PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep learning based mixed-dimensional GMM for characterizing variability in CryoEM

Muyuan Chen, Steven Ludtke

arXiv:2101.10356v2q-bio.BMcs.LG

TL;DR

CryoEM captures molecular variability, but representing conformational changes from particle images remains challenging. e2gmm maps images to a 3-D Gaussian mixture model and latent space, producing interpretable, data-derived structures and motion trajectories. The examples include an unreported spliceosome motion, while GPU memory constrains model scale and resolution.

  • Problem

    Conformational and compositional variability limits the features captured in a single CryoEM structure, and mapping image changes back to molecular motion is difficult.

    Source: Introduction · p. 2, Methods · p. 3

  • Method

    e2gmm uses deep learning to map 2-D particle images and their derivatives through a latent space to a 3-D Gaussian mixture model.

    Source: Introduction · p. 2, Methods · p. 4

  • Results

    The method produced data-derived spliceosome trajectories, including a motion mode not previously reported for the dataset.

    Source: 50S Ribosome assembly · p. 7

  • Takeaways & Limitations

    Each latent-space point can be rendered as a 3-D structure or checked by reconstructing particles nearby in the space.

    Source: Discussion · p. 9

    GPU memory limits model size and resolution; an 11 GB GPU supports up to 3200 Gaussians at 128 x 128 pixels and batch size 8, often limiting proteins to secondary-structure variability.

    Source: Discussion · p. 10

Abstract

from arXiv · show

Structural flexibility and/or dynamic interactions with other molecules is a critical aspect of protein function. CryoEM provides direct visualization of individual macromolecules sampling different conformational and compositional states. While numerous methods are available for computational classification of discrete states, characterization of continuous conformational changes or large numbers of discrete state without human supervision remains challenging. Here we present e2gmm, a machine learning algorithm to determine a conformational landscape for proteins or complexes using a 3-D Gaussian mixture model mapped onto 2-D particle images in known orientations. Using a deep neural network architecture, e2gmm can automatically resolve the structural heterogeneity within the protein complex and map particles onto a small latent space describing conformational and compositional changes. This system presents a more intuitive and flexible representation than other manifold methods currently in use. We demonstrate this method on both simulated data as well as three biological systems, to explore compositional and conformational changes at a range of scales. The software is distributed as part of EMAN2.

Introduction

CryoEM particles capture heterogeneous conformations, but conventional classification works best for few discrete states and can discard data or depend on human choices. e2gmm maps particle images to an interpretable 3-D Gaussian representation and uses a latent space to describe variability.

Introduction

  • CryoEM particles sample different conformational and compositional states, limiting the features visible in any single reconstructed structure.Analyzing the ensemble can instead yield individually more detailed structures.Source: Introduction · p. 2
  • Multi-model classification works best for a small number of states, while discarding particles judged “bad” can impose human bias and leave an incomplete picture.Manifold methods also face challenges linking their representations to biological interpretation.Source: Introduction · p. 2
  • e2gmm uses deep learning to map 2-D particle data to a 3-D Gaussian mixture model, making conformational and compositional variation directly interpretable.A point on the manifold can be visualized without first reconstructing large particle subsets.Source: Introduction · pp. 2–3

Methods

  • The method represents motion as changes in Gaussian locations and ligand presence or absence as changes in amplitude, with width typically fixed to set representation detail.The inverse mapping from a 2-D image to 3-D Gaussians is underdetermined for a single image, so the method uses an image ensemble and an FRC loss.Source: Methods · p. 3
  • A decoder maps a small latent vector to Gaussian parameters, while an encoder maps 2-D images via their derivatives to latent vectors and supports region-specific variability analysis.The network is trained directly from particle data without ground truth.Source: Methods · pp. 3–4
  • Training first fits the decoder to a neutral structure, then trains the encoder on particle images and orientations so latent vectors represent specimen variability.The authors describe optional subsequent orientation refinement against the dynamic GMM, typically requiring one or two iterations.Source: Methods · pp. 4–5
  • The default 4-D latent vector can represent at least four independent variations, and latent-space reduction can support visualization and segmentation.A sufficiently large ensemble of differently oriented images typically contains information to constrain 3-D structure and motion.Source: Methods · pp. 3, 5

Results

Across three public datasets, e2gmm recovered known structural variability and revealed additional compositional and conformational changes. The examples include discrete ribosome states, domain motions, and coordinated movements in the spliceosome and RBDs.

