Does Fast Molecular Dynamics Generalize Across Human Genetic Diversity?

Molecular dynamics (MD) simulation tracks every atom of a protein over time, and it is slow for a brutally simple reason: vibrations of bonded hydrogens force the simulator to take steps of about a femtosecond, while the biologically important motions — domain rearrangements, allosteric shifts, cryptic pocket opening — take microseconds to seconds. A new family of generative models sidesteps this by learning to jump directly between conformational states: DeepJump predicts transitions conditioned on arbitrary time leaps with roughly 1000x less compute than force-field simulation Accelerating Protein Molecular Dynamics Simulation with DeepJump, ProTDyn's 100 ns predictions match Markov state models built from 50 microseconds of continuous MD ProTDyn: A Foundation Protein Language Model for Thermodynamics and Dynamics Generation, and AlphaFLOW-MD reproduces MD-derived per-residue fluctuations with a median Pearson correlation of $r = 0.85$ on the ATLAS benchmark AlphaFold Meets Flow Matching for Generating Protein Ensembles.

Meanwhile, a parallel fairness problem has hardened in human genetics. Causal-inference instruments, clinical thresholds, and reference sequences are overwhelmingly calibrated on European-ancestry cohorts, and transferring them to other ancestries routinely fails: tagging variants lose linkage, instruments drop below the weak-instrument threshold, and patients face diagnostic delay measured in years Precision diagnosis for monogenic diabetes requires ethnicity specific criteria for genetic testing. These two problems share an unexamined root. Fast dynamics models are trained on canonical reference proteins from databases like ATLAS and mdCATH — sequences drawn from the same European-skewed reference world — yet the variants where their predictions would matter most are precisely the population-stratified missense alleles that differ across human ancestries. No published result found on Lacuna evaluates any accelerated dynamics model on ancestry-stratified human variants. This proposal makes that measurement, and then uses what it finds.

Core Claim

The claim has two parts, one diagnostic and one constructive. Diagnostically, I claim that accelerated generative dynamics models carry an ancestry-shaped generalization gap: they predict conformational ensembles less faithfully for missense variants that are common in African, South Asian, or admixed ancestries and rare in Europeans, because such variants are underrepresented in training data and disproportionately alter flexible or disordered regions that static-structure-centric pipelines never saw. Constructively, I claim this gap can be substantially closed by physics-feedback alignment — reweighting the generative model's outputs toward the Boltzmann distribution using force-field energies, the mechanism that lifted RMSF correlation from $r \approx 0.50$ to $r \approx 0.71$ on ATLAS in Aligning Protein Conformation Ensemble Generation with Physical Feedback — because the force field depends on sequence and structure, not on which population database supplied the training trajectory. The payoff is a per-variant conformational disruption score that genetic epidemiology currently lacks: drug-target Mendelian randomization (MR) treats a protein as a scalar abundance dial turned by expression quantitative trait loci, as in the metformin-target analysis that found PRKAG1 protective against colorectal cancer (odds ratio 0.74, 95% CI 0.63–0.87, adjusted $p = 0.001$) using exclusively European-predominant data Evaluating the Impact of Putative Metformin Targets on Cancer Outcomes: A Drug‐Target Mendelian Randomization Study. A dynamics-derived score gives those instruments a physical axis of validity.

Research Program

The program has three workstreams. The first builds the missing benchmark: an ancestry-stratitated panel of missense variants in proteins that already have dynamics ground truth or are MR-relevant. For each protein I will assemble variant sets stratified by ancestry — alleles common in one continental group and rare or absent in others — alongside frequency-matched European-common variants as controls, so that any performance difference reflects variant character rather than protein identity or allele frequency. The natural sources are the 54-gene next-generation-sequencing panels with their variants of uncertain significance from the MY DIABETES multi-ethnic cohort of 1,706 participants (653 White, 639 South Asian, 275 African-Caribbean, 139 Mixed) Precision diagnosis for monogenic diabetes requires ethnicity specific criteria for genetic testing, and the cis-eQTL instrument regions in GTEx v10 used by the metformin MR study Evaluating the Impact of Putative Metformin Targets on Cancer Outcomes. I am explicit that the file indexes no large ancestry-stratified allele-frequency catalog; assembling that variant panel from public sequencing resources is itself a deliverable of the benchmark, and the ATLAS proteins (1,390 monomeric targets with standardized 100 ns trajectories at 1 ns frame spacing) provide native-sequence anchors throughout.

