How Much Missing Heritability Breaks Monotonicity?

Genetic risk models that go beyond adding up allele effects, such as deep networks and boosting methods, beat standard polygenic scores on some diseases. Nobody knows what kind of non-additivity they are exploiting. Some of it could be biologically ordinary: a second copy of a risk allele adds less than the first (dominance), or two risk alleles together are worse than the sum of each alone. Some of it could be exotic: heterozygotes carry less risk than both homozygotes, or an allele is protective in one genetic background and harmful in another. Some could be overfitting to noise in rare genotype combinations.

We propose to separate these by training the same flexible risk model twice. One copy is unconstrained. The other is constrained to be monotone, meaning predicted risk can never go down when you add a copy of a risk allele, with each allele's direction fixed from an independent sample. The monotone model can still learn dominance and interactions that change how large an effect is. It cannot learn effects that flip direction. The held-out accuracy gap between the two copies is therefore an empirical upper bound on how much predictable non-additive heritability needs direction-flipping biology.

Core Claim

Our hypothesis is that, for most complex traits, nearly all of the out-of-sample gain from non-linear genetic models survives a monotonicity constraint. We also expect the constrained model to transfer better to other ancestries and to have more stable explanations, because the forbidden direction flips are where models fit linkage-disequilibrium artifacts. Linkage disequilibrium (LD) is the correlation between nearby variants that makes one SNP stand in for another, with population-specific strength. If the hypothesis holds, monotone non-linear scores are a safer default than unconstrained deep scores. If it fails for particular traits or loci, the failure localizes real sign epistasis or over-dominance, which is a finding in its own right. The expected exception is the HLA region in autoimmune disease.

This differs from the closest existing work in a specific way. GenoBoost fits per-variant genotype scores $(s_0, s_1, s_2)$ for 0, 1 and 2 copies of an allele. It does so without constraint, so it can produce over-dominance from noise at low-frequency variants, and it does not model interactions. i-LDSC detects local SNP-by-SNP interaction heritability in 24 of 25 UK Biobank traits, but it cannot say whether that variance is direction-preserving, and it makes no individual predictions. The DRIVE breast-cancer DNN and DeepPRS show that unconstrained networks gain accuracy, but they never ask which shape class the gain comes from. The one paper that pairs monotonicity with complex-disease data, deep geometric learning for Alzheimer's progression, constrains trajectories over time rather than dose of genotype, and it uses no genetic data.

Research Program

The model is a partially monotone network built from the hidden layers of Scalable Monotonic Neural Networks, which supplies the method. In that design, inputs designated monotone pass through units whose weights are forced non-negative by an exponential. Unconstrained inputs pass through ordinary units. A third kind of unit mixes the two streams while still guaranteeing monotonicity in the designated inputs. We reuse DeepPRS's idea of grouping variants into local sub-networks, one per LD block or gene, whose outputs feed a monotone combining layer. We do not use lattice layers as the default, because Deep Lattice Networks need a projection step whose cost grows as $O(S2^S)$ in the number of inputs per lattice. That rules out large blocks. Lattices remain an option for small blocks.

Orientation is the crux. Each SNP is recoded as a count of its risk allele, oriented using effect signs from a discovery split disjoint from training. A SNP's sign estimated on its own can disagree with its sign once correlated neighbors are accounted for. We therefore compare three orientation rules: marginal signs, signs from a joint penalized regression within each block, and orientation of whole LD blocks. SNPs whose sign is not confidently estimated go into the unconstrained stream. This follows the partial-monotonicity design that both monotone architectures support.

That design creates a dial: the fraction of SNPs that are constrained. Running the dial from 0% to 100% traces accuracy as a function of how much monotonicity is imposed. Finally, we add a local release step. Single LD blocks are moved into the unconstrained stream one at a time, and we test whether held-out accuracy improves. Blocks that improve are flagged as candidate sites of sign epistasis or over-dominance.

Baselines

The naive reference point is an additive score fitted to the same preselected SNPs. On DRIVE this is BLUP, which reports an AUC of 64.2% against LDpred's 62.4% in the breast-cancer DNN paper. Any non-linear model must clear this bar.

