MobiDeep: an AI-based meta-score for scoring non-coding DNA variations

MobiDeep: an AI-based meta-score for scoring non-coding DNA variations | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article MobiDeep: an AI-based meta-score for scoring non-coding DNA variations Abdelhakim Bouazzaoui, Jean-Madeleine de Sainte Agathe, Simon Cabello-Aguilar, and 15 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8823759/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background The interpretation of non-coding variants (NCVs) from genome sequencing represents a major bottleneck in the diagnosis of rare diseases. Existing variant effect predictors (VEPs) show variable performance across different genomic contexts, and a lack of region-specific clinical guidance hinders accurate variant prioritization. This study aimed to rigorously benchmark state-of-the-art VEPs to define region-specific thresholds and to develop MobiDeep, a novel meta-score designed to improve NCV prioritization. Methods We curated a high-confidence dataset of 448 pathogenic NCVs (ClinVar, HGMD, literature) and 38,146 presumed benign NCVs. Critically, variants affecting splicing were excluded to focus on strictly regulatory mechanisms. We benchmarked the performance of ReMM, CADD, GPN-MSA, Cactus241way, and phyloP, both globally and stratified by genomic region (e.g., 5'UTR, 3'UTR). Subsequently, we developed MobiDeep, a neural network integrating these five scores, optimized using Optuna and validated on an independent holdout set of pathogenic NCVs. Results Benchmarking confirmed that no single tool is universally optimal, with performance varying significantly by genomic context; while ReMM excelled in non-coding exons (AUROC = 0.987), GPN-MSA demonstrated superior performance for 3'UTRs (AUROC = 0.901). We established data-driven clinical thresholds, identifying an optimal global cutoff of 10.37 for CADD v1.7, validating previous works of CADD ≥ 10 for regulatory variants and 0.80 for ReMM. Building on these insights, MobiDeep significantly outperformed all individual predictors on an independent test set, achieving an AUROC of 0.973 and an AUPRC of 0.888. In large-scale simulations mimicking a diagnostic, MobiDeep prioritized causal variants effectively, placing 52.0% and 75% within the top 5 and top 20 ranks respectively. Furthermore, the model correctly prioritized all Clinvar pathogenic variants in the recently discovered RNU4-2 non-coding gene. Conclusions Our findings confirm that individual predictors and uniform thresholds are insufficient for interpreting the diverse landscape of non-coding variants. We demonstrate that region-specific calibration is essential for accurate prioritization.. Our meta-score MobiDeep improves classification performance compared to existing tools. This meta-score serves as a robust filter to streamline the identification of high-confidence variants, thereby facilitating focused manual review and subsequent biological validation in diagnostic settings. Genome sequencing Non-coding variants Variant prioritization Meta-score Machine learning Figures Figure 1 Figure 2 Figure 3 Background While Genome Sequencing (GS) is now routinely used in rare disease molecular diagnostics, the yield remains limited, often below 50% of cases 1 . This persistent diagnostic gap suggests that a substantial proportion of disease-causing variants resides within non-coding regions (NCRs). These regions, comprising 98% of the human genome, remain difficult to interpret in clinical practice 2 . Consequently, major genomic medicine initiatives, including the French Plan for Genomic Medicine 2025 3 and the Genomics England 100k Genomes Project 4 , have identified the comprehensive interpretation of Non-Coding Variants (NCVs) in GS data as a critical objective for providing the best care and realizing the full potential of genomic medicine. NCRs harbour a diverse array of regulatory elements, such as promoters, enhancers, and untranslated regions (UTRs), that orchestrate spatiotemporal gene expression. The disruption of these elements is an established, albeit under-characterized, mechanism of Mendelian disease. For instance, pathogenic variants in the 5'UTR of GATA4 are associated with atrial septal defects 5 , while distant-acting enhancer variants affecting SOX9 expression cause Pierre Robin sequence 6 . Despite such examples, regulatory variants are significantly underrepresented in clinical databases. The Human Gene Mutation Database 7 (HGMD; retrieved 03/10/2025) catalogues fewer than 320 regulatory disease-causing variants, while ClinVar 8 (retrieved 09/02/2025) contains only several hundred pathogenic variants annotated within key regulatory elements, a stark contrast to the tens of thousands of well characterized protein-coding variants (64,286). This disparity stems not from a lack of biological relevance, but from fundamental challenges in NCVs interpretation. In contrast to the relatively straightforward application of the universal genetic code in coding regions, regulatory elements operate through complex, context-dependent mechanisms that are difficult to predict computationally. To address this interpretive challenge, several computational variant effect predictors (VEPs) have been developed. Despite their utility, the clinical application of these tools faces several limitations. First, their performance varies substantially across different genomic contexts and is often inadequately benchmarked outside of canonical regulatory elements 9 . Second, tools like ReMM that are trained on known disease variants risk data circularity and inflated performance estimates when evaluated on similar datasets. Finally, the assumption of a universal score threshold is untenable, as recent work highlights the necessity of region-specific calibration for accurate variant classification 9 . In this study, we address these challenges through a multi-faceted approach. First, to provide a practical guide for their clinical use and directly address the need for context-specific thresholds recently highlighted by Villani et al 9 , we conducted a comprehensive and rigorous benchmark on multiple VEPs using a curated dataset of 448 pathogenic and a total 38,146 neutral NCVs. We specifically excluded from the pathogenic datasets variants known or predicted to alter splicing, as the aim of the study was to focus on the other pathological mechanisms, considering the splicing prediction problem has already been adequately addressed 10 , 11 , 12 . We established the VEPs performance and optimal classification thresholds not only globally but also within distinct functional genomic regions (e.g., 5’UTR, 3’UTR, upstream regions). This granular analysis, revealing the context-dependent strengths and weaknesses of each predictor, motivated the development of MobiDeep: a multilayer perceptron meta-score designed to integrate these five tools into a single, calibrated pathogenicity score that achieves superior performance. Furthermore, recognizing that clinical utility extends beyond simple classification, we implemented a ranking simulation framework to evaluate MobiDeep's performance in prioritizing causal variants within large, clinically realistic backgrounds of benign variations. Our aim is to provide not only a new computational tool but also a transparent framework for its responsible implementation, adhering to community-driven best practices for reproducibility and clinical utility 13 . Results Global and stratified benchmarking reveals variable performance of existing VEPs We were able to obtain the scores for all 448 pathogenic variants for the 3 ReMM, phyloP_Primates and Cactus241way VEPs, however, as this study is based on pre-computed datasets, we only scored 400/448 variants for GPN-MSA and 416/448 for CADD v1.7. In the global analysis, which combined all non-coding regions, the integrative models clearly outperformed standalone conservation scores (Fig. 1 A, 1 B; Table 1 , Supplementary Table S1 ). ReMM v0.4 demonstrated the highest overall discriminative ability, achieving a mean AUPRC of 0.814 and an AUROC of 0.953 (Table 1 ), these results were supported by strong F1-score of 0.660 and the highest MCC at 0.623. This was followed by CADD v1.7 (AUPRC = 0.744, AUROC = 0.913) and GPN-MSA (AUPRC = 0.735, AUROC = 0.921), which also performed robustly. The conservation-based tools, Cactus241way (AUPRC = 0.686) and phyloP_Primates (AUPRC = 0.615), provided