Neural spectroscopy of AlphaFold2 reveals encoded protein conformational landscapes

Technical Summary: Neural Spectroscopy of AlphaFold2

Problem Statement
AlphaFold2 (AF2) achieves remarkable accuracy in predicting static protein structures, yet the nature of what its 93 million parameters encode remains opaque. While the model is conventionally treated as a tool for sequence-to-structure conversion, it is unclear whether its weights merely memorize a catalog of known folds or if they internalize a deeper, transferable understanding of protein folding physics and conformational organization. Standard inference yields a single static prediction, compressing the model's internal representational space and obscuring any encoded information regarding conformational heterogeneity, stability ordering, or folding pathways. The central problem addressed is how to systematically probe the learned weights of the Evoformer module to reveal structural constraints and conformational landscapes that are not visible during standard, unperturbed inference.

Methodology: Scaled Gaussian Convolution (SGC)
The authors propose "neural spectroscopy," a protocol involving the systematic, deterministic perturbation of the Evoformer's learned weight tensors to characterize the model's internal encoding. The core method is Scaled Gaussian Convolution (SGC), which modifies the weight tensors ($W_i$) of the Evoformer blocks before inference using three continuous parameters:

1. Scaling Factor ($\lambda$): Uniformly attenuates the magnitude of weights ($\lambda \in (0, 1]$).
2. Gaussian Width ($\sigma$): Applies a Gaussian convolution to smooth fine-grained local weight features while preserving coarse global patterns.
3. Perturbation Depth ($d$): Determines the number of consecutive Evoformer blocks (from block 0 to $d-1$) that are modified.

The perturbation is applied in-place to the model parameters (using OpenFold v2.2.0) and is fully deterministic; the same parameters yield identical outputs. To ensure the observed effects are due to the specific structure of the perturbation rather than mere magnitude, the authors employ matched-power noise controls:

• White Noise: Adds Gaussian noise with the same total energy as SGC but destroys all spectral and phase structure.
• Spectral Noise: Preserves the power spectrum of the SGC perturbation but randomizes the phase coherence.

Experiments were conducted on three distinct protein systems to probe different regimes of learned representation:

• Ubiquitin: A well-folded globular protein with extensive experimental data (NMR, MD simulations) for validation.
• KaiB: A fold-switching protein where AF2 predicts only the active (fold-switched) state.
• $\alpha$-synuclein: An intrinsically disordered protein (IDP) with no single native structure.

To isolate the effect of the learned weights from external coevolutionary constraints, experiments utilized reduced Multiple Sequence Alignment (MSA) depths ($64 \times 64$), making the model more sensitive to weight perturbations.

Key Results

1. Structured Conformational Landscapes:
   SGC induces a tunable, graded structural transition in Ubiquitin. As perturbation depth increases, the model traverses a continuous conformational landscape rather than collapsing randomly. The landscape exhibits an "L-shaped" topology: a deep native basin connected to a denatured ensemble via a hinge of partially unfolded intermediates (molten globules). This topology mirrors the folding funnel observed in millisecond-scale molecular dynamics (MD) simulations, with a 72.6% territorial overlap in the RMSD–Q-factor space.

2. Physically Ordered Unfolding:
   Under progressive perturbation, native contacts in Ubiquitin break in a specific sequence that matches decades of experimental folding studies. The "folding nucleus" (Group 1) is the most robustly encoded and persists longest, while intermediate boundary contacts (Group 2) break first. This ordering is consistent across all five independently trained AF2 models and is not reproduced by noise controls, which produce stochastic, seed-dependent collapses or atomized debris.

3. Correlation with Physical Flexibility:
   Per-residue perturbation sensitivity (the variance in predicted coordinates across recycling iterations under perturbation) correlates strongly with physical flexibility measured by equilibrium MD RMSF. This correlation persists even after controlling for residue burial (Weighted Contact Number), suggesting the weights encode residue-level constraints related to flexibility, not just static geometry.

4. Model Independence and Convergence:
   For Ubiquitin, all five independently trained AF2 models produce the same conformational landscape topology and transition boundaries, indicating the encoding is a property of the architecture and training data rather than a specific weight initialization.

   • KaiB (Convergent Absence): All five models fail to access the ground-state fold under SGC, producing only a one-dimensional denaturation pathway from the predicted fold-switched state. This suggests the specific pathway to the ground state is absent from the learned manifold.
   • $\alpha$-synuclein (Structured Disagreement): The five models produce five distinct but internally coherent landscapes. While they disagree on the specific conformational ensemble (reflecting the underdetermined nature of IDPs), their pooled ensemble covers the core of the conformational space sampled by explicit-solvent MD, suggesting a shared "floor" of constraints.

5. Representational Mechanism:
   Analysis of the Evoformer's pair representation ($z_{ij}$) reveals that SGC shifts pair-activation magnitudes along a single, low-dimensional axis that accounts for 97% of the condition-specific variance. Noise controls project outside this subspace, confirming that the structured response arises from a continuous, weight-aligned modulation of the representational geometry, not generic network failure.

Significance and Claims
The paper claims that AlphaFold2's weights encode a compressed, data-driven representation of protein conformational organization that extends beyond the single static structure predicted during standard inference. By treating the weights as a scientific object rather than just a tool, the authors demonstrate that:

• Emergent Properties: The model learned to predict static structures, yet the conformational organization (stability ordering, flexibility patterns, landscape topology) emerged as a byproduct of that objective, shaped by evolutionary and structural training data.
• Neural Spectroscopy: The approach of reading these encoded constraints via controlled weight perturbation is termed "neural spectroscopy." It allows for the characterization of what the model "knows" about protein folding without requiring explicit training on dynamic or thermodynamic data.
• Limits of Learning: The method reveals the boundaries of the learned manifold. Where training data is strong (Ubiquitin), the landscape is shared and physically accurate. Where data is absent (KaiB ground state), the manifold shows a convergent absence. Where data is ambiguous (IDPs), the manifold shows structured disagreement, with the shared core representing the reliably encoded constraints.

The authors remain modest regarding mechanistic explanations, noting that while the correspondence between weight perturbation and physical processes is robust, the formal theory connecting weight-space geometry to representational construction and conformational emergence is not yet established. They posit that the signal resides in the dominant directions of the learned weights, accessible through global, continuous deformation.