The strongest existing non-linear competitor on DRIVE is that paper's unconstrained DNN, which reports an AUC of 67.4% and 18.8% recall at 90% precision. We re-run it with its published preprocessing (SNPs filtered at $p<10^{-3}$, 80% dropout) on our splits. On UK Biobank diseases, the strongest competitor is GenoBoost. It reports a 2.5% improvement for psoriasis and ranks first or second on 7 of 12 diseases. For Alzheimer's, DeepPRS reports a genotype-only AUC of 0.7245.

The most important baseline is architecture-matched: an identical network with the monotone weight constraint removed. It gets the same SNPs, splits, and hyperparameter-search budget, and its regularization strength is tuned on the same validation fold. This is the ablation that isolates the constraint. A second ablation orients alleles randomly. If random orientation does as well as GWAS orientation, the gain is generic shrinkage of the hypothesis space, not biology.

"Beating" a baseline has a specific meaning here. On the primary question, the constrained model should match the unconstrained twin within confidence intervals from bootstrapping held-out individuals. On secondary outcomes, it should do measurably better: cross-ancestry AUC, and agreement of attributions between explanation methods across retraining runs.

Why not just…?

Why not just constrain GenoBoost so that $s_1$ lies between $s_0$ and $s_2$? This is the most serious threat. If the non-additive gain is mostly per-locus dominance concentrated at HLA, which GenoBoost's own results suggest for autoimmune disease, a constrained GenoBoost could match our network. We build exactly this variant as a baseline. If it ties our model, we learn that monotone interactions add nothing beyond monotone dominance. That is a clean, publishable answer, and our accuracy-versus-constraint curve still holds.

Why not just use gradient-boosted trees with built-in monotone constraints? Standard boosting libraries offer this option, and it is a cheap way to get a monotone learner with interactions. We find no reported result on Lacuna for monotone boosted trees on genotype data. Our claim is about function classes rather than a particular architecture, so we run constrained and unconstrained boosted trees as a second pair of twins. If both pairs show the same gap, the conclusion is robust to the learner.

Why not just rely on heavy dropout or L1/L2 penalties? Penalties shrink effect sizes but do not stop direction flips. The breast-cancer DNN already used 80% dropout, yet its LIME and DeepLift explanations shared only 23% of top SNPs. SMNN found that unconstrained MLPs reached test error nearly seven times that of the monotone model on noisy synthetic data. That result is not genetic, so we test it directly: the unconstrained twin gets a full regularization sweep.

Why not just add lasso-selected pairwise interaction terms within local windows, as i-LDSC models? This captures local interactions but not interactions across blocks. It also does not separate direction-preserving from direction-flipping interactions unless the signs of the interaction coefficients are themselves constrained, so we include it as a baseline.

Proposed Experiments

We start with simulation. We generate genotypes by resampling haplotypes from the 1000 Genomes European panel and plant five architectures: purely additive, dominance, magnitude epistasis, sign epistasis, and over-dominance. The decisive check is that the constrained-versus-unconstrained gap is near zero for the first three architectures and positive for the last two. The local release step should recover the planted blocks. We report how orientation error from LD inflates the gap under each orientation rule.

On real data, the cohorts are DRIVE breast cancer (49,111 subjects), the UK Biobank disease set used by GenoBoost, UK Biobank height as a quantitative trait with strong i-LDSC interaction signal, and Alzheimer's disease. Metrics are AUC, recall at high precision, and $R^2$ for height.

Two further checks decide whether the constraint buys more than accuracy parity. The first is cross-ancestry transfer: we train on European UK Biobank participants and evaluate on participants of other ancestries, where the portability failures noted in this work predict that unconstrained models degrade more. The second is attribution stability: across five random seeds, we measure the overlap of top-ranked SNPs between explanation methods.

We also apply one biological sanity check. Running the local release step on rare-variant gene burdens should not flag genes with known graded dose-response, such as SLC13A1 and SLC26A1 from the metabolomics–exome study. In autoimmune disease, HLA is the expected place where the monotone model loses accuracy. If HLA were not flagged there, we would treat that as a warning that the method lacks power.