a solid baseline but showed lower overall efficacy. The mean optimal global thresholds were found to be 0.800 ± (0.029) for ReMM, 10.373 ± (0.184) for CADD, -2.549 ± (0.312) for GPN-MSA, 0.304 ± (0.081) for PhyloP_Primates and 1.345 ± (0.059) for Cactus241way Vertebrates. However, score distributions for all predictors show significant overlap between pathogenic and benign classes (Fig. 1 C-G), illustrating the challenge of relying on a single threshold for perfect classification. We note that while our data-driven thresholds for GPN-MSA (-2.3) and ReMM (0.81) are optimal, more stringent thresholds of -7.0 and 0.961, respectively, are commonly used to isolate only high-confidence pathogenic variants for PP3 criteria, a stringency supported by our score distribution plot (Fig. 1 E). A context-dependent analysis revealed significant performance heterogeneity across five harmonized genomic regions, indicating that a tool's global performance does not guarantee its utility in a specific locus (Fig. 1 , Supplementary Figure S1 ). The specialized nature of ReMM was evident in its superior performance in Intronic (AUROC = 0.942) and Non-coding exon (AUROC = 0.987) regions, where it most effectively separated pathogenic from benign variants (Table 1 , Fig. 1 and Supplementary Figure S1 ). GPN-MSA showed its key strength in the 3' UTR region (AUROC = 0.901, AUPRC = 0.604), outperforming all other tools in this context (Table 1 ). This makes GPN-MSA the tool of choice for variants located within 3’UTR regions but may necessitate using the 3’UTR specific threshold at -3.59 ± (0.465) (Fig. 1 E, right). CADD v1.7 demonstrated the most consistency and balance across the genomic regions. While it excelled in Non-coding exon regions with an AUROC of 0.959 and an AUPRC of 0.819, its efficacy was markedly reduced in the 5' UTR (AUROC = 0.783, AUPRC = 0.641), where its score distributions for pathogenic and benign variants showed considerable overlap (Supplementary Figure S1 ). The performance of pure conservation scores was also context-dependent; for instance, PhyloP_Primates, which was respectable globally (AUROC = 0.865), showed sharply reduced efficacy in the 5' UTR (AUROC = 0.738, AUPRC = 0.566), suggesting its evolutionary signal is less informative for variants in this specific region. Table 1 Classification Performance of Predictors at Optimal Thresholds in Global and Regional Analyses. Analysis Group Tool AUROC AUPRC F1-Score TP (FN) FP (TN) Global ReMM v0.4 0.953 0.814 0.660 371 (77) 310 (4170) GPN-MSA 0.921 0.735 0.548 318 (82) 448 (4032) CADD v1.7 0.913 0.744 0.648 335 (81) 283 (4197) Cactus241way_V 0.863 0.686 0.617 322 (126) 275 (4196) PhyloP_P 0.865 0.615 0.504 337 (111) 558 (3913) 3' UTR GPN-MSA 0.901 0.604 0.572 32 (10) 38 (422) ReMM v0.4 0.893 0.599 0.568 33 (13) 38 (422) CADD v1.7 0.854 0.654 0.596 31 (11) 31 (429) Cactus241way_V 0.847 0.628 0.674 32 (14) 17 (443) PhyloP_P 0.794 0.394 0.363 30 (16) 94 (366) 5' UTR ReMM v0.4 0.853 0.747 0.667 91 (38) 54 (342) GPN-MSA 0.800 0.646 0.566 61 (51) 44 (352) CADD v1.7 0.783 0.641 0.564 80 (36) 88 (308) Cactus241way_V 0.746 0.623 0.553 69 (60) 53 (343) PhyloP_P 0.738 0.566 0.538 79 (50) 87 (309) Intron ReMM v0.4 0.942 0.814 0.702 67 (17) 40 (800) GPN-MSA 0.922 0.804 0.711 65 (13) 40 (800) CADD v1.7 0.872 0.721 0.655 64 (19) 49 (791) PhyloP_P 0.872 0.681 0.551 66 (18) 89 (748) Cactus241way_V 0.826 0.704 0.659 60 (24) 38 (799) Non_coding ReMM v0.4 0.987 0.931 0.784 69 (4) 35 (695) exon CADD v1.7 0.959 0.819 0.658 59 (7) 55 (675) GPN-MSA 0.952 0.747 0.504 59 (5) 116 (614) Cactus241way_V 0.919 0.704 0.632 63 (10) 64 (666) PhyloP_P 0.900 0.579 0.457 63 (10) 141 (589) Upstream ReMM v0.4 0.959 0.806 0.639 98 (11) 101 (989) GPN-MSA 0.938 0.803 0.586 85 (14) 108 (982) CADD v1.7 0.936 0.797 0.719 90 (14) 57 (1033) Cactus241way_V 0.891 0.733 0.598 90 (19) 102 (986) PhyloP_P 0.892 0.689 0.615 86 (23) 86 (1001) AUROC: Area Under the Receiver Operating Characteristic curve. AUPRC: Area Under the Precision-Recall curve. F1-Score: The harmonic mean of precision and recall. TP (FN): number of true positives (and false negatives) out of the total pathogenic variants for that group. FP (TN): number of false positives (and true negatives) out of the total sampled benign variants for that group. Figures are the average of the 10,000 iterations performed. VEPs are ordered per AUROC performance for each analysed group. Bold: best performing VEP for the considered metric. Global performance (left panel): Mean Receiver Operating Characteristic (ROC) (A) and Area Under the ROC Curve (AUROC) (B), curves, with shaded areas representing ± 1 standard deviation. Area under ROC and PR are reported. Thresholds (left) and Region-specific analyses (right panels): Score distribution and region-specific performance for ReMM v.04 (C), CADD 1.7 (D), GPN-MSA (E), Cactus241way_v (F), and PhyloP_P (G). The distributions for pathogenic (red) and benign (blue) variants are shown. The vertical dashed line indicates the mean optimal classification threshold determined by the Youden's J Index, providing a practical reference for each tool's decision boundary when used globally. Performance was evaluated on a subset of variants located in each genomic region (5’UTR, 3’UTR, Intronic, Non-coding exon and upstream) illustrating the performance variability of tools in specific genomic contexts (violin plots). MobiDeep: An ensemble learning based meta-score The final MobiDeep model is a three-hidden-layer Multilayer Perceptron (MLP) with a [115, 144, 175] neuron architecture and ReLU activation functions (Supplementary Figure S2). A permutation-based feature importance analysis revealed that the integrative models, ReMM and CADD, are the most influential features driving MobiDeep's predictions. GPN-MSA provided the next largest contribution, while the conservation scores served as complementary evidence, confirming that MobiDeep effectively weighs and synthesizes information from all five input VEPs (Supplementary Figure S3D). MobiDeep achieves state-of-the-art classification performance When evaluated on the independent benchmark dataset, MobiDeep demonstrated superior classification accuracy, outperforming all individual VEPs in both AUROC and AUPRC, all metrics are detailed in Supplementary Table S1 . MobiDeep achieved an AUROC of 0.974 (± 0.001) and an AUPRC of 0.889 (± 0.011) (Fig. 2 A-B, Supplementary Table S1 ). This improvement highlights the benefit of integrating multiple informative features within an optimized, non-linear framework. Global, region-specific and clinically relevant thresholds An optimal global threshold of 0.60 was determined by maximizing the Youden's J Index, as shown by the score distributions in Fig. 2 C. Our analysis also yielded the following region-specific thresholds: 0.92 for 5’UTR, 0.60 for 3’UTR, 0.40 for Intronic, 0.85 for Non-coding exonic and 0.61 for Upstream regions., highlighting the performance adaptability across genomic contexts (Fig. 2 D). A second generic threshold has been established, corresponding to a 95% Positive Predictive Value (95% CI: 0.941–0.983), defined as a MobiDeep raw probability score of 0.968. This corresponds to a log-score of > 1.5 and captures approximately 55% of true pathogenic variants in our independent test set. Comparative (A) Receiver Operating Characteristic (ROC) and (B) Precision-Recall (PR) curves for MobiDeep and the five other predictors achieving a state-of-the-art AUROC of 0.973 and an AUPRC of 0.888. (C) score distributions for benign and pathogenic variants confirm the model's high discriminative power with a mean Youden threshold of 0.6 in global (all regions). (D) Youden thresholds for global (0.6) and harmonized regions highlights the specific thresholding needed for each genomic context with specific thresholds of 0.6, 0.92, 0.4, 0.85 and 0.61 for 3’UTR, 5’UTR, Intronic, Non-coding exonic and Upstream regions respectively. MobiDeep robustly prioritizes pathogenic variants in large-scale ranking simulations The primary test of a VEP's clinical utility is its ability to prioritize causal variants within the large lists generated by GS. MobiDeep demonstrated highly scalable performance in these simulations (Table 2 , Fig. 3 ). Table 2 Summary of MobiDeep's absolute ranking performance for 137 independent VOIs against increasing noise backgrounds (5 replicates). Background (N) Median Rank Mean Rank Top 1% Top 1 Top 5 Top 10 Top 20 1,000 1.0 17.9 88.2% 54.7% 