Models that design a protein sequence to fit a given 3D shape, a task called inverse folding, are trained on near-perfect crystal structures. They are then used on predicted or computer-designed shapes, whose errors are not random. Whole regions bend, slide, or shift together. The standard defense, from Learning inverse folding from millions of predicted structures, adds independent 0.1 Å Gaussian jitter to every atom during training. We propose replacing that jitter with perturbations borrowed from high-pressure physics. When the liquid that surrounds a sample in a pressure cell freezes, the sample no longer feels uniform squeezing. It feels a mix of uniform compression, directional shear, and pressure that varies across the sample. Each of these produces smooth, correlated displacement fields rather than independent atom-by-atom noise. Our claim is that multi-chain protein complexes are most fragile exactly where such fields act: at the interface between chains, where a small correlated slide of one chain against another rearranges every contact at once.
In two sentences: inverse folding models are made robust using noise that bears no resemblance to the real errors in their inputs. We train and stress-test them instead with physically shaped strain fields, and ask whether this improves sequence design at multi-chain interfaces, especially on predicted and novel structures, beyond what larger isotropic noise or rigid chain wobbles achieve.
Core Claim
High-pressure physics already separates a stress state into a uniform (hydrostatic) part, a shape-changing (deviatoric, i.e. shear) part, and spatial gradients. It also shows that the non-uniform parts cause the measurement artifacts. Under frozen pressure media they broaden diffraction peaks and hide real lattice anomalies (High pressure route to high-Tc superconductivity in the FeSe-based materials). They shift apparent transition pressures away from idealized calculations (Bi et al.). They can even create new local atomic arrangements through shear-induced defects, as in WB₂ (Creating superconductivity in WB2 through pressure-induced metastable planar defects). The lesson we carry over is that matched-magnitude perturbations with different geometry have very different consequences.
Concretely, we perturb a complex with coordinates $x_i$ (residue $i$, centred on the complex) by
$$ u(x_i) = \left(\epsilon_h I + E_d\right) x_i + \tfrac{1}{2}\, G : (x_i \otimes x_i) + R_{c(i)} . $$
Here $\epsilon_h$ is a scalar uniform compression and $E_d$ is a random symmetric traceless matrix (pure shear). $G$ is a third-order tensor giving a linear strain gradient, contracted against the outer product of the position with itself. $R_{c(i)}$ is a small rigid-body motion applied to the whole chain $c(i)$ containing residue $i$. We sample each term with its own scale and rescale the total so its root-mean-square displacement matches a target value. That lets us compare it to isotropic noise at equal magnitude. The prediction is that, at equal root-mean-square displacement, sequence recovery at interface residues falls much more under the structured field than under isotropic jitter. We further predict that training on the structured field improves recovery on predicted and novel complexes, whose real errors are correlated.
Relation to Prior Work
Three works supply the multi-chain setting and the baselines. Learning inverse folding from millions of predicted structures joins chains into one sequence with gap tokens and trains on 12 million AlphaFold2 structures. It identifies 0.1 Å isotropic jitter as a critical regularizer. We keep its gap-token handling of multiple chains but replace its noise model. ProteinInvBench standardizes evaluation across CATH, multi-chain PDB, and the novel-fold set CASP15 TS45. It reports that recovery drops to about 40% on novel folds, which motivates our focus on inputs that are not crystal structures. KW-Design refines designs using pretrained sequence and structure priors and is the strongest multi-chain recoverer. All three assume a fixed backbone, and none models how correlated deformations of that backbone affect interface design. That gap is listed as an open problem in the file.
Two works inform the mechanics. GenUnfold shows that encoding a protein as a spring network of distance, stiffness, and resistance maps improves prediction of how it unfolds under mechanical force. However, it handles only single chains under pulling and produces a 1D force curve. We use mechanics differently: as a training perturbation for a design model, applied to complexes under compression and shear. Long-Range Neural Atom Learning for Molecular Graphs shows that message passing between distant atoms needs global communication hubs, analogous to Ewald summation. This constrains our design. Local graph models may be unable to "see" a smooth field spanning the complex. We therefore test one local graph model and one model with global attention.
A new benchmark for machine learning applied to powder X-ray diffraction is a cautionary parallel from the materials side. A model trained on clean simulated patterns fell from 82.79% to 35% top-5 accuracy on 20 real samples, largely because of peak broadening and shifts it never saw. This is the failure we want to prevent in protein design: training noise that does not resemble deployment distortions.
Research Program
We first build the perturbation generator from the equation above. It draws shear axes uniformly at random and can optionally align them with the vector between the chain centres. We control the share of total displacement from each term (uniform compression, shear, gradient, rigid chain motion). The generator also offers a "stiffness-aware" variant that scales local displacement down in buried, densely packed regions. Stiffness is estimated from contact density, in the spirit of GenUnfold's stiffness map, so the core moves less than loops and surfaces.
