Mapping Genes That Make or Remove Structure

Genetic mapping of shape asks a simple question: which stretches of chromosome make individuals in a crossbred population look different? The answer depends on how "looking different" is measured. Plant and animal geneticists usually reduce an image to a few numbers, such as total root length or tip count. An alternative is to warp every image onto a shared template and test each template location separately. That warping step assumes every individual has the same parts arranged differently, like a rubber sheet that can stretch but never tear. Root systems break this assumption as a matter of course. One inbred line grows a lateral root where another has none. That is a difference in what exists, not in where it sits.

We propose to split each image into two maps. The first is a smooth warp that records where shared structure has moved. The second is a creation-and-removal map that records tissue with no counterpart in the template. Each map is then tested location by location for linkage to chromosomal markers in a mapping population. The method that makes this split is called metamorphic registration, and it already exists in medical imaging. There, the creation map has only ever been used to hide tumors so healthy anatomy can be aligned. We treat it as a heritable trait instead. We also fix the method's main weakness, which is deciding how much change to attribute to the warp and how much to creation. Current methods settle this with hand-drawn masks. We settle it with genetically identical replicate plants: the right split is the one on which clones of the same line agree most.

Core Claim

Mapping populations of inbred lines grown in replicate come with free supervision for the ambiguous warp-versus-creation decomposition. Among decompositions that fit the images equally well, the one that maximizes the fraction of variance shared by clones (the broad-sense heritability, estimated without looking at markers) separates genetic from non-genetic variation best. Mapping the two resulting fields separately should find, and place in space, loci that scalar traits and pure-warp mapping miss. The clearest example would be a locus that controls whether lateral roots emerge in one zone of the root, as distinct from a locus that controls how long roots grow.

What Exists and What Is Missing

The medical metamorphic methods supply the machinery. Maillard et al. write the metamorphosis evolution equations as a residual network. This gives a 0.16 s runtime per image pair and a ventricle Dice score of 71.9 on BraTS 2021 when a lesion mask is used to regularize. MetaMorph combines a segmentation network with geodesic shooting, the standard way of computing a smooth warp from an initial velocity field. Its regularization is piecewise and bounded by the predicted anomaly mask. It reduces ventricle landmark error to 4.64 mm, against 15.02 mm for VoxelMorph and 10.53 mm for a metamorphic autoencoder. Both methods discard the creation term as nuisance, and both need a mask to stop that term from "painting away" real deformation. We reuse their solvers. We replace the mask with a genetic criterion and keep the creation term as the object of study.

Czolbe et al. motivate the idea that the places where smooth alignment fails carry signal. Their method detects emerging organelles with an AUC of 0.877 and brain tumors with an AUC of 0.797, against 0.616 for tracking the Jacobian determinant (the local volume change of the warp). This is also a constraint on our design: raw warp statistics are poor detectors of topological change, so a pure-diffeomorphic pipeline is unlikely to suffice. Their output is an anomaly heat map, not a quantitative trait, and they never apply it to genetics.

On the genetics side, Biffi et al. is the closest precedent. They align 1,124 human hearts to a fixed-topology mesh and regress wall thickness on SNPs at every vertex. They use threshold-free cluster enhancement (TFCE), which boosts spatially extended signals before permutation testing. This recovered associations for four of six left-ventricular-mass SNPs that the global scalar missed. Their result motivates the spatial approach, but their pipeline requires every heart to share one mesh, and it studies a human cohort rather than a recombinant mapping population. Subramanian et al. show that spatial point-process summaries (Ripley's K) raise correlation with gene expression from 0.74 to 0.81. Those summaries carry no anatomical coordinates, so they cannot say where a genetic effect acts.

For statistics we reuse R/qtl2. It reconstructs founder haplotypes in multi-parent populations with hidden Markov models and runs linear mixed-model genome scans with leave-one-chromosome-out kinship. The application motivation comes from Ma et al.. Across 243 wheat recombinant inbred lines, they mapped five scalar root traits and found 10 QTLs, each explaining 4.5–8.4% of phenotypic variance. The traits pool primary elongation and lateral branching into single numbers. Liu et al. show that eight-way MAGIC populations narrow QTL intervals compared with biparental crosses, which supports using a multi-parent panel. MAGIC lines are populations bred from many founders over several rounds of intercrossing.

Research Program

Registration model. Each individual's image $S_i$ is modeled as $S_i(x) \approx T(\phi_i(x)) + \Delta I_i(x)$. Here $T$ is the population template, $\phi_i$ is a smooth invertible warp produced by geodesic shooting, and $\Delta I_i$ is the creation-and-removal field. We minimize image mismatch plus $\lambda_v \|v_i\|^2 + \lambda_z \|\Delta I_i\|_1$. The term $v_i$ is the initial velocity that generates $\phi_i$, and the two $\lambda$ weights set how expensive warping and creation are. The L1 penalty makes creation sparse, which suits discrete emergence events.

Template construction. The template is built groupwise. We alternate between averaging the images after warping them to the template and subtracting their creation fields, and re-centering the velocities to zero mean as the centrality layer of MultiMorph does. The authors of MultiMorph state that it cannot handle topological change. The subtraction step is our way around that limitation. It lets a template form for a population whose members differ in which parts they have.

