From Point Estimates to Distributions: A Closer Look at EigenFold

Introduction

The success of AlphaFold2 reframed protein structure prediction as a problem that machine learning could, in large part, solve. Yet the framing itself contains a quiet assumption that deserves scrutiny: that a protein sequence maps to a single structure. In a crystal or a cryo-EM map, we observe one conformation, but in the cell a protein samples a distribution of states. It bends, switches folds, and rearranges upon binding partners. Function frequently lives in this conformational flexibility rather than in any single static snapshot. The question, then, is not only "what does this protein look like" but "what is the distribution of shapes this protein adopts."

The paper EigenFold: Generative Protein Structure Prediction with Diffusion Models by Jing, Erives, Pao-Huang, Corso, Berger, and Jaakkola (MLDD workshop, ICLR 2023) addresses precisely this gap. It is, by the authors' framing, the first diffusion generative framework for predicting protein structures and structural ensembles directly from sequence. In this article I want to examine the central methodological idea, the harmonic diffusion process, situate it within the broader landscape of molecular generative models, and offer my own reading of what the results do and do not tell us.

The Core Idea: Diffusion Along Eigenmodes

Standard diffusion models corrupt data with isotropic Gaussian noise and learn to reverse that corruption. For images this works well, but for molecular structure isotropic noise wastes effort: it ignores the strong geometric priors that govern how atoms can move relative to one another. A protein backbone is a connected chain, and neighboring residues are physically constrained. Treating each coordinate as independently noised throws away this structure.

EigenFold's central contribution is harmonic diffusion. The molecule is modeled as a system of harmonic oscillators, essentially a network of springs connecting residues. The forward diffusion process is then defined relative to the eigenmodes (normal modes) of this harmonic system rather than relative to raw Cartesian coordinates. This is the conceptual pivot, and it has an elegant consequence.

When you decompose a harmonic system into its normal modes, the low-frequency modes correspond to large, collective, global motions of the chain, while high-frequency modes correspond to small, local fluctuations. Because the diffusion is defined along these modes, the reverse generative process naturally proceeds in a cascading-resolution fashion:

• Early denoising steps resolve the low-frequency modes, fixing the rough global fold.
• Later steps resolve the high-frequency modes, refining local geometry and detail.

This is reminiscent of coarse-to-fine generation in other domains, but here it emerges from the physics of the harmonic prior rather than being imposed by hand. The practical payoff the authors report is efficiency: EigenFold samples structures in as few as 100 inference steps, scaling to 300 depending on protein size. That is a meaningful number for diffusion models, which often require many more denoising steps.

The other architectural choice worth noting is that EigenFold does not learn structure from scratch. It consumes node and edge embeddings from OmegaFold, in effect converting a powerful deterministic predictor into a generative model. The score model predicts update "forces" at each step, which are then combined with the harmonic constraints to move residue positions. This is a pragmatic decision: rather than relitigating the representation learning problem that AlphaFold-class models already solved, EigenFold reuses those representations and adds the distributional layer on top.

What the Results Actually Show

On recent CAMEO targets, EigenFold reaches a median TMScore of 0.84. This places it in the neighborhood of established methods such as RoseTTAFold for single-structure accuracy. The headline here is not that it beats AlphaFold2; it does not claim to. The headline is that you can obtain near-comparable single-structure quality while also getting a distribution of structures rather than a scalar confidence estimate.

This distributional output is where the more interesting claims live. The authors show that the variation among sampled structures is indicative of model error relative to the ground truth, across several accuracy metrics. This is a genuinely useful property. Existing predictors emit per-residue confidence scores such as pLDDT, but those are learned calibration signals. EigenFold's spread among samples is a more intrinsic form of uncertainty quantification: regions where the model is uncertain produce divergent samples, and that divergence correlates with where the model is actually wrong.

The most ambitious evaluation targets conformational heterogeneity directly, using two datasets:

• A set of fold-switching proteins (Chakravarty and Porter, 2022), which adopt two or more distinct folds.
• A set of ligand-induced conformational changes (Saldaño et al., 2022).

Here the authors are admirably honest. The results are described as mixed. Properties of the sampled structures are moderately correlated with properties of the ground-truth conformations, but they do not predict them to high accuracy. EigenFold does not, by the authors' own assessment, close the gap between single-structure prediction and full ensemble prediction.

My Take: The Framing Matters More Than the Numbers

I think the lasting contribution of EigenFold is conceptual rather than benchmark-topping, and that is fine. Several points stand out to me.

First, harmonic diffusion is a transferable idea. The principle that the noising process should respect the geometry and connectivity of the data, rather than treating coordinates as independent, is broadly applicable. We have seen related moves in torsional diffusion for molecular conformers (Jing et al., 2022) and in other non-Euclidean diffusion formulations. The eigenmode decomposition specifically connects machine learning to the long tradition of normal mode analysis in structural biology, where low-frequency modes have been used for decades to describe functional motions. EigenFold quietly rediscovers, in a learned generative setting, why structural biologists have long cared about these modes.

Second, the dependence on OmegaFold embeddings cuts both ways. It makes the method practical and competitive immediately, but it also means EigenFold inherits the inductive biases of a deterministic single-structure predictor. If those embeddings encode primarily a single dominant conformation, then the generative diversity EigenFold can express may be fundamentally bounded by what the upstream model "believes." This is, I suspect, part of why the fold-switching results are only moderate. You cannot easily sample a second fold if the conditioning information has collapsed onto the first.

Third, there is a deeper question about what distribution is even being modeled. EigenFold is trained on the PDB, which is a collection of experimentally resolved structures. That distribution reflects what is crystallizable and observable, not the true Boltzmann distribution of conformations in solution. A model trained to reproduce PDB structures is learning the distribution of deposited structures, which is a biased and sparse sample of the real conformational ensemble. Bridging to genuine thermodynamic ensembles will likely require integrating physical simulation data or experimental observables such as NMR or hydrogen-deuterium exchange.

Conclusion and Open Questions

EigenFold should be read as an early and thoughtful step toward generative protein structure prediction rather than a finished solution. Its harmonic diffusion formulation is a clean piece of methodology that aligns the noising process with the physics of the molecule, yields an efficient cascading-resolution generative process, and produces uncertainty estimates that track real error. The honest reporting of mixed results on conformational change is, to me, a sign of a serious paper rather than a weakness.

Several open questions follow naturally:

• Can the conditioning be decoupled from a single-conformation predictor, perhaps using sequence-only embeddings or MSA subsampling, to unlock genuine multi-state sampling?
• Can the training signal incorporate dynamical data so the learned distribution approaches the Boltzmann ensemble rather than the PDB distribution?
• How does sample diversity scale with protein size and disorder, and can it serve as a reliable detector of intrinsically disordered regions?

The shift from point prediction to distributional modeling mirrors the trajectory diffusion already traced in image and docking domains. EigenFold marks the moment that trajectory reaches protein structure in earnest. The remaining distance, from sampling plausible alternatives to faithfully reproducing functional ensembles, is where the next round of progress will be measured.