When a patient's genome is sequenced, most rare missense mutations — single amino-acid swaps in a protein — cannot be classified as disease-causing or harmless. Clinical labs resolve these "variants of uncertain significance" by combining four evidence streams: how common the variant is in healthy populations, computational predictions of its impact, inheritance patterns in families, and direct functional experiments (Standardizing Clinical NGS Applications—Based on ACMG Standards). The computational stream, however, is the weakest link: today's predictors are statistical black boxes trained on sequence alignments or clinical text, so they offer no mechanistic account of why a substitution disrupts a protein, and they inherit circularity from the very expert judgments they are meant to assist (From Text to Translation: Using Language Models to Prioritize Variants for Clinical Review; GPN-MSA: an alignment-based DNA language model for genome-wide variant effect prediction).
At the same time, an unrelated line of work in neural representations has solved a problem that looks, structurally, identical. Sinusoidal Representation Networks — coordinate-based networks whose every activation is a sine wave — can fit a continuous 3D field and, crucially, all of its spatial derivatives, because the derivative of a sine network is itself a sine network (Implicit Neural Representations with Periodic Activation Functions). This has been used to solve physics equations from image gradients and to enforce geometric constraints like the Eikonal condition on signed distance fields, including for protein surfaces (DeepReach: A Deep Learning Approach to High-Dimensional Reachability; DSR: Dynamical Surface Representation as Implicit Neural Networks for Protein). The proposal here is to bring that machinery to the variant-interpretation problem: represent the biophysical environment of every residue as a continuous, differentiable field, and score a mutation by how much it changes that field and its gradients.
Core Claim
A single sine-activation network $f_\theta(\mathbf{x}, \mathbf{a})$, mapping a 3D coordinate $\mathbf{x}$ near a residue together with an amino-acid embedding $\mathbf{a}$ to local biophysical field values (electrostatic potential and a signed-distance-like occupancy term), can predict the functional impact of missense mutations at least as accurately as discrete structure-based graph models in the regime where they are most needed — proteins with scarce evolutionary homologs — while using fewer parameters and producing an auditable, physics-grounded quantity that can serve as computational evidence in clinical classification. The mutation score is the integrated change in the field's gradients between the wild-type and mutant amino-acid identities:
$$ s = \int_{\|\mathbf{x} - \mathbf{x}_r\| < R} \big\| \nabla_{\mathbf{x}} f_\theta(\mathbf{x}, \mathbf{a}_{\text{mut}}) - \nabla_{\mathbf{x}} f_\theta(\mathbf{x}, \mathbf{a}_{\text{wt}}) \big\| \, d\mathbf{x}, $$
where $\mathbf{x}_r$ is the mutated residue's position, $R$ a local radius, and $\mathbf{a}_{\text{wt}}, \mathbf{a}_{\text{mut}}$ the embeddings of the original and substituted amino acid. The gradient-difference form is chosen deliberately: the biophysical impact of a substitution is governed by changes in local force and field gradients, and sine networks are, to date, the only standard architecture whose derivatives are well-behaved everywhere (Implicit Neural Representations with Periodic Activation Functions).
The gap this fills is concrete. Variant effect predictors treat protein structures as discrete graphs of residue nodes with distance-thresholded edges: SSEmb: A joint embedding of protein sequence and structure enables robust variant effect predictions masks sequence-model attention by 3D contact maps, and Accurate and Definite Mutational Effect Prediction with Lightweight Equivariant Graph Neural Networks uses equivariant graphs over residue nodes. Neither can query the field, its curvature, or its gradients at arbitrary sub-residue resolution, and neither is supervised on any physical quantity — they are trained purely on discrete fitness labels. Conversely, the two papers that do bring implicit neural representations into biology stop short of variants: Brain-wide interpolation and conditioning of gene expression in the human brain using Implicit Neural Representations fits gene-expression fields at whole-brain scales in healthy donors but never models a mutation, and DSR fits dynamic protein surfaces but uses plain MLPs, no periodic activations, and explicitly lacks any physical energy formulation or variant evaluation. No study in the file parameterizes the local continuous fields around mutated residues with periodic networks and links them to pathogenicity — this proposal does exactly that, and adds a second ingredient: a one-dimensional evolutionary-conditioning axis, adapted from the spectral "gene similarity axis" of the brain-expression work, so the field can borrow signal across related sequences the way the brain model borrows across related genes.