72.3% 80.3% 91.4% 10,000 5.0 171.3 89.6% 21.8% 62.0% 71.5% 74.9% 100,000 45.0 1,697.4 89.1% 5.4% 15.6% 24.7% 34.2% When ranked against a background of 1,000 neutral variants, the median rank for a pathogenic VOI was just 1 (Table 2 ). The prioritization performance was maintained even as the search space expanded: the median rank increased to only 5 against 10,000 variants and 45 against 100,000 variants. The model's ability to enrich for causal variants is further highlighted by its success in placing them at the absolute top of the ranked list, a critical feature for manual review in a clinical setting. Across all 5 replicates, 72.3% of VOIs achieved a median rank in the Top 5 when tested against 1,000 benign variants. This success rate remained high at 62.0% against a 10,000-variant background and 40.1% against a 100,000-variant background. The model's scalability is illustrated in Table 2 and Fig. 3 A. The median rank increases sub-linearly with the background size, demonstrating that the model's performance remains robust even as the number of candidate variants grows substantially. Furthermore, when viewed as a percentage of the total pool, the model's enrichment power is remarkably stable. Nearly 90% of pathogenic VOIs were consistently ranked within the top 1% of their respective pools, a figure that remained nearly constant whether the pool contained 1,000 or 100,000 variants. The cumulative distribution of ranks (Fig. 3 A) illustrates this scalability, with the curves showing only modest rightward shifts as background complexity increases. Figure 3 B further emphasizes this prioritization power by displaying the cumulative percentage of VOIs by rank for N = 10,000. The steep initial rise of this curve demonstrates that MobiDeep efficiently concentrates pathogenic variants at the very top of ranked lists—approximately 62% of VOIs achieve a rank within the top 5 positions, and over 70% are ranked within the top 10. Analysis of individual variants across all replicates for N = 10,000 confirms this overall consistency. The vast majority of VOIs are consistently ranked highly, with only a small subset of variants ranking poorly (Fig. 3 C). Notably, specific variants such as NM_000044.6:c.-547C > T, which disrupts an upstream open reading frame (uORF), consistently rank poorly across all simulation conditions. This demonstrates that while certain pathogenic mechanisms, particularly uORF disruption in 5' UTRs, remain challenging for the model, MobiDeep is nonetheless a robust and reliable tool for drastically reducing the search space for the vast majority of NCVs in a clinical setting. An analysis of rank distribution stratified by VEP-annotated variant effect provides additional mechanistic insight (Fig. 3 D). While most variants in effect categories such as introns and upstream regions were ranked very highly (median rank near 1), variants in the 5_prime_UTR_variant category showed a wider distribution with a notable tail extending to higher ranks. This finding corroborates our earlier observation that 5' UTR variants are more challenging to prioritize and suggests that specific regulatory mechanisms active in this region—such as uORF disruption, ribosome binding site alterations, or other translational regulatory elements—are not fully captured by the current feature set. This represents a clear direction for future model enhancement, potentially through integration of specialized predictors like MORFEE 24 , 25 that explicitly model uORF effects. (A) Cumulative distribution function (CDF) of ranks for 137 pathogenic VOIs against increasing backgrounds (N = 1,000, 10,000, and 100,000 neutral variants). Over 88% of VOIs ranked within the top 1% across all background sizes. X-axis: log scale. (B) Cumulative percentage of VOIs by rank (N = 10,000 background). 89.0% of VOIs ranked within top 1,000; 62.7% within top 10. X-axis: log scale. (C) Individual rank trajectories for 137 VOIs (5 replicates, N = 10,000 background). Most VOIs (X-axis) show consistent high ranking (low rank numbers). Outliers include 5' UTR variants with uORF disruption. Y-axis: log scale. (D) VOI rank distributions by VEP-annotated effect (N = 10,000 background). Most categories show tight distributions near top ranks. The 5_prime_UTR_variant category displays wider distribution with poorer-ranked outliers, reflecting limited capture of uORF-related mechanisms. Complementary test on pathogenic RNU4-2 variants As a final validation step, we scored a set of 30 variants in the RNU4-2 non-coding RNA retrieved from Clinvar VCF at 2025/05/30, which were not present in our first extraction of ClinVar (Supplementary Table S3). The model performed well. Of the 19 variants in this set classified as P/LP in ClinVar, MobiDeep assigned a score exceeding our high-confidence 95% PPV threshold (0.968, log_score > 1.5) and Non-coding Exon threshold (0.85) for all the 19 variants. In contrast, all three "Likely benign" variants were correctly assigned scores below these specific pathogenicity thresholds. Remaining variants were annotated in Clinvar as of unknown clinical significance, and most of them were scored pathogenic by MobiDeep. Discussion and Conclusions The interpretation of NCVs remains a primary bottleneck, limiting the full diagnostic potential of GS. Our study systematically addresses this challenge by first rigorously benchmarking state-of-the-art VEPs and then developing MobiDeep, a meta-score that demonstrates superior performance in NCV prioritization. This work responds directly to the challenges highlighted by community-wide efforts like the CAGI challenge, which reveal the need for more accurate and context-aware tools for variant prioritization in rare disease diagnostics 26 . Our initial benchmarking of existing VEPs revealed two critical findings. First, no single tool is universally optimal; performance is highly context dependent. While integrative tools like ReMM and CADD showed strong global performance, stratified analysis uncovered specific strengths and weaknesses. For instance, ReMM demonstrated clear dominance in intronic (AUROC = 0.942) and non-coding exon regions (AUROC = 0.987), whereas GPN-MSA showed superiority in 3' UTRs (AUROC = 0.901) (Table 1 , Fig. 1 ). Conversely, the performance of all predictors was notably diminished in 5' UTRs, suggesting that key pathogenic mechanisms in this region, such as uORF disruption, may not be adequately captured by these models. Second, our derivation of optimal thresholds confirms the necessity of region-specific calibration. Our optimal global CADD 1.7 threshold of 10.37 closely replicates the CADD ≥ 10 threshold proposed by Villani et al. 9 for assigning moderate evidence of pathogenicity (PP3) to regulatory variants. Similarly, our optimal ReMM threshold of 0.80 aligns with their REMM ≥ 0.86 threshold. Recent work by Tenywa et al. 27 provides further independent validation of this region-specific approach, deriving CADD v1.7 thresholds for distinct noncoding regions using ClinVar data. Our independently derived region-specific CADD thresholds (Supplementary Table S1 , Supplementary Figure S1 ) show notable concordance with their findings: both studies identify 5'UTR regions as requiring substantially higher thresholds (17.37 in our dataset vs. 16.79 in theirs), while thresholds for 3'UTR and non-coding exonic regions cluster in the 11–13 range. The substantial threshold variation observed across genomic contexts—ranging from 8.39 for intronic regions to 17.37 for 5'UTRs in our analysis, and from 11.08 to 25.5 (for splicing variants, excluded in our study) across different noncoding categories in Tenywa et al 27 . This demonstrates, empirically, the inadequacy of applying a single genome-wide cutoff. This convergence of independently derived, region-specific thresholds across multiple independent studies confirms that missense-variant thresholds are not applicable to NCVs and establishes genomic context-aware calibration as an essential principle for accurate noncoding variant interpretation. (Table 1 , Fig. 1 ). The significant overlap in score distributions between pathogenic and benign variants (Fig. 1 C-G), even for the best-performing VEPs, illustrates the inherent difficulty in achieving perfect classification. This presents a major clinical challenge. In a diagnostic setting, a stringent threshold chosen to minimize false positives (i.e., increase specificity) will inevitably increase the