When a genome-wide study flags a genetic variant that raises risk for two diseases at once, the finding is statistical: the variant tracks with both diagnoses in a cohort. What almost never follows is a physical measurement of what that variant does to the protein it encodes — whether it bends the backbone, destabilizes the fold, or changes how the protein behaves in a membrane. The tool that could answer this, circular dichroism (CD) spectroscopy — an optical technique that reads out a protein's secondary-structure mix (helix, sheet, disordered coil) and its thermal stability in solution — is cheap and direct, but it is low-throughput and notoriously sensitive to buffer, pH, temperature, and lipid content. At the same time, the curated repositories that aggregate thousands of experimental stability measurements blend results from CD, calorimetry, and fluorescence assays across wildly different conditions, and the computational models trained on them throw all of that variation away as noise.
My claim is that this discarded variation is not noise but structure, and that the right statistical tool for it is the hierarchical linear model — a regression that partitions effects across nested tiers (measurements within experimental conditions within proteins), partially pooling sparse groups toward the population mean so that under-sampled proteins borrow strength from well-sampled ones. Doing this yields two payoffs. First, it calibrates and improves existing computational stability predictors, which currently pool everything. Second, it produces a standardized, condition-aware biophysical effect size per variant — exactly the quantity needed to test, with CD experiments, whether proteins encoded at pleiotropic risk loci share a measurable folding phenotype. Shared genetic architecture then stops being a purely statistical observation and becomes a testable claim about protein physics.
Core Claim
In pooled protein-stability data, a substantial fraction of the disagreement between computational predictions and experimental measurements is attributable to unmodeled experimental context — assay type, buffer, temperature, lipid composition — rather than to the variants themselves. A three-level mixed model that explicitly partitions this variance will (a) improve held-out prediction of mutational stability effects on heterogeneous benchmarks, (b) expose where current zero-shot models are condition-blind, and (c) supply the calibrated per-variant biophysical estimates needed to correlate genetic effect sizes at pleiotropic loci with protein-folding consequences, validated by a targeted CD titration experiment.
This is grounded in a documented gap. Zero-shot protein stability prediction by inverse folding models: a free energy interpretation shows that inverse folding models — sequence-to-sequence networks whose likelihoods proxy for folding free energy — reach only a Pearson correlation of 0.66 on the Protein G benchmark (0.72 after corrections using molecular dynamics ensembles and disordered-protein statistics), and it explicitly notes that experimental conditions are treated as unmodeled latent variables. Meanwhile FireProtDB: database of manually curated protein stability data compiles 13,274 measurements across 237 proteins from heterogeneous assays and labs, with no hierarchical structure imposed. And on the genetics side, Detecting local genetic correlations with scan statistics (LOGODetect) identified 227 pleiotropic genomic segments across seven neuropsychiatric traits, replicating 17 of 24 discovery regions in an independent cohort — yet no pipeline connects such loci to the biophysics of the encoded proteins. The mixed-modeling machinery itself is proven in adjacent domains: Mixed Effects Neural Networks (MeNets) With Applications to Gaze Estimation cut gaze error from 6.04° to 4.90° by separating population-wide from individual-specific effects, and One model to rule them all: unification of voltage-gated potassium channel models via deep non-linear mixed effects modelling unified 87,640 patch-clamp sweeps across 20 channel types with fixed plus random effects. None of these touches spectroscopy, sequence structure, or genetics — that three-way opening is what this proposal takes.
Research Program
The program has three linked components. The first is a re-analysis of FireProtDB with an explicit measurement-error hierarchy. For a stability measurement $y$ (a stability change $\Delta\Delta G$ or a melting-temperature shift $\Delta T_m$) of variant $v$ in protein $p$ under condition $c$, I fit
$$y_{vpc} = \beta_0 + \beta_1 \hat{y}^{\mathrm{IF}}_{v} + u_p + u_{c(p)} + u^{\mathrm{slope}}_{p}\,\hat{y}^{\mathrm{IF}}_{v} + \epsilon_{vpc},$$
where $\hat{y}^{\mathrm{IF}}_{v}$ is a zero-shot inverse folding prediction treated as a fixed-effect covariate, $u_p$ is a protein-level random intercept, $u_{c(p)}$ is a condition-level random intercept (assay, pH, buffer, temperature) nested within protein, $u^{\mathrm{slope}}_{p}$ allows the calibration of the computational score to vary by protein, and $\epsilon$ is per-measurement error, all random terms Gaussian. The immediate deliverable is a variance decomposition: what share of total variance sits at the condition tier versus the protein tier versus the residual? No published analysis of these repositories reports this, and the intraclass correlations it yields are the empirical justification (or refutation) for the whole program.