Results

  • The three public EMPIAR datasets showed mostly known variability plus additional motions consistent with system understanding; experimental validation was beyond the study's scope.The authors also included simulated data to demonstrate representation of multiple motions.Source: Results · p. 5
  • Six visible particle clusters in the 50S ribosome dataset recapitulated known ribosome assembly states.The clusters were identified after UMAP reduction of the 4-D latent vectors, and cluster particles were reconstructed separately.Source: 50S Ribosome assembly · pp. 5–6
  • Within ribosome clusters, reconstructions along selected paths showed introduction of the same 23S rRNA helices, revealing subtler compositional changes.The paths were examined both in a cluster with the central protuberance and in a different cluster without it.Source: 50S Ribosome assembly · p. 6
  • Restricting analysis to particles with the central protuberance present identified an approximately 8-degree tilt of the domain.The analysis allowed Gaussian positions in that domain to vary.Source: 50S Ribosome assembly · p. 6
  • In the spliceosome, localized helicase and SF3b motions could be combined into same-direction or separating trajectories, including a mode not previously reported.The trajectories were represented by 3-D reconstructions of particles near each manifold point, with relative populations assessable within noise limits.Source: 50S Ribosome assembly · p. 7
  • At the RBD, the residue 335–344 helix tilted about 11 degrees; opposite motion in an untargeted neighboring subunit suggested coordination, while the focused subunit had stronger reconstructed density.The neighboring subunit's motion was less clear along the second PCA eigenvector.Source: 50S Ribosome assembly · p. 8

Discussion

The Gaussian representation reduces model size while supporting interpretable descriptions of structural variability. Its use remains bounded by orientation and neutral-structure assumptions and current GPU memory limits.

Discussion

  • At 13 Å, 2048 Gaussians require 10,240 variables versus 592,704 voxel parameters, while average FSC remains above 0.95 at the considered spatial frequencies.The authors describe this as a good representation of the density map.Source: Discussion · p. 8
  • Gaussian parameter changes represent ligand binding and domain motion as simple trajectories, reducing variability-model complexity and easing encoder-decoder training.The authors describe Gaussians as natural models of low-resolution CryoEM maps.Source: Discussion · p. 8
  • Gaussian properties avoid common image-processing artifacts and preserve smoothness during continuous motion, including rotation and non-integral translation.Selecting domain Gaussians for updating leaves all Gaussian functions in projection images; constraints can also be applied in real or Fourier space.Source: Discussion · p. 9
  • Each conformational-space point maps to Gaussian parameters for a complete 3-D structure, allowing Gaussian maps and nearby particle reconstructions to be compared.The authors note this can help visualize putative changes when too few particles are available for a sufficiently resolved reconstruction.Source: Discussion · p. 9
  • e2gmm requires particle orientations and a sufficiently rigid molecular portion to support a reasonable neutral structure and orientation assignment.Severe heterogeneity can make initial alignment fail, leaving projection mismatch that may reflect misalignment rather than conformational change.Source: Discussion · p. 9
  • An averaged particle structure may omit regions broadened by large domain motion, leaving no Gaussians there to move; ligand-variable models need density for every ligand in the training volume.A stronger neutral structure can address this; building the neutral Gaussian model from aligned particles is another option, but costs time and may be less robust.Source: Discussion · pp. 9–10
  • With 11 GB of GPU memory, the reported setup supports up to 3,200 Gaussians at 128 × 128 sampling and batch size 8, limiting many proteins to secondary-structure-scale variation.The authors say this would suffice for a 50S ribosome at about 8 Å resolution, or smaller proteins at proportionally higher resolution.Source: Discussion · p. 10
  • e2gmm is distributed in EMAN2, while adapting it to CryoET remains technically challenging because of noisy individual tilts and increased coordinated image data.The paper identifies CryoET as a possible future application.Source: Discussion · p. 10

Figures

The figures trace e2gmm’s training design and show its use to explore compositional and conformational variability across ribosome, spliceosome, and spike-protein datasets.

Figures

  • Fig. 1 contrasts decoder training for the neutral map with full-network training for system heterogeneity.Source: Figure 1 · p. 11
  • Fig. 2 pairs a clustered 2-D embedding of particles from a 4-D latent space with 8 Å-filtered class-average structures.Source: Figure 2 · p. 12
  • Fig. 3 links sampled locations in the ribosome conformation space to reconstructed structures and central-protuberance motion trajectories.Source: Figure 3 · p. 13
  • Fig. 4 shows spliceosome particle distributions, two Gaussian-coordinate motion vectors, and reconstructed structures along those vectors and their combinations.Source: Figure 4 · p. 14
  • Fig. 5 follows the target spike-protein RBD along two eigenvectors and shows the neighboring, untargeted RBD along the first.Source: Figure 5 · p. 15

Gaussian representation of protein density maps

The method represents density as a sum of Gaussian functions and projects that model into particle-image coordinates. It scores agreement between projections and images using Fourier ring correlation.