Research Program
Three components will be built. First, the field model itself: a SIREN with the standard scaled initialization, taking as input a local 3D coordinate (in the frame of the residue's structural neighborhood, drawn from structures in CATH as used by LGN) plus an amino-acity identity code $\mathbf{a}$, and outputting electrostatic potential and an occupancy value. The occupancy head is regularized with an Eikonal constraint $\|\nabla_{\mathbf{x}} f\| = 1$, following its use for protein-surface SDFs in DSR and for robotic mapping in HIO-SDF: Hierarchical Incremental Online Signed Distance Fields; the electrostatic head is supervised on computed potentials and, where available, on their gradients, which is exactly the derivative-only supervision regime sine networks were shown to handle (Implicit Neural Representations with Periodic Activation Functions). Second, the conditioning axis: a spectral embedding over an amino-acid similarity graph, concatenated as an extra input coordinate, extending the gene-similarity-axis construction of Brain-wide interpolation from genes to protein families; a per-protein evolutionary-depth coordinate will let one shared field model serve many proteins while exposing how much evolutionary signal it is using. Third, the readout: the gradient-difference score $s$ above, either used zero-shot or lightly calibrated against Deep Mutational Scanning assays — high-throughput experiments that measure the functional effect of thousands of substitutions at once (ProteinGym: Large-Scale Benchmarks for Protein Fitness Prediction and Design).
The parameter-efficiency claim rests on theory already in the file: From Activation to Initialization: Scaling Insights for Optimizing Neural Fields shows that smooth periodic activations reach global convergence with linear parameter scaling $O(N)$, whereas ReLU networks with Fourier positional encodings need $O(N^{1.5})$ or worse. If this transfers to molecular fields, the same accuracy should be reachable at markedly smaller model sizes — relevant because clinical deployment of per-protein field fitting otherwise inherits the slow-fitting problem seen in COIN: COmpression with Implicit Neural representations, which needed 15,000 optimization steps per image.
Baselines
The naive reference point is the current clinical default: a variant with no decisive evidence stays a VUS, which as a predictor is equivalent to random ranking — Spearman correlation 0, AUROC 0.5 — and is not a strawman, since it is what actually happens to most patients: in a 172-patient pediatric epilepsy cohort interpreted under ACMG rules, 76.7% of patients received no molecular diagnosis at all (Novel and de novo mutations in pediatric refractory epilepsy). Any model that cannot beat random ranking on held-out assays is useless; any model that only matches it leaves the clinical bottleneck untouched.
The strongest competitor in the regime this proposal targets is SSEmb, which combines a structure-masked MSA Transformer with a geometric-equivariant graph network and is specifically evaluated on proteins with shallow alignments (SSEmb). On the low-MSA subset of ProteinGym ($N_{\text{eff}}/L < 1$, i.e. fewer effective homologs than sequence length), SSEmb reports a Spearman correlation of 0.45, against 0.39 for a standard MSA Transformer ensemble — the paper's own demonstration that sequence statistics collapse under alignment scarcity and structure must carry the load. SSEmb will be run from its released setup on the identical ProteinGym split and stratification, with the same input structures, and given an equal tuning budget (the same number of random-search configurations and the same GPU-hours) as the field model; beating it means exceeding 0.45 Spearman on the low-MSA subset with non-overlapping bootstrap confidence intervals. The strongest structure-based model overall is LGN, a lightweight equivariant graph network that reports an average Spearman correlation of 0.841 across 19 DMS assays for deep mutations of up to 28 simultaneous substitutions, outperforming the supervised baseline ECNet by 4.7% and self-supervised baselines by 83.3% (Accurate and Definite Mutational Effect Prediction with Lightweight Equivariant Graph Neural Networks). LGN will be re-run under its own 19-assay protocol with the field model inserted, matched on training data and tuning budget; matching 0.841 there is a stretch goal, not the headline claim, since LGN's number is for multi-substitution effects with a few hundred fine-tuning samples per assay. Finally, the key ablation is the same field architecture with the sine activations replaced by ReLU units plus Fourier positional encodings, and separately the sine field with the derivative and Eikonal losses removed: the first isolates whether periodic activations are doing the work, the second whether derivative supervision is. All comparisons include a parameter-matched and a FLOP-matched variant, since the efficiency claim is central.
Why not just…?
Why not just use an equivariant graph neural network? This is the most serious objection, and LGN's 0.841 Spearman across 19 DMS assays makes it a real threat rather than a rhetorical one (LGN). Graph models already encode 3D geometry, are architecturally mature, and are fast. Where they fall short is on the two things this proposal actually claims: continuous physical fields and their gradients cannot be queried from a graph with contact-mask edges at all — there is no $\nabla f$ to supervise or audit — and the graph literature has not demonstrated an efficiency advantage of the kind the initialization theory predicts for periodic activations (From Activation to Initialization). The head-to-head on LGN's own 19-assay protocol and the parameter-matched sweep will settle this empirically; if the graph model wins at every model size and every MSA depth, the proposal's claim fails and should be reported as failing.