The second workstream measures and repairs the gap. I will evaluate existing surrogates — DeepJump, ProTDyn, AlphaFLOW-MD, and the discrete-token ESMDiff, whose public codebase is available as ESMDiff Codebase — by editing the variant amino acid into the input sequence and comparing predicted ensembles against newly run classical MD replicates at ATLAS scale (100 ns per variant, a few hundred variants, which is feasible precisely because it is not millisecond simulation). The repair step is the key ingredient: Boltzmann alignment in the style of EBA, which treats ensemble generation as a preference-optimization problem pushing the model toward low-energy structures scored by the CHARMM36m force field, applied here per-variant. I expect alignment to matter most out of distribution, where the model's learned priors are least reliable — the same reason DeepJump's own authors report that large 100 ns jumps degrade folding accuracy relative to 1 ns and 10 ns jumps, and that the model is biased toward globular conformations DeepJump.

The third workstream carries the score into causal inference. For each missense variant $s$, define the conformational disruption score as the divergence between the model's ensemble for the mutated sequence and for the reference sequence:

$$D(s) = \mathrm{JSD}\big(q_{\mathrm{ref}},\; q_s\big),$$

where $q_s$ is the predicted distribution over structural descriptors (per-residue fluctuation, radius of gyration, pocket-region geometry) given the mutated sequence, and JSD is the Jensen-Shannon divergence, chosen because it is bounded and symmetric. The MR experiment then asks whether $D(s)$ separates instruments that transfer across ancestries from those that do, and whether filtering on $D$ complements purely statistical pleiotropy detection such as HEIDI-outlier, which removed 16 pleiotropic SNPs in the APOE region and made an apparent LDL–Alzheimer causal link disappear in Causal associations between risk factors and common diseases inferred from GWAS summary data but, as its authors note, remains bounded by European GWAS data.

Proposed Experiments

Experiment 1 is the fidelity measurement. For roughly 200–300 variants across 15–25 proteins, I will run matched classical MD (ATLAS protocol) and compare surrogates on RMSF correlation, ensemble-distribution overlap in radius of gyration and RMSD (the zero-shot evaluation ProTDyn used on unseen proteins ProTDyn paper summary), and recovery of metastable states (the BPTI Cluster-3-style test where ESMDiff succeeded Structure Language Models for Protein Conformation Generation). The primary endpoint is the ancestry gap: the difference in RMSF correlation between European-common and non-European-common variant strata. Experiment 2 is the repair: the same backbone with and without Boltzmann alignment, reporting how much of the gap closes and at what compute cost. Experiment 3 is clinical: rank the MY DIABETES variants of uncertain significance by $D(s)$ and test whether high-disruption VUS enrich for later molecular confirmation in the Exeter registry. Experiment 4 is the MR integration: re-run the metformin drug-target design across ancestry-stratified instrument sets, testing whether instruments with high-$D$ coding variants show divergent cross-population causal estimates, and whether a $D$-based filter changes the instrument set relative to HEIDI-outlier alone.

Baselines

The naive reference point is current practice: treat the drug target as scalar expression, using cis-eQTL instruments with $F$-statistics above 10, exactly as the metformin study did for 11 targets in GTEx v10 and UK Biobank Evaluating the Impact of Putative Metformin Targets on Cancer Outcomes. For variant interpretation, the naive baseline is single-structure prediction with standard self-consistency, which Rethinking Self-Consistency in Protein Generative Models showed rejects 44% of real native proteins, rescuing recall from 74% to 82% (and 50% to 65% in the most flexible bin) only by switching to ensemble scoring — so the ensemble baseline is included too. The strongest existing competitor is AlphaFLOW-MD, which reports a median RMSF correlation of $r = 0.85$ against MD on ATLAS with a 10x wall-clock speedup AlphaFold Meets Flow Matching; I will run it mutation-conditioned (sequence edited, same variant set, same per-variant compute, and the same tuning budget allocated on European-common variants so no method sees non-European variants during development). DeepJump and ProTDyn run under the same protocol, with jump size fixed per their own reported best practice (1–10 ns leaps). The ablation isolating my key ingredient is identical with and without Boltzmann alignment; EBA's published move from $r \approx 0.50$ to $r \approx 0.71$ on native ATLAS sequences Aligning Protein Conformation Ensemble Generation with Physical Feedback sets the expected effect size on in-distribution data, and the open question is whether the effect grows on out-of-distribution variants. Beating the baselines means two things at once: matching the best baseline's RMSF correlation on European-common variants, and cutting its ancestry gap by at least half. Classical MD is not a competitor but the referee, since even a 100-microsecond run failed to fully sample cryptic pockets of SARS-CoV-2 main protease Matteo Carli et al., making proteome-scale classical MD for thousands of variants arithmetically impossible.

