Short tandem repeats are stretches of DNA where a one- to six-letter motif is repeated many times (for example CAGCAGCAG…). They mutate very fast because DNA polymerase slips when copying them, so a single locus can carry many different lengths across a population, or be missing entirely in some individuals. Two pieces of recent work point the same way. In Arabidopsis, these repeats explain a substantial share of trait variation that ordinary single-letter variants miss (Mutations of short tandem repeats explain abundant trait heritability in Arabidopsis). In barley, simple and tandem repeats are among the most active enhancers measured (Identification of Highly Repetitive Enhancers with Long-range Regulation Potential in Barley via STARR-seq). Yet the heritability study treats each repeat as a plain number, its length, entered into a linear model. Meanwhile the deep DNA sequence models now used to interpret non-coding variants are built and benchmarked almost entirely on single-letter changes (Benchmarking DNA Sequence Models for Causal Regulatory Variant Prediction in Human Genetics).
We propose to close that gap with a concrete operation we call in-silico repeat titration. For every polymorphic repeat, we build the actual DNA sequence of each length allele seen in the population, including the "absent" allele, with its real flanking context. We pass each of these sequences through a sequence model and read off a predicted functional value, which gives a per-locus curve of predicted effect against repeat count. Each individual's genotype is then recoded through that curve before it enters standard variance-component and prediction models. The claim is specific and falsifiable. If regulatory repeat effects depend on sequence context and are often non-linear in length (saturating, peaked, or reversed by flanking motifs), then curve-encoded genotypes should predict gene expression in held-out individuals better than raw length. They should also do better than cheap annotation-based weights. If they do not, raw length is enough and the deep model adds nothing.
Core Claim and Positioning
The closest existing work is the Arabidopsis study, which genotyped repeat length and presence/absence across 1,168 accessions. It mapped 3,871 expression-associated repeats and 651 splicing-associated repeats, and estimated average repeat heritability of 0.111 for expression, 0.143 for splicing and 0.101 for organism-level traits using a linear mixed model (LDAK). That paper motivates our proposal and also supplies the data. What it does not do is model how repeat sequence and context set the size or shape of an allele's effect. Each locus gets one linear slope in length. The Caragana genic-repeat study points to the same limitation (The Potential Role of Genic-SSRs in Driving Ecological Adaptation Diversity in Caragana Plants). It found only 264 of 3,666 polymorphic genic repeats with a linear length–expression relationship, which is consistent with, though not proof of, non-linear effects.
On the modeling side, three lines of work provide the scorers we reuse. The first is GPN, a masked DNA language model trained on Brassicales genomes that scores variants zero-shot by likelihood ratio and beats phyloP and phastCons in Arabidopsis (DNA language models are powerful predictors of genome-wide variant effects). The second is AlphaGenome, a one-megabase supervised model predicting thousands of human and mouse regulatory tracks (AlphaGenome: advancing regulatory variant effect prediction with a unified DNA sequence model). The third is the gene-centric evaluation that anchors predictions at the target gene's transcription start site rather than at the variant (Gene-centric evaluation of causal variant prediction for DNA models). All three were evaluated on single-letter variants. None scores a series of length alleles at one locus, and none feeds its scores into heritability estimation.
TraitGym is a constraint on the design rather than a motivation. It shows that sequence models discriminate poorly among distal complex-trait variants even for single-letter changes, so we cannot assume model scores carry signal for repeats. Our first experiment therefore tests whether the models respond sensibly to repeat length at all before anything is built on top of them.
Research Program
The primary setting is Arabidopsis, for three reasons: it is the only system in our material with population repeat genotypes, expression data and trait heritability together; it has a plant-trained DNA language model; and the published linear estimates give a direct reference point. Human transfer is a secondary aim.
The first step is building allele sequences. For each polymorphic repeat with at least three observed alleles, we construct the reference-coordinate sequence for every observed repeat count and for the deletion allele. Lengthening or shortening a repeat shifts everything downstream in a fixed-length model window. To avoid this, we keep the window centred on the repeat and trim the flanks symmetrically, which holds the repeat's distance to the target gene's start site constant. We also run a second version anchored at the start site, following the gene-centric design.
The second step is scoring. We use two scorers. The first is zero-shot: the GPN pseudo-log-likelihood of each allele sequence relative to the reference. The second is supervised: a small expression probe trained on GPN embeddings at the gene's start site. Each probe is fit on genes from chromosomes 1–3 using single-letter expression associations only, and applied to repeats on chromosomes 4–5, and vice versa. The repeat-encoding function therefore never sees repeat expression data from the chromosomes it is scored on.
The third step is encoding and estimation. For locus $i$ and individual $j$ with repeat count $L_{ij}$, the genotype entry becomes
$$X_{ij} = f_i(L_{ij}),$$
where $f_i$ is the model-predicted response curve for locus $i$, standardized across individuals. Raw-length encoding is the special case where $f_i$ is the identity. We then refit the same LDAK partitioned models as the original study, with single-letter variants, small indels, transposable elements and repeats as separate components, swapping only the repeat encoding. We also fit penalized predictors of expression for held-out accessions.
The fourth step is a small experimental validation of the curves. We take several hundred repeat loci that sit in predicted enhancers or promoters, synthesize each at a ladder of repeat counts, and assay them in plant protoplasts with the self-transcribing reporter design used for barley. This gives measured dose–response curves to compare against predicted ones.