Why not just use sequence models? When many homologous sequences exist, they should be used, and they win: ProteinGym's own benchmarks found the hybrid evolutionary model TranceptEVE the overall zero-shot substitution winner, and the supervised ProteinNPT reached a Spearman correlation of 0.613 (ProteinGym). But SSEmb's low-MSA result shows the failure mode: at $N_{\text{eff}}/L < 1$ the MSA Transformer ensemble drops to 0.39 Spearman, and even with structural help only reaches 0.45 (SSEmb). Sequence-only models also cannot explain a mechanism, and the clinical need is precisely for variants without deep alignments or literature trails — 60.5% of the pathogenic variants in the epilepsy cohort were novel (Novel and de novo mutations in pediatric refractory epilepsy). The field model is scoped to the scarce-homolog regime and to interpretability, not to dethroning TranceptEVE on well-aligned proteins; the MSA-depth-stratified evaluation makes that boundary explicit rather than argued around.
Why not just use ReLU networks with Fourier positional encodings? If all one needs is a coordinate-conditioned function, standard encodings are easier to optimize. The file contains direct evidence against this shortcut in the exact setting that matters here: ReLU networks with positional encodings produce poor derivatives and failed Eikonal-type constraints, while sinusoidal networks solved the same problems, and the scaling analysis shows ReLU-PE needs $O(N^{1.5})$ parameters where sine activations need $O(N)$ (Implicit Neural Representations with Periodic Activation Functions; From Activation to Initialization). Because the entire score $s$ is built from gradients, derivative quality is not incidental — it is the mechanism. The ReLU-PE ablation under the identical training pipeline will test this directly, and if the encoded ReLU variant matches the sine variant's Spearman at equal parameters, the honest conclusion would be that the field formulation, not the periodicity, deserves the credit.
Why not just fine-tune a clinical text model? ClinVar-BERT reaches an AUROC of 0.927 by reading submitted clinical evidence text, with significant validation against MaveDB assays on BRCA1, TP53, and PTEN (From Text to Translation). But that method can only classify variants that already have clinical write-ups, and the paper itself flags template leakage and circularity — it partially re-reads expert conclusions that already incorporated computational scores. It is complementary rather than competing: the field model is designed for the novel, undocumented variants where text models have no input, and its output could eventually serve as one auditable input criterion alongside text-derived evidence rather than a replacement for it.
Proposed Experiments and Decisive Checks
The first experiment evaluates zero-shot on the ProteinGym substitution benchmark, stratified by alignment depth, with Spearman correlation as the primary metric; the decisive cell is the low-MSA stratum, where the target is to exceed SSEmb's reported 0.45 (against its 0.39 sequence-only baseline). The second experiment runs the field model under LGN's protocol across the 19 DMS assays on CATH-derived structures, reporting average Spearman for both single and deep mutations, with LGN's 0.841 for deep mutations as the reference number. The third is a parameter-scaling sweep — accuracy as a function of network width for the sine field, the ReLU-PE ablation, and the graph baseline — testing whether the predicted linear-versus-$O(N^{1.5})$ scaling separation from From Activation to Initialization appears in the molecular setting. The fourth is clinical: score variants in ClinVar and check monotonic separation of the five ACMG tiers by $s$, then score the novel pathogenic and de novo variants documented in the epilepsy cohort and the MYT1L case report (Novel and de novo mutations in pediatric refractory epilepsy; De novo missense mutation in MYT1l leading to autosomal dominant intellectual disability 39 and autism spectrum disorder: a case report) — variants with no prior literature, where text-based methods are structurally inapplicable. Throughout, fitting-time per protein will be reported, since thousands-of-steps overfitting is a known cost of this architecture family (COIN) and a clinical tool must fit a protein in minutes, not hours. The claim is falsified if, at matched parameters and tuning budget, the graph baseline matches the field model on the low-MSA stratum and the ReLU-PE ablation matches the sine field's gradient quality — and the write-up will say so plainly if that is what the numbers show.
References
- Standardizing Clinical NGS Applications—Based on ACMG Standards
- From Text to Translation: Using Language Models to Prioritize Variants for Clinical Review
- GPN-MSA: an alignment-based DNA language model for genome-wide variant effect prediction
- Implicit Neural Representations with Periodic Activation Functions
- DeepReach: A Deep Learning Approach to High-Dimensional Reachability
- DSR: Dynamical Surface Representation as Implicit Neural Networks for Protein
- SSEmb: A joint embedding of protein sequence and structure enables robust variant effect predictions
- Accurate and Definite Mutational Effect Prediction with Lightweight Equivariant Graph Neural Networks
- Brain-wide interpolation and conditioning of gene expression in the human brain using Implicit Neural Representations
- HIO-SDF: Hierarchical Incremental Online Signed Distance Fields
- ProteinGym: Large-Scale Benchmarks for Protein Fitness Prediction and Design
- From Activation to Initialization: Scaling Insights for Optimizing Neural Fields
- COIN: COmpression with Implicit Neural representations
- Novel and de novo mutations in pediatric refractory epilepsy
- De novo missense mutation in MYT1l leading to autosomal dominant intellectual disability 39 and autism spectrum disorder: a case report