rate of false negatives, causing true disease-causing variants to be missed. The clinical and personal cost of a missed diagnosis is high, therefore, the primary goal for a clinically useful tool must be to manage this trade-off, with a strong emphasis on maximizing sensitivity to avoid discarding potential causal candidates. The fact that no single VEP excels across all genomic regions argues strongly for an integrative strategy. We hypothesized that by combining the distinct, and at times complementary, predictive signals from different tools, a meta-score, could achieve a more robust and accurate classification. Our development of MobiDeep was a direct test of this hypothesis. This approach proved effective, as MobiDeep outperforms each individual predictor on our independent test set (AUROC = 0.973, AUPRC = 0.888; Fig. 2 A-B). The feature importance analysis (Supplementary Figure S3D) provides further insight: while it confirms the strong predictive power of integrative tools like ReMM and CADD, it also shows that conservation-based scores, though weaker on their own, still contribute meaningfully to the final prediction. This implies that the most accurate classification arises from a synergistic combination of all features, which MobiDeep is designed to leverage. The model was carefully optimized to avoid overfitting (Supplementary Figure S3), ensuring that this performance reflects a genuine ability to generalize. The clinical utility of a VEP is ultimately determined by its performance in prioritizing variants from large candidate lists. In the simulation results, the low median rank, which scales sub-linearly with the background size (Table 2 , Fig. 3 A), suggests that the typical pathogenic variant will be placed near the top of a ranked list. However, the stark divergence between the median and the much higher mean rank (e.g., 45.0 vs. 1,697.4 at N = 100,000) points to a significant issue: a tail of pathogenic variants that are ranked very poorly. This suggests that while MobiDeep is reliable for the majority of variants, it can fail substantially in a minority of cases. This trade-off is further highlighted by the model's performance on the most stringent success criterion: placing the causal variant at rank 1. This metric diminished from 54.7% to 5.4% as the background set grew from 1,000 to 100,000 variants. Therefore, while MobiDeep consistently places nearly 90% of pathogenic variants in the top 1% of the list, clinical biologists can use the top-ranked candidate as a strong hypothesis requiring verification. The tool is best understood as a method for generating a highly enriched shortlist (e.g., the top 20–50 candidates) that facilitate focused manual review. The outlier analysis in our results (Fig. 3 C, 3 D) directly links the poorly ranked variants to specific biological contexts, such as 5' UTRs, underscoring that these performance characteristics reflect specific biological contexts (e.g. 5’UTRs) that benefit from complementary approaches. Importantly, our simulations represent a deliberately challenging 'filter-free' scenario, that does not incorporate the standard filtering steps used in clinical practice (e.g. restricting analysis to a gene panel, filtering by allele frequency, or leveraging patient phenotype terms). This context is essential for understanding how MobiDeep is designed to be implemented: not as a replacement for the established diagnostic pipeline, but as a flexible and powerful component within it. Its utility is best realized through a dual-threshold strategy tailored to different stages of the workflow. For initial, high-throughput screening where maximizing sensitivity is paramount, the global threshold of 0.60 (or the even more granular region-specific thresholds) can be applied to ensure potential candidates are not prematurely discarded. Conversely, for the final step of prioritizing pre-filtered list of challenging VUS, the high-confidence threshold (calibrated to a 95% PPV) can help to pinpoint the most likely candidate for manual review and functional validation. However, this flexibility embodies a critical trade-off. While the high-confidence threshold provides very strong evidence for the variants it flags, it only captures approximately 55% of the pathogenic variants in our independent test set. Careful review of candidates that fall below this stringent cutoff remains essential. Ultimately, our simulations serve as a proxy for clinical reality, and real-world implementation is the necessary next step to confirm the usefulness of this integrated strategy. Despite its overall strong performance, MobiDeep has areas for future enhancement that define our roadmap. Its current scope focuses on single nucleotide variants (SNVs) and its predictive power is constrained by its five input features. For instance, our ranking analysis revealed that a subset of variants, particularly in 5' UTRs, were consistently ranked poorly (Fig. 3 B). Post-hoc analysis of these variants, like NM_000044.6:c.-547C > T, confirmed they disrupt upstream Open Reading Frames, a mechanism not explicitly modeled by the input VEPs. As shown by Meguerditchian et al. 25 , uORF-altering variants represent a significant fraction of functional NCVs. The future integration of features from specialized tools like MORFEE 25 is a clear priority to improve MobiDeep's mechanistic coverage and address this specific weakness. Similarly, deep intronic variants without predicted splicing effects represent another mechanistic frontier requiring enhanced interpretation frameworks. The example of NM_003024.3(ITSN1):c.3661 + 1376T > C (Supplementary Figure S2), located + 1376 nucleotides from the canonical splice donor, illustrates this challenge. At such distances, conventional splicing mechanisms cannot account for pathogenicity. Several non-mutually-exclusive mechanisms may explain the functional impact of these variants. First, some annotated "intronic" variants may in fact affect regulatory elements within alternative 3' UTRs of minor transcript isoforms that terminate within what is annotated as an intron of the canonical, longer transcript. This phenomenon of overlapping isoforms with differential 3' UTR usage could be particularly relevant given the critical role of 3' UTR composition in determining mRNA stability, localization, and translational efficiency. Alternative mechanisms include disruption of intronic enhancers, polyadenylation signals, or non-coding RNA elements that regulate expression in a tissue or developmental-stage-specific manner. Distinguishing among these mechanisms for individual variants requires systematic functional characterization combined with improved transcript-level annotation that captures tissue-specific and cell-type-specific isoform expression. As with uORF-disrupting variants, the future integration of specialized predictive features that explicitly model these regulatory mechanisms would enhance interpretability for this important class of variants. Collectively, these examples underscore that while MobiDeep achieves strong overall performance through integration of existing tools, further improvements will require expanding its mechanistic scope through incorporation of specialized features that capture variant effects not adequately represented in current VEPs. Furthermore, this work represents a comprehensive retrospective validation and serves as proof of concept; prospective validation on large, diverse cohorts of undiagnosed patients is the necessary next step to confirm the model's real-world clinical impact. Methods Dataset curation We established a rigorous data curation pipeline to generate distinct, high-confidence datasets for model training, testing and independent evaluation. All variant coordinates were based on the GRCh38/hg38 human reference genome. We aggregated a set of pathogenic non-coding variants (NCVs) from several sources. The primary source was the ClinVar VCF file (release 2025-02-09) 8 . This was supplemented with variants from the Human Gene Mutation Database (HGMD Professional v.2024.1), from which we extracted "Disease Causing Mutations" (DM class) located in regulatory regions such as non-coding RNA, miRNA, and TFBS (queried March 10, 2024) 7 . The dataset was further enriched with manually curated NCVs from published literature and functionally validated variants provided by collaborating French diagnostic genetics laboratories. All variants (SNVs), regardless