The second component embeds the hierarchy as a calibration layer for zero-shot predictors. The model above already does this for point predictions; I extend it to condition-conditional prediction by adding assay-specific random slopes, so the system can answer "what is the predicted CD-measured effect under near-physiological pH" versus "under thermal-scan conditions" — questions that Improving Inverse Folding models at Protein Stability Prediction without additional Training or Data, which achieves 77% Success@10 on the Tsuboyama dataset with ProteinMPNN-DDG (up from 66%), cannot address because its input is sequence and structure only.
The third component is the genetics bridge. I map the pleiotropic hub genes from LOGODetect's 227 regions, together with the lipid-metabolism loci implicated in neurodegeneration (Circulating Cholesterol Levels May Link to the Factors Influencing Parkinson's Risk; ABCA7 variants impact phosphatidylcholine and mitochondria in neurons), onto proteins with stability data in FireProtDB, and regress hierarchical-model biophysical effect sizes against GWAS effect sizes at shared loci. I am deliberate about the limits: coverage will be partial (only some pleiotropic genes have stability measurements), and the association is observational, subject to linkage-disequilibrium confounding in the same way Gastroesophageal reflux GWAS identifies risk loci that also associate with subsequent severe esophageal diseases is. The correlation study therefore serves mainly to prioritize targets, not to prove causation.
Proposed Experiments
The computational experiments run on FireProtDB (13,274 entries, 237 proteins, 43% destabilizing and 14% stabilizing mutations), the S2648 and S669 stability sets, and — deliberately as a negative control — the Protein G and Tsuboyama datasets, which are single-protein and near-uniform in condition, so the hierarchy has almost nothing to pool over there. Evaluation is held-out-variant Pearson correlation and RMSE for $\Delta\Delta G$, a Success@10 ranking metric to compare against the reported Tsuboyama numbers, and a new condition-stratified evaluation: predict measurements in a condition entirely held out from training. The variance-partitioning analysis of FireProtDB is itself a reported result, with intraclass correlations per tier.
The wet-lab experiment tests the most concrete mechanistic hypothesis in the file's Gap 4: the lipid axis. A Role of Cholesterol in Modulating the Binding of α-Synuclein to Synaptic-Like Vesicles showed by CD titration that vesicle cholesterol suppresses α-synuclein's helical folding, and Modulation of the Interactions Between α-Synuclein and Lipid Membranes by Post-translational Modifications reviewed the same folding transition. I will measure far-UV CD spectra and thermal melts for a panel of variants from lipid-transport pleiotropic loci (ABCA7 and cholesterol-pathway genes) across a graded vesicle-cholesterol composition, fitting the folding transition with cholesterol content as a random slope within variant. The decisive readout is a variant-by-cholesterol interaction: do risk-variant proteins lose (or gain) folding sensitivity to membrane composition relative to wild type? The Protein Circular Dichroism Data Bank is named in the literature as a repository of raw CD spectra with metadata, but no reported result was found on Lacuna for it, so spectra from this experiment will be released with full condition metadata as a small benchmark in the form the field currently lacks.
Baselines
The naive reference point is complete pooling: ordinary least squares on all 13,274 FireProtDB measurements at once, treating every observation as independent — the implicit practice whenever repository data is regressed or averaged without structure, and the practice that inflates effective sample size and understates standard errors, as the multilevel literature (The Dynamics of (Not) Unfollowing Misinformation Spreaders) documents. The second baseline is no pooling: a separate regression per protein, which will be reported alongside to show where sparse proteins (most of the 237 have few measurements) produce volatile estimates.
The strongest existing competitor is the zero-shot inverse folding pipeline itself, used exactly as published: Zero-shot protein stability prediction by inverse folding models: a free energy interpretation reports a Protein G Pearson correlation of 0.66 for standard practice and 0.72 with molecular-dynamics ensembles plus disordered-protein statistics, and Improving Inverse Folding models at Protein Stability Prediction without additional Training or Data reports 77% Success@10 on Tsuboyama for ProteinMPNN-DDG. These run at essentially zero fitting cost, so the fair comparison gives my hierarchy a modest, pre-registered tuning budget (random-effects structure selected by information criteria, nothing else), identical held-out splits, and the same raw predictions as input. Beating the baseline means higher held-out correlation on the heterogeneous sets (FireProtDB, S2648) and, more importantly, any nonzero accuracy in the condition-stratified evaluation, where the zero-shot baseline has no input signal at all and will perform at chance. The key ablation strips the hierarchy down to fixed effects plus a single global offset — same covariates, no random intercepts or slopes — which isolates whether the gains come from partial pooling itself rather than from merely having the computational score as a regressor. A second ablation removes only the assay tier to test whether condition structure, not just protein structure, carries the improvement.