Why not just…?

Why not just use static AlphaFold2 or ESMFold to score variant effects? This is the strongest simple alternative, and static predictors will catch gross fold-destabilizing variants. But the evidence in the file says they fail exactly where this proposal lives: single-structure metrics reject 44% of genuinely flexible native proteins Rethinking Self-Consistency, static predictors regress to one answer and cannot model ensembles from a single sequence AlphaFLOW, and adding explicit MD representations to a protein foundation model improved thermostability ($p < 0.01$), protein-interaction ($p < 0.03$), and ion-binding ($p < 0.006$) prediction over static structure tokens alone When Protein Dynamics Matter. To be fair, Experiment 1 includes a static-structure variant-effect baseline so the ensemble advantage is measured, not asserted; if static scores match ensemble scores on the benchmark, that is a decisive negative result worth publishing.

Why not just fix ancestry bias statistically, with harmonized multi-ancestry GWAS and filters like HEIDI-outlier? Necessary but insufficient. Statistical filters detect heterogeneity after the fact — HEIDI removed the 16 APOE-region SNPs GSMR — but cannot rescue an instrument whose tagging SNP has lost linkage in another ancestry, and cannot say why a variant is pleiotropic. The monogenic diabetes study shows what pure statistical recalibration on homogeneous data costs in practice: a ten-year diagnostic delay for non-White patients MY DIABETES. My filter is designed to be complementary and is tested as an addition to, not a replacement for, statistical harmonization in Experiment 4.

Why not just run classical MD on the variants of interest? Because the arithmetic does not close. Slow conformational events need microsecond-to-millisecond sampling; 200 ns missed SH3 folding transitions Xavier Périole et al., 1.5 microseconds missed rare p53 states Marco Gaetano Lepre et al., and 100 microseconds still missed cryptic pockets Carli et al.. Classical MD therefore appears in this proposal only as ground truth at ATLAS scale, and the single-case-study precedent — docking plus slow MD on ALK missense mutations ALK missense MD study — is exactly what does not scale to population-level variant panels.

Why trust generative surrogates at all, given their documented biases? This is a real threat, not a straw man: DeepJump fails on small and non-globular proteins, ESMDiff's resolution is capped by its discrete tokenizer, and EBA is single-chain only. That is why Experiment 1 measures out-of-distribution fidelity directly against MD before anything downstream is attempted, and why the key ingredient is precisely the mechanism (force-field feedback) that corrects a generative model's drift when its learned priors fail. If the ancestry gap survives alignment, the honest conclusion — that current accelerated dynamics is unsafe for population genetics — is itself the paper.

Decisive Checks

The proposal lives or dies on three numbers. First, the measured ancestry gap in RMSF correlation for the best unaligned surrogate on held-out MD ground truth — if it is near zero, the diagnostic claim is false and the constructive workstream collapses to an ensemble-variant benchmark. Second, the fraction of that gap closed by Boltzmann alignment at matched compute, judged against EBA's published 0.50-to-0.71 in-distribution effect. Third, whether $D(s)$ adds anything demonstrable in Experiment 4: enrichment of confirmed pathogenic VUS over unconfirmed, and cross-ancestry instrument behavior that HEIDI-outlier-style filtering alone cannot explain. Each check is falsifiable on data already identified in the file, and each result — positive or negative — answers a question no published work has yet touched.