Microwave filter designers have a precise language for how a network stops a signal. A "transmission zero" is a frequency at which the output of a two-port circuit is exactly zero. It arises in two ways. A local resonant branch can shunt the energy away, or two paths can carry the signal and cancel each other out of phase. Proteins are also networks that carry signals. A ligand binding at one site shifts atoms at a distant active site, a phenomenon called allostery, and many disease mutations damage this long-range communication without unfolding the protein. We propose to model a protein as a mass–spring network and to compute its frequency-dependent transfer function between functional sites. We then score a missense variant by how far it moves the zeros of that transfer function toward, or into, the low-frequency band where functional motion lives. That score becomes a covariance kernel inside a Gaussian-process variant-effect predictor.
The opportunity is concrete. TriFit showed that residue-level dynamics from an elastic network add real signal on top of sequence and structure: +2.2 AUROC points, against +0.2 for adding structure to sequence. But it feeds those dynamics into a neural network as raw vectors: fluctuation amplitudes, mode shapes and correlation matrices. Nobody has asked the question a filter engineer would ask, namely whether a substitution places a zero in the functional band between two specific sites. If that question carries information that static coupling measures do not, the gain should be largest where current supervised predictors fail most. That is at sequence positions never seen in training.
Core Claim
Our central hypothesis is specific and falsifiable. After controlling for sequence embeddings, the static structural environment, and static (zero-frequency) coupling between a mutated residue and functional sites, the frequency-resolved zero structure of the inter-site transfer function still predicts deep-mutational-scan fitness and clinical pathogenicity. The effect should be strongest on ProteinGym's position-held-out splits. The note is organized around the ablation that could kill this claim. If a plain static perturbation-response kernel matches the frequency-resolved one, the transmission-zero framing adds nothing, and we will report that.
The Model
We build an anisotropic elastic network from each AlphaFold2 structure. Residues are point masses and nearby residues are joined by springs. This gives a stiffness matrix $K$, a diagonal mass matrix $M$, and a damping matrix $C$. For a stimulus site $a$ and an active site $b$, the transfer function is
$$G_{ab}(s) = \left[(s^2 M + s C + K)^{-1}\right]_{ab},$$
where $s$ is complex frequency. Its zeros are the values of $s$ at which a force at $a$ produces no motion at $b$. Filter theory gives two ways to produce zeros here. The first is local trapping: a residue whose own resonance absorbs the drive, the analogue of the quarter-wave stub in small-reflection stub filter design. The second is multipath cancellation: two contact pathways from $a$ to $b$ arriving out of phase, the analogue of the input–output cross-coupling that creates tracking zeros in tunable microstrip filters and of the mixed electric/magnetic coupling in hybrid SIW–CPW filters. We define the "passband" as the frequency range spanned by the lowest normal modes, which carry the large-scale functional motion.
A substitution at residue $r$ is modeled as a local, low-rank change to $K$ and $M$. The springs incident on $r$ get new stiffness and $r$ gets a new effective mass. These are parameterized by a small table indexed by wild-type amino acid, mutant amino acid and burial, and fit on training folds only. Because the change is low-rank, the Woodbury identity lets us update $G_{ab}$ for every variant without refactoring the full matrix. That makes all ProteinGym substitutions tractable.
From the perturbed transfer function we extract three features for each variant. The first is the shift of the nearest zero relative to the passband edge, in units of passband width. The second is the change in integrated in-band transmission. The third is a count of zeros that cross into the passband. The first feature borrows a constraint from dual-DRAG spectral notch pulses. That work found that a notch placed within about 20 MHz of a 25-ns carrier eats into the carrier and degrades fidelity. Our analogue predicts that zeros moving close to the passband, not only into it, are damaging.
Protein motion is heavily damped, so zeros sit off the real frequency axis. We therefore measure distances in the complex plane rather than assuming sharp real notches.
Functional sites $a$ and $b$ come from ligand-contacting residues in the structures where these exist. As a sensitivity analysis, we also use sites chosen by zero-shot sequence-model conservation, so that the method does not depend on hand annotation.
Two prior works motivate this setup without supplying it. Architecture and coevolution of allosteric materials evolved spring networks whose response is carried along a narrow channel, drops to zero in the bulk, and is amplified at the active site. That is evidence that null transmission is a real organizing feature of allosteric networks, although the study used idealized networks and no real variants. AlloPool prunes molecular-dynamics interaction graphs down to minimal signaling pathways. We use those pathways as a qualitative check that our multipath zeros coincide with the channels it finds. AlloPool itself neither scores variants nor treats pathways as frequency-dependent channels.
Research Program
The predictor is built on Kermut. Kermut is Gaussian-process regression with an ESM-2 sequence kernel, a ProteinMPNN structural kernel and an ESM-2 zero-shot mean. We add a third kernel, which measures the similarity between two variants' zero-shift signatures, with an automatically weighted amplitude.
We chose this host over a neural fusion model for three reasons. The Gaussian process learns an interpretable weight for the new kernel. It trains in minutes, so every ablation can use the same tuning budget. And it keeps calibrated uncertainty.
