Physics-Inferred Mechanism Priors for Unsolved Congenital Hydrocephalus

More than half of patients with a suspected inherited disorder remain undiagnosed after exome sequencing (Beyond the exome). Part of the reason is that software ranking the thousands of variants in a patient's genome learns about the patient only through coarse clinical labels such as "hydrocephalus." Two children can carry that label for different physical reasons. In one, the brain's surface cilia fail to push cerebrospinal fluid (CSF). In another, a narrow channel is blocked. In a third, the fluid is not drained away. These reasons point to different gene families. We propose to recover that mechanistic information from each patient's own brain images. We do this by fitting a fluid simulation to the patient's ventricles and CSF flow measurements. The few fitted physical parameters, such as wall propulsion strength, channel resistance, and outflow rate, then become a prior that reweights which genes the variant ranker favors.

The simulator is the Lattice Boltzmann Method (LBM). LBM is a fluid solver that moves and mixes virtual particle populations on a regular grid. It runs directly on image voxels and is easy to make differentiable. The application is primary congenital hydrocephalus. In this disease, Jacquemin et al. found a significant burden of rare variants in cilia genes (p = 0.0082), yet obtained a definitive diagnosis in only 18% of 28 probands. The scientific question is whether a patient-specific physical phenotype contains diagnostic information about gene families that standard phenotype labels and literature knowledge graphs do not. The answer is checkable, and it is useful even if it is negative.

Core Claim

Variant rankers for multi-gene ("oligogenic") disease rank causal genes better when the patient description includes a mechanism inferred from physics rather than a symptom label. The leading oligogenic ranker is Hop (Prioritization of oligogenic variant combinations in whole exomes). It combines a variant-pair pathogenicity classifier with a random walk over a literature-built knowledge graph. Its authors acknowledge that the random walk favors well-studied genes.

We do not try to simulate individual mutations. Mapping a missense change onto a ciliary beat parameter is unsolved, and pretending otherwise would make the work unfalsifiable. Instead, we assign genes coarsely to mechanism families, using structural annotation that also exists for poorly studied genes. The first family is motile-cilia machinery, which drives wall propulsion. The second is primary-cilia and intraflagellar transport signaling, which acts through development and should not directly reduce propulsion. The third covers structural, vascular and extracellular factors, which act through channel narrowing or impaired drainage. The physics fit decides which family a patient's flow pattern is consistent with.

Section 2 of our research file found no study linking LBM to variant prioritization. The closest LBM work builds patient-specific physiological models without genotype. Geddes et al. used the HARVEY solver on pulmonary arteries, Rapaka et al. simulated cardiac electrophysiology, and Audigier et al. simulated liver ablation. We reuse their image-to-grid pipeline and point the output at a different consumer, a gene ranker.

Research Program

The first component is a differentiable CSF model built in Lettuce. Lettuce is an LBM library written in PyTorch, which makes the whole simulation available to automatic differentiation. The ventricles are segmented from MRI and voxelized directly onto the grid. Ciliary propulsion is represented as a tangential slip velocity on the ventricular wall. The model does not resolve individual cilia. This is a deliberate coarsening, because the claim concerns family-level mechanism.

CSF production enters as a source term in the choroid plexus regions. The narrow channel between ventricles (the aqueduct) and the outlets carry learnable resistances, and drainage is a learnable outflow condition. Gradient-based fitting matches simulated flow to measured aqueductal flow waveforms from cine MRI, where a patient has them, and to the observed pattern of ventricular enlargement.

Rapaka et al. found that boundaries on a regular grid produce staircase artifacts. We therefore report how sensitive the fitted parameters are to grid resolution. Following Geddes et al., who found that low geometric detail was enough for their pressure predictions, we start with coarse ventricular geometry and add detail only if it changes the fitted parameters.

The second component is an identifiability study, done before any genetics. We simulate synthetic patients with known perturbations of propulsion, resistance and drainage on real ventricular shapes. We then ask whether the inverse fit recovers the perturbed parameter well enough to separate the three mechanism families. Wall slip and channel resistance may trade off against each other given only imaging data. If they cannot be distinguished, the project stops here and reports that negative result. That outcome would be informative, because it would mean imaging alone cannot carry this signal.

The third component is integration with the ranker. The posterior over mechanism families becomes a family-weighted restart distribution in Hop's random walk. A simpler version adds the same weights as a term in a reranking step. Gene-to-family assignment uses cilia annotations such as the SYScilia list of 304 genes that Jacquemin et al. used, split into motile and primary subsets. We avoid literature co-citation so that the prior does not inherit the knowledge-graph bias it is meant to offset.