of their source, were processed through a uniform quality control pipeline. To ensure variants were strictly non-coding and to minimize confounding effects from strong splicing alterations, we used Ensembl Variant Effect Predictor (VEP v103) 14 to annotate them against all available transcripts, without applying transcript selection filters (i.e., --flag_pick_allele was not used) or GENCODE Basic restrictions. A variant was excluded from our dataset if any of its annotated consequences across any transcript matched protein-coding categories (missense_variant, stop_gained, stop_lost, start_lost, frameshift_variant, inframe_insertion, inframe_deletion, or protein_altering_variant) or if it was located at canonical splice sites (± 1, ± 2). This stringent filtering approach ensured that variants with mixed consequences, for instance, annotated as 5_prime_UTR_variant on multiple transcripts but as missense_variant on a single transcript, were conservatively excluded to maintain a strictly non-coding dataset. Additionally, variants with a predicted SpliceAI 12 delta score > 0.2 (maximum across acceptor gain, acceptor loss, donor gain, and donor loss) were removed from the collection (based on SpliceAI pre-computed dataset v1.3) to further minimize splicing-related confounding effects.. This filtered collection of NCVs was then partitioned into two subsets. The primary training and benchmarking cohort was defined as the high-confidence Pathogenic / Likely Pathogenic (P/LP) set (N = 448), composed exclusively of variants present in ClinVar with a review status of "criteria_provided,_multiple_submitters,_no_conflicts", "reviewed_by_expert_panel", or "practice_guideline" (i.e., ≥ 2 stars). This gold-standard set was primarily distributed across 5' UTR (n = 129), upstream (n = 109), intronic (n = 84), non-coding exons (n = 73), 3’UTR (n = 46) and other (i.e., intergenic, n = 7) regions, based on the VEP annotation. All other variants that passed our filtering criteria, including those sourced from HGMD, literature, collaborators, or those in ClinVar with a < 2-star status, were compiled into the independent P/LP holdout test set (N = 137). This set was held in reserve for further evaluation of the model on unseen data and for the ranking simulations. For benign variants, a pool of presumed benign variants was derived from the TraitGym mendelian_traits_full dataset 15 , a resource containing ~ 4.95 million variants. From this pool, we curated a large-scale set of 38,146 unique benign NCVs that were used as the source for negative controls in all subsequent region-specific analyses. A ratio of 1–10 was used at benchmarking (bootstrapping) and model training (4,480 fixed neutral variants). It is important to note that these variants are relatively frequent and therefore have a very low prior probability of being pathogenic for rare Mendelian diseases. While this does not definitively prove their functional neutrality on an individual basis, it provides a robust and conservatively selected set of negative controls for a rare disease context. While most of our pathogenic variants were sourced from the public ClinVar database (release 2025-02-09), ensuring full transparency, the dataset was also supplemented with variants from the HGMD Professional database (v.2024.1) and collaborating French diagnostic laboratories. Due to commercial licensing restrictions for HGMD Pro and regulatory constraints for patient data from clinical laboratories, these specific variant lists cannot be shared publicly, with the exception of those from Clinvar. However, to ensure maximum reproducibility, we have meticulously documented the version numbers of the databases. This allows for the precise reconstruction of our methodology. Model training Our model was trained and evaluated on a fixed dataset constructed from the core variant sets described previously. We combined the entire set of 448 high-confidence P/LP variants with 4,480 benign variants, which were randomly sampled once from our benign pool to create a static cohort. This final dataset of 4,928 variants, maintaining a strict 1:10 ratio, was then deterministically partitioned into a 90% training set (403 P/LP and 4,032 benign variants) and 10% hold-out test set (45 P/LP and 448 benign variants). Independent dataset For a final, completely unbiased assessment of model performance, we utilized the independent P/LP test set (N = 137), which had no overlap with the training data. This set was used in two validation experiments: The first is an independent benchmarking where the 137 P/LP variants were paired with 1,370 newly randomly sampled benign variants (maintaining the 1:10 ratio) to directly compare the performance of the final MobiDeep model against existing VEPs on unseen data. The second is an in-silico ranking simulation where they served as the Variants of Interest (VOIs) in our large-scale ranking simulations to evaluate MobiDeep's ability to prioritize causal variants within large, clinically realistic backgrounds of neutral variants. Benchmarking the VEPs Choice of VEPS Five VEPs were selected for benchmarking (CADD 16 (v1.7), ReMM 17 (v0.4), GPN-MSA 18 , PhyloP_Primates 19 , and Cactus241way 20 ), based on their published performance, resolution (1 base pair), diverse underlying methodologies and availability of pre-computed datasets covering the whole human genome. The Regulatory Mendelian Mutation score (ReMM v.0.4) 17 , is a random forest model supervised on clinically annotated variants. It was specifically trained to distinguish known pathogenic non-coding variants from HGMD against a background of common variants, using a curated feature set optimized for regulatory function, including epigenomic data such as DNase I hypersensitivity, key histone modifications (H3K4me1/3, H3K27ac), and eQTL data. In contrast, Combined Annotation Dependent Depletion (CADD 1.7) 16 is an integrative model trained to distinguish observed human variation from a simulated genomic background, independent of clinical labels. Its power derives from the comprehensive integration of over 90 diverse annotations, combining foundational conservation metrics with a vast catalogue of ENCODE 21 and Roadmap Epigenomics data 22 , as well as other genomic and transcript-level features. Representing a newer paradigm, GPN-MSA (genomic pretrained network with multiple-sequence alignment) 18 is a sequence-based deep learning model that leverages a transformer architecture to learn complex evolutionary constraints directly from the DNA context within large-scale multiple sequence alignments of 100 vertebrate species. Finally, to provide a foundational and orthogonal measure of evolutionary constraint, we included phyloP conservation scores, which quantify nucleotide-level selection. We utilized scores derived from two distinct alignments: a primate-specific alignment (phyloP_Primates) 19 and a broad 241-mammals alignment (Cactus241way) 20 , thereby capturing constraint across different evolutionary timescales. Pre-computed, genome-wide scores for GRCh38 were downloaded from their official sources. Scores for each variant were extracted using tabix and home-made python scripts. Benchmarking VEPS The benchmarking analysis was performed over 10,000 bootstrap iterations. In each iteration, a test dataset was constructed by combining the fixed set of 448 high-confidence P/LP variants with 4,480 benign variants selected via sampling with replacement from the global benign pool, maintaining a strict 1:10 pathogenic-to-benign ratio. For region-stratified analysis, a similar bootstrap was performed within each harmonized genomic region (5' UTR, 3' UTR, Intron, Upstream, Non-coding exon), sampling benign variants only from the corresponding regional pool. For each of the 10,000 datasets, multiple performance metrics (Confusion matrix, AUROC, AUPRC, F1-score, MCC) were calculated, and an optimal classification threshold was determined by maximizing the Youden's J Index (Sensitivity + Specificity − 1). This process generated mean and standard deviation for performance estimates for each VEP, both globally and across specific genomic contexts. Model development Model selection We developed MobiDeep, a deep learning classifier based on a Multilayer Perceptron (MLP) architecture implemented using scikit-learn v1.6.1 library. The MLP architecture was selected for its well-established capability in handling complex, non-linear classification