Choosing the warp-versus-creation balance. The ratio $\lambda_z/\lambda_v$ decides how much variation goes to each field. We choose it on a grid by maximizing the median, across template locations, of heritability estimated from within-line versus between-line variance of both fields among replicate plants. No marker data is touched in this step. Genome-wide permutation of genotype labels therefore remains a valid null for all later tests. As a guard against degenerate solutions, we also require image fit to stay within a fixed tolerance of the best fit on the grid.

Mapping. At each template location we fit an LMM genome scan in R/qtl2 to two traits. The first is the log-Jacobian of $\phi_i$, which is smooth, so TFCE applies as in Biffi et al. The creation field violates the smoothness that TFCE assumes, so we split it into a binary presence trait (whether $|\Delta I_i|$ exceeds a threshold) and a magnitude trait conditional on presence. We test both. To control error jointly over markers and locations, we permute lines and record the maximum statistic over the whole genome-by-space scan. To keep computation feasible, the template is downsampled to roughly $10^4$ locations, and scans across locations share a single kinship decomposition.

Baselines

The naive reference point follows Ma et al.. We map scalar traits (total length, area, volume, tip count, and lateral root count from ChronoRoot 2.0) in R/qtl2 on the same lines, with the same kinship correction and the same permutation threshold. On their wheat data, Ma et al. report 10 QTLs explaining 4.5–8.4% of variance each. No image-based QTL result is reported on Lacuna for the population we will phenotype, so the comparison will be internal. Beating this baseline means finding loci at the same genome-wide error rate that the scalars do not, and assigning each locus a spatial footprint that the scalars cannot give.

The strongest existing spatial method is a pure-diffeomorphic pipeline in the style of Biffi et al., built on a MultiMorph template: voxel-wise log-Jacobian mapping with TFCE. It is the direct competitor because it already outperformed scalars in the one imaging-genetics study in the file. It will use the same images, the same template resolution, the same GPU budget, and the same grid search over its own regularization weight, tuned by the same heritability criterion. The fairness condition is that its tuning gets the same number of grid points as ours.

A graph baseline builds per-line traits from root skeletons: branch counts per zone along the primary root, insertion angles, and lengths by branch order. Zones are defined by fixed fractions of primary root length.

A learned-embedding baseline trains an autoencoder on the root images and maps each latent dimension. Its training budget matches our registration budget.

Two ablations isolate what is new. The first fixes $\lambda_z/\lambda_v$ at a hand-set value instead of tuning it by replicate heritability; this tests whether the genetic criterion matters. The second maps only $\phi$ or only $\Delta I$; this tests whether the decomposition matters. For registration quality, MetaMorph's 4.64 mm ventricle landmark error on BraTS 2021 is the bar. Our mask-free, heritability-tuned variant cannot be tuned that way on BraTS, which has no replicates. There it will run with the nearest grid value, and it should stay within a reported margin of that bar rather than beat it.

Why not just…?

Why not just map scalar root traits? Scalars answer "how much" but not "where", and they mix mechanisms together. In Biffi et al., four SNPs were invisible to global left-ventricular mass and detectable spatially. Our simulation, described below, plants a zone-specific emergence locus. It will show directly whether the scalar lateral root count detects that locus at equal power.

Why not just use diffeomorphic registration with strong smoothing? When a structure is missing, the warp must crush or stretch tissue to make up the difference. On BraTS, this left VoxelMorph with 15.02 mm landmark error against 4.64 mm for MetaMorph (Wang et al.). Czolbe et al. found Jacobians to be weak detectors of topological change (AUC 0.616). The diffeomorphic baseline is run at full strength, so if it matches our power, the experiments will show it.

Why not just extract skeleton graph traits? This is the real threat for roots, because a skeleton already represents branching explicitly. Its limits are that zone boundaries must be chosen by hand, and that it does not separate branch position from branch presence in a common coordinate frame. If graph traits match our method on real and simulated data, we will report that. The contribution then narrows to the replicate-based choice of decomposition and to organisms whose structures are not tree-like, such as leaf venation. We will not claim more than that.

Why not just map autoencoder latents? Latent dimensions are not spatially indexed and are not identifiable across training runs. A locus found this way is hard to interpret and hard to replicate. We test this by measuring locus stability across five random seeds for each method.

Proposed Experiments

The first experiment is a semi-synthetic power study that settles most of the baseline questions under known ground truth. We take real founder-haplotype genotypes from the Arabidopsis 19-way MAGIC lines (677 lines) distributed with R/qtl2. We plant three kinds of simulated QTL: a zone-specific lateral root emergence locus, an elongation locus, and a null. Each line gets rendered 2D root images from a stochastic growth simulator with four replicates. Every method is scored on detection power at a genome-wide error rate of 0.05, on interval width, and on spatial localization error of the planted effect. The decisive check is that emergence loci are found in $\Delta I$ presence maps but not in Jacobian maps, and that elongation loci are found in Jacobian maps. If the two effect types are not separated this way, the decomposition claim fails.

The second experiment checks that the registration method still works on standard data. We run it on BraTS 2021 against MetaMorph's landmark error and on the Learn2Reg OASIS task (Siebert et al.). The file reports no numeric OASIS baseline for Learn2Reg; its overall winner was ConvexAdam.

The third experiment uses real data: time-lapse ChronoRoot imaging of a subset of the Arabidopsis MAGIC lines with replicates. We ask whether spatial loci replicate across two independent growth batches, and whether heritability-tuned decompositions give higher replication rates than hand-tuned ones. If replicate-based tuning does not beat a hand-set balance on replication, the core claim is falsified.