Second, we run a diagnostic study with no retraining. We measure how released inverse folding models degrade as perturbation size grows from 0.1 to 1.5 Å root-mean-square, comparing isotropic noise against each structured component. We break results down into core, surface, and interface residues, defining interface residues as those with a Cα atom within 8 Å of a Cα on another chain.
Third, we retrain two architectures under identical budgets with each noise model. One is a local graph model in the ProteinMPNN/PiFold family from the ProteinInvBench codebase. The other is a GVP-Transformer-style model with global attention.
Fourth, we evaluate on inputs whose errors are naturally correlated. These are AlphaFold2- or ESMFold-predicted structures of held-out multi-chain complexes, plus the novel TS45 set.
Baselines
The naive reference point is the same architecture trained with no coordinate noise. It tells us how much any augmentation helps.
The incumbent regularizer is isotropic Gaussian jitter at 0.1 Å, as in Learning inverse folding from millions of predicted structures. A tuned variant searches isotropic noise from 0.05 to 1.0 Å with the same number of tuning runs we give our own scale search. Structured noise beats it only if it wins against the best isotropic setting, not just the 0.1 Å default.
The strongest existing system for multi-chain recovery is KW-Design. ProteinInvBench reports it reaching "up to 66%" recovery on the PDB multi-chain task, and the KW-Design paper reports above 60% recovery across CATH, TS50, TS500, and PDB. For the global-attention reference, the GVP-Transformer paper reports 51% native recovery on CATH and a 0.69 Spearman correlation on SARS-CoV-2 receptor-binding-domain binding data. We found no reported result on Lacuna for ProteinMPNN's multi-chain recovery under predicted-structure inputs, so we measure it ourselves. KW-Design is run on the identical splits and inputs. Its refinement modules are left unchanged, and we test whether adding structured augmentation to its base model helps. This isolates the noise model from the pretrained priors.
Two ablations isolate the key ingredient. The first uses rigid chain motion only: $R_{c(i)}$ on, all strain terms off. The second uses strain only, with no chain motion. We count as a win a gain of at least 2 points of interface recovery on predicted complexes over the best-tuned isotropic model, consistent across three seeds, with no loss above 0.5 points on clean CATH.
Why not just…?
Why not just turn up the isotropic noise? Larger isotropic noise blurs every local geometry equally. It degrades the precise backbone angles that core packing depends on, so it should cost clean-structure recovery for little interface gain. The tuned isotropic baseline settles this directly. If the best isotropic level matches structured noise on predicted complexes, the geometric claim fails and we will report that.
Why not just train on predicted structures? Hsu et al. already did this at scale, and recovery on novel folds still sits near 40% per ProteinInvBench. Predicted training structures teach the predictor's own typical errors. They do not teach the model to be robust to the shear and inter-chain slides that occur in designed or poorly predicted interfaces. We test this by fine-tuning on predicted complexes with and without structured augmentation.
Why not just jitter whole chains rigidly? This is the most serious threat. Much of the interface fragility may come from chains sliding as rigid units, and rigid jitter is simpler than any strain tensor. The rigid-only ablation is designed to expose this. If it captures most of the gain, the proposal reduces to "augment with inter-chain rigid motion." That is still a useful and testable finding, and we would state it plainly rather than credit the strain terms.
Proposed Experiments
The decisive check is on held-out multi-chain PDB complexes, comparing their crystal structures with their AlphaFold2/ESMFold predictions. We test whether training with structured noise narrows the gap in interface recovery between crystal and predicted inputs, more than tuned isotropic noise or rigid chain jitter does. Secondary checks cover TS45 recovery, perplexity on CATH 4.2 (to rule out clean-data regression), and zero-shot Spearman correlation on SARS-CoV-2 binding data.
A mechanistic check tests whether the deformations that hurt most resemble real prediction errors. For each predicted complex we fit the equation above to the true difference between predicted and crystal coordinates and report how much of that error the strain and rigid terms explain. If real errors contain little shear, we expect, and would report, that the rigid ablation wins.
We also run a comparison between the local graph model and the global-attention model. If only the global-attention model benefits, that would support the Neural Atoms argument that smooth, complex-wide fields need long-range communication to be learned.
References
- Learning inverse folding from millions of predicted structures
- High pressure route to high-Tc superconductivity in the FeSe-based materials
- Pressure-induced structural transitions in europium to 92 GPa
- Creating superconductivity in WB2 through pressure-induced metastable planar defects
- ProteinInvBench: Benchmarking Protein Inverse Folding on Diverse Tasks, Models, and Metrics
- KW-Design: Pushing the Limit of Protein Design via Knowledge Refinement
- GenUnfold: Rapidly Predict Protein Mechanical Unfolding Trajectory via a Physics-Guided Diffusion Model
- Neural Atoms: Propagating Long-range Interaction in Molecular Graphs through Efficient Communication Channel
- A new benchmark for machine learning applied to powder X-ray diffraction