tasks and its effectiveness in learning hierarchical feature representations from high-dimensional biological data. The model was optimized using Optuna, maximizing Area Under the Precision-Recall Curve (AUPRC) on 5-fold cross-validation to ensure robust performance. Feature selection and final model optimization Five VEPs, CADD 1.7, GPN-MSA, ReMM v.0.4, Cactus241way, and Phylop_Primates were chosen as inputs for MobiDeep: to prepare these features and address the 1:10 class imbalance, we constructed a multi-step pipeline using the scikit-learn v1.6.1 library. This structure is critical as it ensures that all preprocessing steps, especially resampling, are applied only to the training data within each fold of our cross-validation, preventing data leakage. The pipeline consisted of three sequential steps: 1) Iterative Imputation where missing values were imputed using an IterativeImputer with a RandomForestRegressor estimator; 2) Feature Scaling where all features were subsequently normalized to a [0, 1] range using a MinMaxScaler; 3) Class Imbalance Resampling: The Synthetic Minority Over-sampling Technique (SMOTE) was applied to the training data to generate synthetic examples of the minority (pathogenic) class. The final MobiDeep architecture and its hyperparameters were determined via a 200-trial Optuna study that optimized this entire pipeline for AUPRC. This entire workflow was encapsulated in a single scikit-learn pipeline object to ensure that all data transformations are consistently and reproducibly applied during any prediction task. Performance evaluation framework After training was complete, the final, optimized pipeline was evaluated a single time on the unseen 10% hold-out test set. We calculated standard classification metrics (AUROC, AUPRC, confusion matrix, F-scores) on this set to obtain an unbiased estimate of the model's generalization performance. Clinical thresholds For clinical applicability, we determined optimal probability thresholds for global and genomic region-specific analyses using bootstrap resampling (10,000 iterations with replacement) on 448 pathogenic and 38,146 benign variants. Thresholds were selected by maximizing Youden's J index on the ROC curve to optimize sensitivity and specificity. We also establish a more stringent threshold to achieve a Positive Predictive Value (PPV) of 95% for high-confidence predictions for manual curation. In silico ranking simulation To assess MobiDeep's utility in a realistic filtering scenario, we conducted a large-scale ranking simulation using the independent test set (n = 137). For each VOI, we created test pools by mixing it with N neutral variants bootstrapped from the whole TraitGym dataset, where N was 1,000, 10,000, and 100,000. To ensure robustness, this process was replicated 5 times for each VOI at each background size using different random seeds. In every simulation, all N + 1 variants were scored by the MobiDeep pipeline and ranked. We recorded the absolute rank of the VOI and calculated the mean/median rank across the 5 replicates. We also calculated the percentage of VOIs that ranked in the Top 1, 5, 10, and 20 absolute positions, and within the Top 1%, 5%, and 10% of the total pool size. Complementary test on pathogenic RNU4-2 variants To test the model's generalization capabilities on a specific gene and disease mechanism not represented in our training data, we curated a set of 30 variants in the RNU4-2 gene from ClinVar (accessed 2025-05-30). We applied the trained MobiDeep model to these variants and evaluated its ability to assign high, pathogenic-range scores to the known P/LP variants using the region specific (non-coding exon) and the stringent PPV 95% thresholds. Open-source availability and adherence to FAIR principles To promote reproducibility and community use, and to ensure our framework is Findable, Accessible, Interoperable, and Reusable (FAIR), the complete MobiDeep framework, including the trained model pipeline is available on GitHub at ( https://github.com/mobidic/MobiDeep ). MobiDeep is provided for academic and non-commercial use only, under a GNU GPL (GPLv3) license. An Apptainer container is also provided to ensure a fully reproducible environment. For users without a computational background, MobiDeep is accessible via a user-friendly web application, MobiDetails 23 . Finally, to maximize utility and interoperability, we are also releasing a pre-computed dataset of MobiDeep scores for 8,773 billion single nucleotide variants (SNVs) covering 94.7% of all genomic positions across the GRCh38p14 reference genome. The dataset includes the raw score and a log-transformed score for easier interpretation. Declarations Acknowledgements We would like to thank the ANPGM (Association Nationale des Praticiens en Génétique Moléculaire), Fondation Groupama « Vaincre les Maladies Rares » and the FHU GenOMedS for their support. We would also like to express our gratitude to the MoBiDiC platform (Montpellier Bioinformatics for Clinical Diagnosis). Our thanks also go to the Montpellier Laboratory of Computer Science, Robotics, and Microelectronics (LIRMM), especially to Charles Lecellier, Laurent Bréhélin, and Elliot Butz for their valuable expertise. Finally, we thank Roselyne Vallo (Biostatistician / Data manager, Pathogenesis & Control of Chronic & Emerging Infections unit, Inserm) for her expertise in biostatistics. 7. Funding This work was supported by Fondation Groupama « Vaincre les Maladie Rares », the French Association of Molecular Genetics Practitioners ANPGM (Association Nationale des Praticiens en Génétique Moléculaire) and FHU GenOMedS. 8. Author information Contributions Study design: AB, DB, SCA Data collection and management: AB, DB Mobidetails web app implementation: DB Expertise: JM-SA, SCA, VK, A-FR Data sharing: OE, JN, MK, LQ, TC, SC, PB, YJ, CR, MC Analysis and results interpretation: AB, DB, SCA, JM-SA Bioinformatics infrastructure and support: DB, CVG, OA Writing: AB, DB, all other authors contributed to the writing and review of the manuscript The authors read and approved the final manuscript Corresponding authors Correspondence to David BAUX at [email protected] or Abdelhakim BOUAZZAOUI at [email protected] 9. Ethics declarations Ethics approval and consent to participate This research only used publicly available data. The remaining patients’ variants underwent a full anonymization process. Consent for publication Not applicable Competing interests The authors declare no competing interests References Pandey R, et al. A meta-analysis of diagnostic yield and clinical utility of genome and exome sequencing in pediatric rare and undiagnosed genetic diseases. Genet Med. 2025;27:101398. Ellingford JM, et al. Recommendations for clinical interpretation of variants found in non-coding regions of the genome. Genome Med. 2022;14:73. Abadie C, et al. PFMG2025 –integrating genomic medicine into the national healthcare system in France. Lancet Reg Health - Eur. 2025;50:101183. Turnbull C, et al. The 100 000 Genomes Project: bringing whole genome sequencing to the NHS. BMJ. 2018;361:k1687. Mohan RA, et al. A mutation in the Kozak sequence of GATA4 hampers translation in a family with atrial septal defects. Am J Med Genet A. 2014;164A:2732–8. Gordon CT, et al. Identification of novel craniofacial regulatory domains located far upstream of SOX9 and disrupted in Pierre Robin sequence. Hum Mutat. 2014;35:1011–20. Stenson PD, et al. The Human Gene Mutation Database (HGMD®): optimizing its use in a clinical diagnostic or research setting. Hum Genet. 2020;139:1197–207. Landrum MJ, et al. ClinVar: public archive of interpretations of clinically relevant variants. Nucleic Acids Res. 2016;44:D862–868. Villani RM, McKenzie ME, Davidson AL, Spurdle AB. Regional-specific calibration enables application of computational evidence for clinical classification of 5’ cis-regulatory variants in Mendelian disease. Am J Hum Genet. 2024;111:1301–15. de Sainte Agathe J-M, et al. SpliceAI-visual: a free online tool to improve SpliceAI splicing variant interpretation. Hum Genomics. 2023;17:7. Leman R, et al. SPiP: Splicing Prediction Pipeline, a machine learning tool for massive detection of exonic and intronic variant effects on mRNA splicing. Hum Mutat. 