Gene expression is governed by the DNA sequence, which is read out through complex interactions between transcription factors (TFs), co-activators, and chromatin. Massively Parallel Reporter Assays (MPRAs) provide a high-throughput framework for functionally characterizing how regulatory DNA sequences impact the expression of a model gene. MPRAs have also proven to be useful for measuring the effects of genetic variation, where each allele is typically tested in the center of ~200 bp of genomic context cloned into the MPRA, but the impact of variant position and local context remains largely unexplored. In this study, we systematically investigate how shifting the position of a variant within an MPRA probe influences its regulatory activity using models that predict expression in MPRAs from DNA sequence. We find that while the direction of variant effects is usually preserved across positions, the magnitude of expression changes can vary substantially depending on where the variant is placed within the construct. This positional bias appears to be largely explained by the strong position-dependent activity of TFs whose binding the variants perturb. In a subset of cases, interactions consistent with cooperativity between TFs also contribute to position-specific effects. ~1% of variants appear to disrupt RNA polymerase III (Pol III) promoters within Alu elements, resulting in position-specificity because both A and B boxes are required for function and exclusion of either motif due to window shifts disrupts the variants' effects. However, we saw little evidence to support the hypothesis that the positional dependence of variant effects resulted from the redundancy of motifs. Overall, our study demonstrates the complexity of cis-regulatory grammar and how it can confound the interpretation of regulatory variants. ### Competing Interest Statement R.T. has filed intellectual property related to MPRA and MPRA models. The other authors declare no competing interests.

The rapid expansion of genomics datasets and the application of machine learning has produced sequence-to-activity genomics models with ever-expanding capabilities. However, benchmarking these models ...