Baselines

The naive reference point is Hop run with only the phenotype terms ("hydrocephalus" plus any listed malformations). This is how unsolved cases are processed today. On its 50,400-exome synthetic benchmark, Hop reports that the true causal combination appears in the top 50 in 71% of cases, against about 23% for Exomiser and about 19% for OligoPVP. Because Hop is clearly the strongest existing method on this task, it is the baseline to beat. Exomiser and CADD (about 6% top-50) are included only as reference points.

All arms share the same exomes, the same VarCoPP2.0 pathogenicity scores and the same gene-to-family table. Each arm gets the same budget for tuning the weight given to the prior, chosen by nested cross-validation over patients. Beating Hop means a statistically supported gain in the rank of the causal gene for solved cases, using paired comparisons on the same patients. It also requires that genes with sparse literature coverage, defined as those in the lowest third of knowledge-graph connectivity, do not get worse.

The key ablation is a radiological label prior. It uses the same family-weighted restart, but the weights come from the radiologist's classification of the hydrocephalus as obstructive or communicating, with no simulation. If this ablation matches the physics prior, the simulation adds nothing.

A second ablation replaces the 3D LBM with a lumped compartment model. This model treats the ventricles, aqueduct and drainage as a few linked reservoirs with production, resistance and outflow, fitted to the same flow waveform. A third ablation randomly permutes the family assignments, to show that gains come from the mechanism content and not from adding an extra restart term.

Why not just…?

Why not just use the radiologist's obstructive-versus-communicating label? This is the most serious threat. That label already separates channel blockage from drainage failure, and it costs nothing. What it cannot do is separate weak wall propulsion from the other two causes. Wall propulsion is the motile-cilia signal that matters most for the cilia burden Jacquemin et al. reported. Whether the simulation adds information beyond the label is an empirical question, and the radiological-label ablation answers it directly. If the two priors perform the same, we will report that the label is sufficient.

Why not just use a lumped compartment model or a standard finite element solver? A compartment model can represent resistance and drainage, but it has no ventricular wall on which to place propulsion. Its predictions about propulsion therefore depend on a modeling assumption rather than on the patient's anatomy. A finite element solver could represent the wall, but it requires body-fitted meshes. LBM has been reported to run 10 to 45 times faster than finite elements for cardiac electrophysiology (LBM-EP) and about 60 times faster for liver ablation (Audigier et al.). Lettuce also provides gradients without extra engineering, which matters because we fit parameters for every patient. The claim of the proposal does not depend on LBM specifically, and the compartment-model ablation tests whether 3D wall physics is needed at all.

Why not just run RNA-seq or long-read sequencing on the unsolved cases? These assays find variants that exome sequencing missed. Reported RNA-seq yields range from 7.5% to 34% (Beyond the exome). They do not say which of the many rare variants a patient carries in cilia genes actually matters. The physics prior works on the ranking problem, so it complements these assays and does not compete with them. Where RNA-seq solves a case, that case becomes extra labeled evaluation data for us.

Proposed Experiments

The first experiment tests identifiability on synthetic patients. The pass criterion is that the inverse fit classifies the perturbed mechanism family well above chance, together with a resolution-sensitivity curve showing that this holds as the grid changes.

The second experiment tests ranking mechanics on a modified version of Hop's synthetic benchmark. We restrict spiked-in combinations to OLIDA entries in cilia and hydrocephalus-related families from OLIDA, and attach a simulated mechanism posterior to each synthetic exome. This experiment is partly circular, because we generate the phenotype from our own family map. It can therefore test only how sensitive ranking is to an imperfect posterior. It cannot test the underlying biology.

The decisive biological test needs real patients with both imaging and exome data. The Jacquemin cohort is the natural pilot: 28 probands, with 5 solved and 3 strong candidates. We know of no reported result on Lacuna for ranking hydrocephalus genes from imaging, and a cohort this small can only show direction, not significance. We will therefore pool additional solved cases with available MRI. The primary metric is the change in causal-gene rank against phenotype-only Hop and against the radiological-label ablation.

A second test needs no diagnoses. Among probands carrying cilia-gene variants, we ask whether carriers of motile-cilia variants show lower fitted propulsion than carriers of primary-cilia variants. If propulsion does not differ, the burden Jacquemin et al. found would point to a developmental mechanism rather than reduced flow at the ventricular wall. That conclusion would be worth reporting whether or not the ranking gain appears.