2022;43:2308–23. Jaganathan K, et al. Predicting Splicing from Primary Sequence with Deep Learning. Cell. 2019;176:535–e54824. Livesey BJ et al. Guidelines for releasing a variant effect predictor. ArXiv arXiv:2404.10807v1 (2024). McLaren W, et al. The Ensembl Variant Effect Predictor. Genome Biol. 2016;17:122. Benegas G, Eraslan G, Song YS, Benchmarking. DNA Sequence Models for Causal Regulatory Variant Prediction in Human Genetics. BioRxiv Prepr Serv Biol. 2025. 10.1101/2025.02.11.637758 . 2025.02.11.637758. Schubach M, Maass T, Nazaretyan L, Röner S, Kircher M. CADD v1.7: using protein language models, regulatory CNNs and other nucleotide-level scores to improve genome-wide variant predictions. Nucleic Acids Res. 2024;52:D1143–54. Schubach M, Nazaretyan L, Kircher M. The Regulatory Mendelian Mutation score for GRCh38. GigaScience 12, giad024 (2022). Benegas G, Albors C, Aw AJ, Ye C, Song YS. GPN-MSA: an alignment-based DNA language model for genome-wide variant effect prediction. BioRxiv Prepr. Serv. Biol. 2023. 10.10.561776 (2024) doi:10.1101/2023.10.10.561776. Kuderna LFK, et al. Identification of constrained sequence elements across 239 primate genomes. Nature. 2024;625:735–42. Christmas MJ, et al. Evolutionary constraint and innovation across hundreds of placental mammals. Science. 2023;380:eabn3943. ENCODE Project Consortium. An integrated encyclopedia of DNA elements in the human genome. Nature. 2012;489:57–74. Chadwick LH. The NIH Roadmap Epigenomics Program data resource. Epigenomics. 2012;4:317–24. Baux D, et al. MobiDetails: online DNA variants interpretation. Eur J Hum Genet EJHG. 2021;29:356–60. Dylan, Aïssi et al. MORFEE: a new tool for detecting and annotating single nucleotide variants creating premature ATG codons from VCF files. bioRxiv 2020.03.29.012054 (2020) 10.1101/2020.03.29.012054 Meguerditchian C, et al. Enhancing the annotation of small ORF-altering variants using MORFEE: introducing MORFEEdb, a comprehensive catalog of SNVs affecting upstream ORFs in human 5’UTRs. NAR Genomics Bioinforma. 2025;7:lqaf017. Stenton SL, et al. Critical assessment of variant prioritization methods for rare disease diagnosis within the rare genomes project. Hum Genomics. 2024;18:44. Tenywa J-F, et al. Genome region aware CADD thresholds for noncoding variant prioritization. NAR Genomics Bioinforma. 2025;7:lqaf157. Additional Declarations No competing interests reported. Supplementary Files Supplementarydata.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8823759","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":603487861,"identity":"ab577a3f-b70e-49c3-84b8-b94f3e3f0709","order_by":0,"name":"Abdelhakim Bouazzaoui","email":"","orcid":"","institution":"Centre Hospitalier Universitaire de Rennes, ERN ITHACA, Hôpital Sud","correspondingAuthor":false,"prefix":"","firstName":"Abdelhakim","middleName":"","lastName":"Bouazzaoui","suffix":""},{"id":603487862,"identity":"7f815d12-611e-4e6b-b283-0313e838e2d0","order_by":1,"name":"Jean-Madeleine de Sainte Agathe","email":"","orcid":"","institution":"Department of Genetics, Assistance Publique-Hôpitaux de Paris, Sorbonne Université, Pitié-Salpêtrière University Hospital, Paris, France","correspondingAuthor":false,"prefix":"","firstName":"Jean-Madeleine","middleName":"de Sainte","lastName":"Agathe","suffix":""},{"id":603487863,"identity":"116a1ea8-6b1b-43b6-8109-14413f926b73","order_by":2,"name":"Simon Cabello-Aguilar","email":"","orcid":"","institution":"PMMG : Plateau de Médecine Moléculaire et de Génomique, Montpellier BioInformatique pour le Diagnostic Clinique (MoBiDiC), CHU Montpellier, F-34000 Montpellier, France.","correspondingAuthor":false,"prefix":"","firstName":"Simon","middleName":"","lastName":"Cabello-Aguilar","suffix":""},{"id":603487864,"identity":"ce954b39-4efe-47b0-abc9-a200c43f6fa6","order_by":3,"name":"Ophélie Evrard","email":"","orcid":"","institution":"Laboratoire de Génétique Constitutionnelle, CHU Amiens Picardie, Amiens, France. ; Université Picardie Jules Verne, HEMATIM UR4666, Amiens, France","correspondingAuthor":false,"prefix":"","firstName":"Ophélie","middleName":"","lastName":"Evrard","suffix":""},{"id":603487865,"identity":"41aac081-fe8e-43da-99c7-b87f317c2c12","order_by":4,"name":"Juliette Nectoux","email":"","orcid":"","institution":"Service de Médecine Génomique des Maladies de Système et d'Organe, Fédération de Génétique et de Médecine Génomique, Centre - Université Paris Cité, Hôpital Cochin, APHP, Paris, France.","correspondingAuthor":false,"prefix":"","firstName":"Juliette","middleName":"","lastName":"Nectoux","suffix":""},{"id":603487866,"identity":"c7c698c3-d369-4b21-9f08-0ad4b426b841","order_by":5,"name":"Marina Konyukh","email":"","orcid":"","institution":"Département de Génétique Médicale, Hôpital Henri Mondor, Assistance Publique–Hôpitaux de Paris, Créteil, France ; Laboratoire de biologie médicale multisites SeqOIA – PFMG2025, Paris, France.","correspondingAuthor":false,"prefix":"","firstName":"Marina","middleName":"","lastName":"Konyukh","suffix":""},{"id":603487867,"identity":"e6f4e85f-42df-46b8-9be0-6ba0ae49ba14","order_by":6,"name":"Leila Qebibo","email":"","orcid":"","institution":"Pediatric Neurogenetics Laboratory, Department of Genetics, Reference Center for Cerebellar Malformations and Congenital Diseases, Armand-Trousseau Hospital, AP-HP","correspondingAuthor":false,"prefix":"","firstName":"Leila","middleName":"","lastName":"Qebibo","suffix":""},{"id":603487868,"identity":"2be85ba6-d1d1-4b38-94de-3285c7ccc9c6","order_by":7,"name":"Thibault Coste","email":"","orcid":"","institution":"Département de Génétique, Hôpital Bichat Claude-Bernard, AP-HP, Paris, France. ; Institut du Cerveau (ICM), INSERM UMR1127, Université Sorbonne, Paris, France","correspondingAuthor":false,"prefix":"","firstName":"Thibault","middleName":"","lastName":"Coste","suffix":""},{"id":603487869,"identity":"0946cfc2-304e-41a4-a0bc-bbd40f9343e2","order_by":8,"name":"Sandrine M. Caputo","email":"","orcid":"","institution":"Department of Genetics, Institut Curie, Paris, France and PSL Research University, Paris, France.","correspondingAuthor":false,"prefix":"","firstName":"Sandrine","middleName":"M.","lastName":"Caputo","suffix":""},{"id":603487870,"identity":"dfafa194-c7ca-4882-8a88-1fb925830b9c","order_by":9,"name":"Perrine Brunelle","email":"","orcid":"","institution":"CHU Lille, Univ. Lille, Institut de Génétique Médicale, ULR7364 RADEME, F-59000 Lille, France","correspondingAuthor":false,"prefix":"","firstName":"Perrine","middleName":"","lastName":"Brunelle","suffix":""},{"id":603487871,"identity":"923b99b7-7eaf-452e-be6c-027098fdbdf2","order_by":10,"name":"Yohann Jourdy","email":"","orcid":"","institution":"Service d'hématologie biologique, Hospices Civils de Lyon, Groupe Hospitalier Est, Bron, France","correspondingAuthor":false,"prefix":"","firstName":"Yohann","middleName":"","lastName":"Jourdy","suffix":""},{"id":603487872,"identity":"d5877eb1-7cc5-4316-a150-0e37143216a7","order_by":11,"name":"Cécile Rouzier","email":"","orcid":"","institution":"Medical Genetic Department, Archet University Hospital, Université Côte d'Azur, Nice, France; Inserm U1081, CNRS UMR7284, IRCAN, Université Côte d'Azur, Nice, France","correspondingAuthor":false,"prefix":"","firstName":"Cécile","middleName":"","lastName":"Rouzier","suffix":""},{"id":603487873,"identity":"550ca58e-d530-4d7c-9475-433f89bcc3d9","order_by":12,"name":"Mireille Cossée","email":"","orcid":"","institution":"PMMG : Plateau de Médecine Moléculaire et de Génomique, Laboratoire de Génétique Moléculaire des Maladies Rares, Université de Montpellier, CHU Montpellier, F-34000 Montpellier, France.","correspondingAuthor":false,"prefix":"","firstName":"Mireille","middleName":"","lastName":"Cossée","suffix":""},{"id":603487874,"identity":"71af99e9-c780-4875-b837-11d4ca58ec3f","order_by":13,"name":"Charles Van Goethem","email":"","orcid":"","institution":"PMMG : Plateau de Médecine Moléculaire et de Génomique, Laboratoire de Génétique Moléculaire des Maladies Rares, Université de Montpellier, CHU Montpellier, F-34000 Montpellier, France.","correspondingAuthor":false,"prefix":"","firstName":"Charles","middleName":"Van","lastName":"Goethem","suffix":""},{"id":603487875,"identity":"a2f44c9b-45cc-44e6-bce7-a7609fec9997","order_by":14,"name":"Olivier Ardouin","email":"","orcid":"","institution":"PMMG : Plateau de Médecine Moléculaire et de Génomique, Laboratoire de Génétique Moléculaire des Maladies Rares, Université de Montpellier, CHU Montpellier, F-34000 Montpellier, France.","correspondingAuthor":false,"prefix":"","firstName":"Olivier","middleName":"","lastName":"Ardouin","suffix":""},{"id":603487876,"identity":"ec3f802a-518e-4a4f-ba0f-576a3ab71637","order_by":15,"name":"Vasiliki