Gaussian representation of protein density maps

Neural network structure and parameter selection

The network uses an encoder and decoder, with adjustable training parameters and Gaussian-model size. Defaults specify network architecture and optimization settings, while model detail is increased iteratively toward the target resolution.

Neural network structure and parameter selection

Tests on simulated datasets

Simulations test recovery of linear motion, cyclic rotation, and combined compositional and conformational variability. They also identify an orientation-related bias in particle-level conformation estimates.

Tests on simulated datasets

  • The linear-motion simulation’s latent variable agreed well with the known position along the trajectory.Source: Tests on simulated datasets · p. 20
  • Off-diagonal particle estimates can be biased toward the neutral conformation when some orientations make the simulated planar motion unobservable.The passage states that sufficiently diverse orientations can still constrain the manifold accurately.Source: Tests on simulated datasets · p. 20
  • For simulated cyclic rotation, the latent-space particle distribution roughly formed a circle, whose polar angle correlated well with the ground-truth rotation angle.Source: Tests on simulated datasets · p. 20
  • In the mixed simulation, two roughly parallel latent-space curves distinguished particles with and without added density, while position along each curve represented linear domain motion.The compositional difference and domain motion were independent in this simulation.Source: Tests on simulated datasets · p. 20

Additional data processing details for tests on real data

The real-data processing setup varied by dataset: ribosome particles were split into independent subsets, while spliceosome and spike data used their respective supplied particles and structures for refinement.

Reproducibility

The authors tested reproducibility by independently processing even and odd subsets of EMPIAR 10076. Despite stochastic differences in learned spaces, the resulting clusters and matched class structures were consistent.

  • The entire pipeline, from single-model refinement through Gaussian-model generation and heterogeneity analysis, was run independently on evenly split subsets using the same GMM parameters and training process.Source: Reproducibility · p. 21
  • The even/odd EMPIAR 10076 subsets yielded equivalent cluster counts, cluster arrangements, and particles per cluster; matched class averages were highly consistent, with corresponding FSC extending beyond 4Å.Source: Reproducibility · p. 21
  • The authors state functional reproducibility is possible but not guaranteed, and suggest even/odd splitting as a reasonable test of biological conclusions.Source: Reproducibility · p. 21

Computational requirements

The reported workloads ranged from about two to three GPU-hours for heterogeneity analysis, with additional CPU time for embedding, clustering, and map reconstruction.

  • For EMPIAR-10076, each of two heterogeneity-analysis rounds took about 3 GPU-hours, followed by less than 1 hour on a 12-core workstation for embedding, clustering, and map reconstruction.Source: Computational requirements · p. 21
  • For EMPIAR-10180, heterogeneity analysis required about 3 GPU-hours and about 30 CPU-hours to reconstruct density maps along four motion trajectories.Source: Computational requirements · p. 21
  • For EMPIAR-10492, heterogeneity analysis required about 2 GPU-hours and about 10 CPU-hours, using the provided Relion alignment.Source: Computational requirements · p. 22

Comparison to existing methods

The paper compares e2gmm with CryoDRGN and 3DVA on shared datasets, noting differences in speed, representation, and the heterogeneity each method captures. It emphasizes e2gmm’s interactive latent-space interpretation and reports additional motions identified in the ribosome and spliceosome.

  • 3DVA’s linear-subspace model is fast and accurate when variability is linear in voxel intensity, but its linearity makes large-scale nonlinear trajectories difficult to represent.On the spliceosome, 3DVA took about 3 GPU-hours, similar to e2gmm and 20x faster than CryoDRGN; on the ribosome assembly dataset, six of its seven classes matched e2gmm’s six classes.Source: Comparison to existing methods · p. 22
  • CryoDRGN can generate structures without a reference 3D density map and captured a 70S ribosome impurity comprising less than 1% of particles, which 3DVA and e2gmm did not immediately detect.Source: Comparison to existing methods · p. 22
  • Across the benchmark datasets, e2gmm was roughly as fast as 3DVA and 10–20x faster than CryoDRGN while producing qualitatively similar results, despite tests using a consumer GPU.The e2gmm tests used a GTX2080 with about one-third the memory and substantially lower performance than hardware used in the other benchmarks.Source: Comparison to existing methods · p. 22
  • With e2gmm, users can interactively inspect decoded Gaussian coordinates and apply a smooth mask to model selected Gaussians without edge artifacts.The authors report identifying the ribosome central-protuberance tilt and independent helicase and SF3b movements in the spliceosome, which earlier work had not described.Source: Comparison to existing methods · p. 23
Loading 2101.10356v2…