Quasi-Static Deformation Libraries for Cryo-EM Heterogeneity

Cryo-electron microscopy (cryo-EM) freezes thousands of copies of a protein in thin ice and photographs each one as a very noisy 2D projection. Because the copies are caught in different shapes, the images also hold information about how the molecule moves. The most physically grounded current method deforms a known atomic model along its "normal modes", the small harmonic wiggles computed from the curvature of the energy around one fixed structure. Those wiggles are linear, so the method cannot represent a domain that snaps from one stable pose to another. Glass physics has a simple procedure built for exactly this situation. Athermal quasi-static (AQS) deformation pushes a disordered solid by a tiny amount, lets it relax to the nearest energy minimum, and repeats. Along the way it records the points where a minimum disappears and the system jumps to a new one.

We propose to run the AQS procedure offline on a coarse model of the protein. Each run drives the protein slowly along its softest directions, forward and backward. The result is a library of shapes: smooth branches joined by recorded snaps, each stretch made of physically relaxed structures. We then fit cryo-EM images by locating each particle on this branched library, plus a small local correction, instead of on a straight line of normal-mode amplitudes. The claim is that when molecules make large or switch-like motions, this library recovers the true atomic shapes more accurately and more robustly to noise than linear modes or unconstrained neural decoders. In this proposal glass physics supplies the method, and cryo-EM heterogeneity reconstruction is the application.

Core Claim

aNiMAte maps each image to amplitudes of the lowest normal modes of an initial atomic model $X_0$. It projects the deformed model and compares it with the image directly, which is why it explains 74% of spliceosome variance with two components on EMPIAR-10180. Its stated limitation is that the modes come from one Hessian at $X_0$, so large hinge motions and barrier crossings are outside its reach. In glass physics, AQS work shows that such crossings have a specific structure. Under slow driving a basin flattens until its curvature vanishes, an event called a saddle-node bifurcation. The system then drops into a neighbouring basin, and forward and reverse driving trace different branches, a hysteresis loop (Networks and Hierarchies: How Amorphous Materials Learn to Remember).

Our claim is that this piecewise-smooth, branch-and-jump structure is the right shape prior for large-motion cryo-EM heterogeneity. It can be computed cheaply ahead of time with a coarse anharmonic protein model. What is new relative to aNiMAte is that the deformation space is a graph of relaxed structures connected through bifurcations, not a linear span around $X_0$. What is new relative to CryoDRGN is that the latent coordinate indexes physically relaxed atomic models. CryoDRGN's latent space is learned purely to reconstruct images, and its interpolated densities need not be physical.

Research Program

The first component is the library builder. We start from a deposited atomic model and use a Cα-level elastic network whose springs can break and reform. Breakable springs make the energy anharmonic and allow multiple minima; a plain harmonic network has a single minimum by construction and could never produce a snap. For each of the lowest $k$ normal modes of $X_0$ (we expect $k$ around 5 to 10), we define a driving coordinate. For a hinge mode, this is the projection of the displacement onto that mode, measured between the centroids of the two domains the mode separates.

One AQS step has two parts. First we impose an affine increment $\Delta\gamma$ of the driving coordinate, meaning every atom is shifted proportionally along the mode. Then we minimize the energy of all other degrees of freedom with that coordinate held fixed, using the FIRE minimizer standard in jamming studies (Emergent SO(3) Symmetry of the Frictionless Shear Jamming Transition). After each relaxation we track the lowest nonzero eigenvalue $\lambda_{\min}$ of the constrained Hessian. A drop of $\lambda_{\min}$ to zero, followed by a discontinuous jump in coordinates, marks a bifurcation. We store the structures just before and just after it.

We sweep each driving coordinate forward and back, and then reverse it from intermediate points to expose nested sub-loops. This produces a directed graph of relaxed structures, analogous to the mesostate networks of the Mungan et al. work above. A modest compute budget should cover a few thousand AQS steps per mode on a few-thousand-bead model.

The second component is the image model. It keeps aNiMAte's differentiable projector and its image-domain loss. The encoder outputs three things: a soft choice of branch $b$; a continuous position $s$ along that branch, decoded by interpolating between neighbouring relaxed structures; and a few small amplitudes $a$ along local normal modes recomputed at that branch point. The deformed model is

$$X(b,s,a) = X_b(s) + \sum_{j=1}^{m} a_j\, u_j^{(b,s)},$$

where $X_b(s)$ is the relaxed library structure at position $s$ on branch $b$, $u_j^{(b,s)}$ are the $m$ lowest local modes there, and $a_j$ are kept small by a prior. When the library contains only one branch with $s \equiv 0$, this reduces exactly to aNiMAte. That reduction makes the comparison clean.

The third component is the analysis. We test whether recovered branch occupancies and jump locations match ground truth on synthetic data. On experimental data, we check whether branches correspond to recognizable discrete switches rather than arbitrary cuts.