Baselines
The naive reference point is the single-letter-variant-only model. We fit LDAK and the expression predictor using only single-letter variants and small indels, on the same accessions and the same cross-validation folds. This measures how much repeat information adds at all, and it directly tests whether nearby single-letter variants already tag the repeats.
The strongest existing method is the linear-length encoding of the Arabidopsis study: repeat length as a number plus a presence/absence indicator, entered into partitioned LDAK. Its reported averages are 0.111 for expression, 0.143 for splicing and 0.101 for traits. We are aware of no reported out-of-sample expression-prediction accuracy for repeats; no reported result found on Lacuna for that metric. We will rerun this baseline ourselves with identical genotype calls, accession splits and tuning budget. That means the same LDAK settings, and for the predictor, the same penalty grid searched over the same folds as the curve-encoded model. Beating it means a higher mean held-out expression $R^2$ across genes with a repeat in the cis window, and a higher held-out $R^2$ on the repeat-carrying subset, with gene-level paired tests and a pre-registered minimum gain. A rise in in-sample heritability alone will not count. Recoding can inflate variance components without improving prediction, so prediction on held-out accessions is the deciding metric.
The cheap competitor most likely to tie us is an annotation-weighted linear model. Each repeat's linear length effect is scaled by prior weights from motif class, unit size, distance to the start site, overlap with accessible chromatin, and whether the GPN score of the reference allele is extreme. This gets most of what a sequence model "knows" without running titration, and it receives the same tuning budget.
Two ablations isolate the key ingredient, which is the context-specific shape of $f_i$. A linear projection of the curve keeps only the model's predicted slope per locus and discards curvature. A shuffled-flank control recomputes curves after scrambling flanking sequence while keeping the repeat motif. If the shuffled-flank curves predict as well as the real ones, the gain comes from motif identity rather than regulatory context.
Why not just…?
Why not just keep linear length, as the Arabidopsis study did? That approach already shows that repeats carry heritability, and it is our main baseline. Its weakness is structural. One slope per locus cannot represent an effect that peaks at intermediate length or depends on what flanks the repeat. The Caragana result, where most repeats in differentially expressed genes show no linear length correlation, hints that such cases are common. Whether they are common enough to matter for prediction is exactly what the held-out comparison measures. If the curve-encoded model does not beat this baseline, we will report that linear length is sufficient.
Why not just use annotation weights instead of a deep model? This is the real threat. Location in open chromatin and distance to the gene may carry most of the signal, and they cost nothing to compute. The annotation-weighted baseline is designed to settle this directly. If it matches curve encoding, the defensible conclusion is that titration adds nothing, and we would redirect the work toward annotation priors for repeat heritability. The ablation separating curve shape from slope tells us whether any deep-model advantage comes from non-linearity or simply from better per-locus weighting.
Why not just rely on single-letter variants to tag the repeats? Repeat mutation rates of $1.93 \times 10^{-2}$ to $4.40 \times 10^{-3}$ per locus per generation mean recurrent length changes that erode linkage with neighbouring single-letter variants. The joint Arabidopsis models still attributed variance to repeats after including all single-letter variants, indels and transposable elements. We will check this again in our own folds with the single-letter-only baseline.
Why not just fine-tune a sequence model on the known expression-associated repeats? With roughly 3,900 such loci, many in linkage with each other, a fine-tuned model would be trained and evaluated on overlapping signal and would leak into the heritability estimate. The chromosome-split probe trained only on single-letter associations avoids that leakage. A fine-tuned variant restricted to the training chromosomes can be added as an extra arm.
Proposed Experiments and Decisive Checks
The first decisive check comes before any heritability work: do the models respond to repeat length meaningfully? On the protoplast titration ladder, we compute the per-locus Spearman correlation between predicted and measured activity across repeat counts, and compare it with a null obtained by permuting counts. If predicted curves are flat or uncorrelated with measured ones, typical of models trained on genomes where repeats are masked or rare, the program stops at a clean negative result about DNA model blind spots. That result would itself be useful for the field.
The second check is held-out expression prediction across the 413 RNA-seq samples. We split by accession and test across chromosome folds, comparing all baselines and ablations. Success means curve encoding beats both linear length and annotation weights on repeat-carrying genes. The same comparison runs for splicing, reusing the published splicing-associated loci.
The third check concerns organism-level traits, using AraPheno traits including Fruit Number, for which repeats were reported to explain up to 65.97% of variation. With only about 200 phenotyped accessions, we treat this as a consistency check: estimates should move in the same direction as expression without losing calibration. It is not the primary evidence.
The fourth, secondary, check is human transfer. We apply AlphaGenome titration to repeats inside Open Targets GWAS credible sets and GTEx-linked regions referenced by AlphaGenome. We ask whether curve-derived scores separate fine-mapped loci from matched controls, using a TraitGym-style matched-control design. TraitGym itself contains only single-letter variants, so this is a new repeat benchmark that we would construct and release, and we report no comparison number for it in advance.
References
- Mutations of short tandem repeats explain abundant trait heritability in Arabidopsis
- Identification of Highly Repetitive Enhancers with Long-range Regulation Potential in Barley via STARR-seq
- Benchmarking DNA Sequence Models for Causal Regulatory Variant Prediction in Human Genetics
- The Potential Role of Genic-SSRs in Driving Ecological Adaptation Diversity in Caragana Plants
- DNA language models are powerful predictors of genome-wide variant effects
- AlphaGenome: advancing regulatory variant effect prediction with a unified DNA sequence model
- Gene-centric evaluation of causal variant prediction for DNA models