Why not just…?
Why not just use the inverse folding models zero-shot, without experiments or hierarchy? This is the real threat, and I will not argue it away: on single-protein, uniform-condition benchmarks like Protein G and Tsuboyama, the zero-shot models are excellent and my hierarchy has little to add — I expect a tie there, and I say so in the design by naming those datasets negative controls. The evidence for where they fall short is already published: even the corrected version explains only about half the variance (r = 0.72 on Protein G, from the free-energy interpretation paper), and both that paper and the tied-decoding paper note the models cannot represent pH, temperature, ionic strength, or lipid presence — precisely the variables CD measures directly. The condition-stratified evaluation is the experiment that settles it: if the hierarchy adds nothing even there, the proposal's first component fails and I will report that.
Why not just run ordinary regression per protein, or one big pooled regression? Both fail for documented reasons. Complete pooling treats nested measurements as independent and understates uncertainty; no pooling overfits the sparse proteins that dominate FireProtDB's long tail. The mixed-model alternative has direct performance evidence in neighboring domains: MeNets improved gaze estimation by 10–20% across benchmarks precisely by adding a mixed-effects layer to a standard network, and the potassium-channel NLME model beat seven separately tuned Hodgkin-Huxley models across 87,640 sweeps. The per-protein-regression baseline will be run, so the comparison is empirical rather than rhetorical.
Why not just do standard single-disease GWAS or a phenome-wide scan, skipping the pleiotropy machinery and the biophysics? Because the genetic side of this proposal deliberately reuses existing pleiotropy tools rather than replacing them — LOGODetect and the multi-organ analysis in Brain–heart–eye axis revealed by multi-organ imaging genetics and proteomics already showed that local, locus-resolved correlation finds hub regions (227 segments, with 17 of 24 replicated) that global single-trait or genome-wide-correlation analyses dilute away. My contribution is not better pleiotropy detection; it is what happens downstream of it. A reviewer who says "just use GWAS" is agreeing with the proposal's inputs and objecting to its outputs, which is the mechanistic SNP-to-protein gap the GWAS literature itself leaves open.
Why not just run molecular dynamics on the variants instead of CD experiments? The closest precedent, Computational compensatory mutation discovery approach: Predicting a PARP1 variant rescue mutation, predicted a PARP1 rescue mutation purely in silico and its authors still note the need for experimental structural confirmation; molecular dynamics cannot cheaply scan a cholesterol-composition gradient, which is a solution condition, not a coordinate set. The CD titration is the only component here that measures the membrane-composition dependence directly.
Decisive Checks
Three results would make or break this. First, the FireProtDB variance decomposition: if the condition tier explains only a few percent of variance, the premise fails and the calibration component is moot. Second, the condition-stratified prediction gap: the hierarchy must beat zero-shot where conditions vary and tie with it where they do not; anything less means partial pooling is not the operative ingredient. Third, the CD titration: a reproducible variant-by-cholesterol interaction in proteins from lipid-pleiotropic loci, with uncertainty that survives the mixed-model standard errors, would be the first mechanistic foothold connecting shared genetic architecture to a measured folding phenotype — and its absence would sharpen the field's picture of where the SNP-to-biophysics gap actually resists closure.
References
- Zero-shot protein stability prediction by inverse folding models: a free energy interpretation
- FireProtDB: database of manually curated protein stability data
- Detecting local genetic correlations with scan statistics
- Mixed Effects Neural Networks (MeNets) With Applications to Gaze Estimation
- One model to rule them all: unification of voltage-gated potassium channel models via deep non-linear mixed effects modelling
- Improving Inverse Folding models at Protein Stability Prediction without additional Training or Data
- Circulating Cholesterol Levels May Link to the Factors Influencing Parkinson’s Risk
- ABCA7 variants impact phosphatidylcholine and mitochondria in neurons
- Gastroesophageal reflux GWAS identifies risk loci that also associate with subsequent severe esophageal diseases
- A Role of Cholesterol in Modulating the Binding of α-Synuclein to Synaptic-Like Vesicles
- Modulation of the Interactions Between α-Synuclein and Lipid Membranes by Post-translational Modifications
- The Dynamics of (Not) Unfollowing Misinformation Spreaders
- Brain–heart–eye axis revealed by multi-organ imaging genetics and proteomics
- Computational compensatory mutation discovery approach: Predicting a PARP1 variant rescue mutation