Massively parallel characterization and predictive modelling of neuronal regulatory variation

Technical Summary: Massively Parallel Characterization and Predictive Modelling of Neuronal Regulatory Variation

Problem Statement
A significant challenge in genomics is understanding the functional consequences of noncoding variants residing in cis-regulatory elements (CREs). While large-scale initiatives have cataloged millions of candidate CREs, the precise mechanisms encoding cell-type-specific regulatory programs remain incompletely understood. A central question is how the functional impact of regulatory variants relates to their population allele frequency (AF). While rare variants are often assumed to have larger effects due to purifying selection, this relationship is confounded by demographic history and drift. Furthermore, current computational models based on genomic location or transcription factor binding sites (TFBS) provide incomplete predictions of regulatory "grammar." There is a need for large-scale functional assays in relevant cellular contexts to quantify regulatory activity, assess the sparsity of active elements, and evaluate the ability of state-of-the-art models to predict variant effects at single-nucleotide resolution.

Methodology
The authors performed a large-scale lentivirus-based Massively Parallel Reporter Assay (lentiMPRA) in human excitatory neurons differentiated from WTC11 iPSCs using NGN2 induction.

• Library Design: The study targeted 525 disease-associated genes (neurological, cardiac, and clinically actionable) and 25 random genes. From these loci, 27,209 unique genomic regions (cCREs) within 50 kb of Transcription Start Sites (TSSs) were selected from the ENCODE SCREEN v3 registry.
• Variant Selection: The library included 46,374 naturally occurring single nucleotide variants (SNVs) from gnomAD v3.1.2. This comprised over 22,000 common variants (AF > 1%) and over 24,000 rare and singleton variants. Rare variants were prioritized using an Enformer-based strategy to enrich for those with predicted activating or repressive effects on chromatin accessibility, alongside a randomly sampled subset to serve as an unbiased control.
• Experimental Execution: Sequences (270 bp oligos) were synthesized, cloned into lentiviral reporter constructs with unique barcodes, and transduced into NGN2-derived neurons in three biological replicates. Regulatory activity was quantified via RNA-to-DNA barcode ratios.
• Analysis and Modeling: The authors analyzed allelic effects using BCalm, stratified by allele frequency and mutational class. They benchmarked epigenomic annotations (ENCODE SCREEN v4) and transcription factor binding (ReMap2022). Finally, they evaluated the predictive performance of several sequence-to-function models (Enformer, AlphaGenome, CADD, and a cell-type-matched NGN2 chromBPNet) against the MPRA data. They also fine-tuned an AlphaGenome encoder on the MPRA dataset to assess transfer learning capabilities.

Key Results

• Regulatory Activity and Sparsity: Only a minority of tested cCREs exhibited significant activity relative to scrambled controls (2.1% increased, 4.8% decreased). However, elements annotated as brain cCREs were enriched for activity and displayed a near-balanced distribution of activating and repressive effects. Conversely, elements lacking brain annotations were strongly biased toward transcriptional repression in this neuronal context.
• Allele Frequency and Functional Impact: Significant allelic effects were detected across the entire allele frequency spectrum (common, rare, and singletons) at comparable rates (~2.6–3.0%). While there was a modest monotonic decrease in the proportion of activating variants from singletons to very common variants, the magnitude of effect sizes showed no strong dependence on allele frequency. This suggests that population frequency provides limited information regarding the per-variant regulatory impact within measurable MPRA effects.
• Context Dependence: The detectability and characteristics of variant effects were governed primarily by the baseline activity of the enclosing regulatory element and local sequence context, rather than allele frequency alone. Variants in low-activity elements were more likely to show increasing effects, while those in highly active elements showed decreasing effects.
• Transcription Factor Architecture: Regulatory effects were distributed across numerous transcription factors rather than concentrated in a few master regulators. Enrichment analyses identified neurodevelopmental regulators (e.g., SOX4, ASCL1, PAX6) in brain-annotated active elements, while non-brain annotated elements were associated with corepressor complexes.
• Model Performance: State-of-the-art sequence models (Enformer, AlphaGenome) showed modest discrimination (ROC-AUC ~0.57–0.58) for predicting significant allelic effects across all variants. Performance improved when restricted to variants within active elements (dCREs) or when using cell-type-matched models (NGN2 chromBPNet). Crucially, fine-tuning the AlphaGenome encoder on the MPRA data significantly improved performance (ROC-AUC ~0.78–0.81), demonstrating that pre-trained sequence representations can be effectively adapted to MPRA-specific signals.

Significance and Claims
The authors establish a large-scale functional variant catalog in human excitatory neurons, providing a benchmark for evaluating models of noncoding regulatory variation. The study claims that:

1. Frequency is a weak proxy: Population allele frequency carries limited information about the regulatory impact of individual variants; local regulatory context is a stronger predictor of effect detectability.
2. Combinatorial Architecture: Neuronal regulatory variation follows a combinatorial enhancer architecture involving many TFs rather than a small set of dominant regulators.
3. Model Limitations and Potential: While current general-purpose sequence models provide modest enrichment for functional variants, their performance is significantly enhanced by cell-type matching and task-specific fine-tuning.
4. Resource Utility: The dataset serves as a resource for mechanistic follow-up, offering a list of candidates for targeted CRISPR perturbation and allele-specific expression analyses, particularly for variants overlapping fine-mapped eQTLs and GWAS signals.

The paper concludes that while sequence models and epigenomic annotations are valuable heuristics, large-scale functional assays remain essential for resolving fine-grained allelic effects in disease-relevant cellular systems and for training the next generation of context-aware predictive models.