Baselines

The naive reference point is a single consensus reconstruction with no heterogeneity model. It sets the floor for per-image Fourier shell correlation (FSC), the standard measure of agreement between a reconstruction and ground truth at each spatial frequency. Any method that cannot beat the consensus map on per-image FSC is modelling noise.

The strongest physically grounded competitor is aNiMAte. It shares our forward model, so the only difference is the deformation space. On EMPIAR-10180 it reports 74% of structural variance explained with two principal components, against about 30% for map-based methods. We will rerun it on every dataset with the same atomic starting model, the same fixed consensus poses, the same encoder architecture, the same number of free deformation parameters (matching $k$ linear modes against branch, position and $m$ local modes), and the same training epochs and hyperparameter search budget.

The strongest unconstrained competitors are CryoDRGN and the methods ranked in CryoBench. The best number in the file for that suite is RECOVAR's Adjusted Rand Index of 0.968 on Ribosembly. Ribosembly involves subunits appearing and disappearing, which our method cannot represent, so we report it only as an honest failure case. For the continuous sets IgG-1D, IgG-RL and Spike-MD there is no reported result on Lacuna for per-image FSC or for neighbourhood similarity (a CryoBench measure of how well the learned latent space preserves which particles are close in true conformation). We will run CryoDRGN and RECOVAR ourselves using CryoBench's protocol.

Two ablations isolate the key ingredient. The first, "extrapolated modes", builds the library by stepping along the $X_0$ modes without relaxation; it tests whether the energy-minimization step matters. The second, "smooth relaxed paths", keeps relaxation but forbids branching by taking only the forward sweep and removing every structure on the far side of a detected bifurcation; it tests whether the hysteresis structure matters.

Beating a baseline means three things. First, lower Cα RMSD to the ground-truth atomic models on synthetic data, with confidence intervals across three seeds. Second, higher per-image FSC. Third, the advantage must grow, or at least not shrink, as the signal-to-noise ratio (SNR) falls toward 0.001.

Why not just…?

Why not just use more normal modes, or recompute them iteratively, inside aNiMAte? This is the most serious threat. Adding modes buys flexibility at the cost of noise-fitting. Iteratively recomputing modes from the current fit is close to our smooth-relaxed-paths ablation, done during training instead of beforehand. If that ablation matches the full method, the branch-and-jump structure contributes nothing, and the honest conclusion is "relaxed nonlinear paths suffice." The decisive data are the switch-like benchmarks described below, where linear spans must blur two basins together.

Why not run molecular dynamics, or enhanced-sampling simulation, and fit images to the resulting ensemble? Large-machine rearrangements happen on timescales that molecular dynamics struggles to reach. AQS also recovers the slow-driving limit of finite-temperature dynamics at far lower cost (Plasticity and dynamical heterogeneity in driven glassy materials). The cost of our choice is that a coarse potential may miss the true barrier. We therefore evaluate on Spike-MD, which is generated by all-atom molecular dynamics, so the test does not reward agreement with our own potential.

Why not just add a steric-clash penalty to CryoDRGN? A penalty removes overlapping atoms but does not stop intermediate densities from blending two poses. CryoBench reports coordinate networks failing on the flexible-linker IgG-RL set even without noise. We will run CryoDRGN with an added clash-penalty term as an extra comparison on IgG-1D and IgG-RL.

Why not discrete 3D classification? Discrete classes split continuous branches arbitrarily and discard intermediates (CowScape). Our model keeps continuous motion along each branch and uses discreteness only where the physics produces it. We include a K-class baseline on the switch benchmarks, with K equal to the true number of basins.

Proposed Experiments

The decisive synthetic test is a switch benchmark that our potential did not generate. We take pairs of deposited endpoint structures, such as the open and closed gating states in the CLC-2 chloride channel study. We connect them with an independent all-atom targeted-molecular-dynamics trajectory and render particles weighted toward the two ends, so most mass sits in two basins with sparse intermediates. Images are simulated at the CryoBench noise levels down to SNR 0.001.

We will also run IgG-1D, a large rigid rotation that is strongly nonlinear in Cartesian coordinates, and Spike-MD. The prediction is that the full method beats aNiMAte on Cα RMSD and per-image FSC on the switch benchmark and IgG-1D. It should also beat the smooth-paths ablation specifically on the switch benchmark, and at least equal aNiMAte's 74% variance explained on EMPIAR-10180. If the full method fails to beat the smooth-paths ablation on the switch benchmark, we drop the branching claim.

IgG-RL is a stress test that we expect may remain unsolved, since a disordered linker has no small set of switches. If the method fails there, that is informative about the limits of the prior.

On experimental data we fit EMPIAR-10180 and EMPIAR-10028 (the P. falciparum ribosome). On the ribosome we check whether the head swivel and L1-stalk motion fall onto separate branches. As a held-out check, we test whether branch structures fit half-maps reconstructed from their assigned particles better than equally sized random particle sets.