94.3%.The comparison used EMD-6656 at 3.5 Å resolution.
  • Missing wedge restoration: The EMD-5106 visualization provides an additional qualitative case for evaluating CryoLVM missing wedge restoration.The map was processed with a wedge-shaped mask.
Loading 2602.02620v2…