The program has three stages. The first is a synthetic validation. We regenerate evolved spring networks following the allosteric-materials procedure and check that zero placement predicts task failure under random bond perturbations. This establishes that the features measure what they claim before any protein data is used. The second stage fits the substitution-to-spring table and runs the full ProteinGym evaluation. The third is a zero-shot variant for clinical data. Here the transmission score is combined with a sequence-model log-likelihood ratio and evaluated on the ClinVar subset of ProteinGym, where no assay labels exist.
Baselines
The naive reference point is zero-shot ESM-style scoring. On the binarized ProteinGym benchmark, TriFit reports ESM3 at AUROC 0.769. We rerun ESM-2 zero-shot on our exact splits so that it anchors the scale.
The strongest direct competitor is Kermut. It is the host model, and it isolates exactly what our kernel adds. Kermut reports an average Spearman correlation of 0.662 over 217 substitution assays, above ProteinNPT's 0.613. TriFit reports Kermut at AUROC 0.864 on its binarized benchmark.
The strongest existing dynamics-aware method is TriFit, which reports AUROC 0.897 (top/bottom 30% binarization, 217 assays). We will reproduce TriFit's binarization so the two numbers are comparable, and we will report Spearman correlation separately, because the two metrics are not interchangeable. ProteinNPT is included for completeness, using its reported numbers, because its reported cross-validation cost of over 400 hours rules out rerunning it under our budget.
The key ablations are three. The first replaces the frequency-resolved kernel with a static perturbation-response kernel, i.e. $G_{ab}$ evaluated at $s=0$. The second keeps the frequency dependence but uses only integrated in-band transmission, with no zero-location features. The third substitutes TriFit-style fluctuation-amplitude and correlation features into the same Kermut slot.
All variants use the same AlphaFold2 structures, the same ESM-2 650M embeddings, the same folds, and the same hyperparameter search budget. The substitution-to-spring table is always fit within training folds.
Beating Kermut means a paired, per-assay Spearman improvement that is significant under a sign test across assays. Beating TriFit means exceeding 0.897 AUROC under its binarization. We regard the second as a stretch goal. The decisive result for the claim is the gap over the static ablation on the Modulo and Contiguous splits.
Why not just…?
Why not just use a large protein language model? Sequence models already absorb much structure. TriFit's own ablation, though, shows dynamics adding 2.2 AUROC points beyond sequence plus structure, which argues that part of the signal is not captured by sequence alone. Our Kermut host already contains the ESM-2 kernel and mean, so any gain we report is by construction over and above the language model.
Why not just feed elastic-network features into a neural network, as TriFit does? This is a genuine competitor. TriFit's fusion model may learn whatever our zero features encode. The third ablation tests this directly. If TriFit-style features in the same kernel slot match ours, the zero framing is redundant. A further advantage would remain, because our feature is tied to specific site pairs and is interpretable, but we would not claim a predictive one.
Why not just use static coupling between the mutated residue and the active site? This is the most serious threat. Perturbation-response scanning is cheap and well motivated, and it is the zero-frequency limit of our transfer function. Heavy damping may flatten the frequency structure until only this static term matters. The first ablation settles the question. We commit in advance to treating parity with the static kernel as a negative result for the transmission-zero hypothesis.
Why not just rely on alignment-based conservation such as GEMME? Conservation is strong where alignments are deep. ProteinGym documents that these models degrade when alignments are shallow, and the Lacuna file reports no GEMME number we can cite. We will therefore stratify all results by alignment depth. Our hypothesis predicts the largest gains on proteins with shallow alignments and a known allosteric function.
Decisive Checks
The first check is synthetic: zero placement must predict functional failure in evolved spring networks, or the pipeline is not measuring transmission at all. The second, which is the main test, is the frequency-resolved versus static ablation on the Modulo and Contiguous splits of the 217 ProteinGym substitution assays. We report per-assay Spearman correlation and binarized AUROC, stratified by whether the protein has annotated allosteric or ligand sites. The third is the clinical transfer: whether adding the transmission score to zero-shot sequence scoring improves AUROC on ProteinGym's ClinVar set, with calibration error reported alongside TriFit's stated 0.044.
A positive outcome would be a consistent gain over the static ablation, concentrated in allosteric proteins and at held-out positions. That would give variant predictors a physically interpretable handle — "this substitution notches out the signal between site $a$ and site $b$." A null outcome would be informative too, because it would show that damping reduces protein signaling to its static limit.
References
- TriFit: Trimodal Fusion with Protein Dynamics for Mutation Fitness Prediction
- A Method Based on the Theory of Small Reflections to Design Arbitrary Passband Microstrip Filters
- A Compact Tunable Microstrip Bandpass Filter with Tuning Range and Bandwidth Enhanacement
- Compact filter based on a hybrid structure of substrate integrated waveguide and coplanar waveguide
- Suppressing Spurious Transitions Using Spectrally Balanced Pulse
- Architecture and coevolution of allosteric materials
- AlloPool: An Adaptive Graph Neural Network for Dynamic Allosteric Network Prediction in Protein Systems
- Kermut: Composite kernel regression for protein variant effects
- ProteinNPT: Improving Protein Property Prediction and Design with Non-Parametric Transformers
- ProteinGym: Large-Scale Benchmarks for Protein Fitness Prediction and Design