Kalatzis","email":"","orcid":"","institution":"Institute for Neurosciences of Montpellier (INM), Université de Montpellier, Inserm, 34091 Montpellier, France","correspondingAuthor":false,"prefix":"","firstName":"Vasiliki","middleName":"","lastName":"Kalatzis","suffix":""},{"id":603487877,"identity":"ac31f00e-a668-4a49-bb38-0cd791a84fcd","order_by":16,"name":"Anne-Françoise Roux","email":"","orcid":"","institution":"PMMG : Plateau de Médecine Moléculaire et de Génomique, Laboratoire de Génétique Moléculaire des Maladies Rares, Université de Montpellier, CHU Montpellier, F-34000 Montpellier, France.","correspondingAuthor":false,"prefix":"","firstName":"Anne-Françoise","middleName":"","lastName":"Roux","suffix":""},{"id":603487878,"identity":"d4a44a87-a4fe-47e4-a909-07ea91f270bc","order_by":17,"name":"David Baux","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABBElEQVRIiWNgGAWjYDCCAzwg0obBAC7CztxwAEjxE9CSBtMCpJgZwVokG/BrOYyqhQGfFr4DvAc/87adlzeXSGDd8HHHnzx+ZsbGAx/3MEiY49AjeYAvWZrnzG3DnTMS2G7OPGNQLNnM2HBwxjMGCZkD2LUY3H9jIM1TcTvB4Eb+t9u8bQaJGw4zNhzmOcBQJ4HDYQYHeIx/8xicA2pJYANr2Q/S8ucAgwQeLWZAWw4gtGwAev8wAx4tQL+kWc45k2y44cwDoF/ajIslgLYc7DkggVMLMMQO33jbZidvcDyB7cbHNrk8/vbmwx9+HLDBqQUDJEBpojUgtIyCUTAKRsEogAMAenddkL48mM4AAAAASUVORK5CYII=","orcid":"","institution":"PMMG : Plateau de Médecine Moléculaire et de Génomique, Laboratoire de Génétique Moléculaire des Maladies Rares, Université de Montpellier, CHU Montpellier, F-34000 Montpellier, France.","correspondingAuthor":true,"prefix":"","firstName":"David","middleName":"","lastName":"Baux","suffix":""}],"badges":[],"createdAt":"2026-02-08 19:08:52","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8823759/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8823759/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":104443265,"identity":"aabd49fd-61c8-456e-b361-9f403c466d88","added_by":"auto","created_at":"2026-03-11 19:06:49","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":547121,"visible":true,"origin":"","legend":"\u003cp\u003ePerformance of Variant Effect Predictors\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8823759/v1/e01504523826b76ab824ebd5.png"},{"id":104443264,"identity":"ab85011b-cba4-4fe7-bee4-5cadf83c5cd3","added_by":"auto","created_at":"2026-03-11 19:06:49","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":243882,"visible":true,"origin":"","legend":"\u003cp\u003eMobideep performance against other VEPs on independent dataset and region-specific thresholds\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8823759/v1/f25d66d83456ef04388685e4.png"},{"id":105562372,"identity":"0e8baaa3-91c2-49b7-a9b4-e6432f886c63","added_by":"auto","created_at":"2026-03-27 12:29:25","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":675256,"visible":true,"origin":"","legend":"\u003cp\u003eMobiDeep's Performance in Large-Scale Ranking Simulations\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8823759/v1/cdef091b501190fba3fb4451.png"},{"id":107487652,"identity":"004e8fea-2535-444a-9210-c35538c7b094","added_by":"auto","created_at":"2026-04-22 02:42:31","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1753447,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8823759/v1/37acc6a7-7f3c-4a1d-925f-6c0c61819588.pdf"},{"id":104780351,"identity":"5254caf3-b5f0-4908-8cbc-942fb241076b","added_by":"auto","created_at":"2026-03-17 07:52:27","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":1311274,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementarydata.docx","url":"https://assets-eu.researchsquare.com/files/rs-8823759/v1/cdf9854316309688732d7644.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"MobiDeep: an AI-based meta-score for scoring non-coding DNA variations","fulltext":[{"header":"Background","content":"\u003cp\u003eWhile Genome Sequencing (GS) is now routinely used in rare disease molecular diagnostics, the yield remains limited, often below 50% of cases\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e. This persistent diagnostic gap suggests that a substantial proportion of disease-causing variants resides within non-coding regions (NCRs). These regions, comprising 98% of the human genome, remain difficult to interpret in clinical practice\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e. Consequently, major genomic medicine initiatives, including the French Plan for Genomic Medicine 2025\u003csup\u003e3\u003c/sup\u003e and the Genomics England 100k Genomes Project\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e, have identified the comprehensive interpretation of Non-Coding Variants (NCVs) in GS data as a critical objective for providing the best care and realizing the full potential of genomic medicine.\u003c/p\u003e \u003cp\u003eNCRs harbour a diverse array of regulatory elements, such as promoters, enhancers, and untranslated regions (UTRs), that orchestrate spatiotemporal gene expression. The disruption of these elements is an established, albeit under-characterized, mechanism of Mendelian disease. For instance, pathogenic variants in the 5'UTR of \u003cem\u003eGATA4\u003c/em\u003e are associated with atrial septal defects\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e, while distant-acting enhancer variants affecting \u003cem\u003eSOX9\u003c/em\u003e expression cause Pierre Robin sequence\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e. Despite such examples, regulatory variants are significantly underrepresented in clinical databases. The Human Gene Mutation Database\u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e (HGMD; retrieved 03/10/2025) catalogues fewer than 320 regulatory disease-causing variants, while ClinVar\u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e (retrieved 09/02/2025) contains only several hundred pathogenic variants annotated within key regulatory elements, a stark contrast to the tens of thousands of well characterized protein-coding variants (64,286). This disparity stems not from a lack of biological relevance, but from fundamental challenges in NCVs interpretation. In contrast to the relatively straightforward application of the universal genetic code in coding regions, regulatory elements operate through complex, context-dependent mechanisms that are difficult to predict computationally.\u003c/p\u003e \u003cp\u003eTo address this interpretive challenge, several computational variant effect predictors (VEPs) have been developed. Despite their utility, the clinical application of these tools faces several limitations. First, their performance varies substantially across different genomic contexts and is often inadequately benchmarked outside of canonical regulatory elements\u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e. Second, tools like ReMM that are trained on known disease variants risk data circularity and inflated performance estimates when evaluated on similar datasets. Finally, the assumption of a universal score threshold is untenable, as recent work highlights the necessity of region-specific calibration for accurate variant classification\u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eIn this study, we address these challenges through a multi-faceted approach. First, to provide a practical guide for their clinical use and directly address the need for context-specific thresholds recently highlighted by Villani et al\u003csup\u003e9\u003c/sup\u003e, we conducted a comprehensive and rigorous benchmark on multiple VEPs using a curated dataset of 448 pathogenic and a total 38,146 neutral NCVs. We specifically excluded from the pathogenic datasets variants known or predicted to alter splicing, as the aim of the study was to focus on the other pathological mechanisms, considering the splicing prediction problem has already been adequately addressed\u003csup\u003e\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e,\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e,\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e. We established the VEPs performance and optimal classification thresholds not only globally but also within distinct functional genomic regions (e.g., 5\u0026rsquo;UTR, 3\u0026rsquo;UTR, upstream regions). This granular analysis, revealing the context-dependent strengths and weaknesses of each predictor, motivated the development of MobiDeep: a multilayer perceptron meta-score designed to integrate these five tools into a single, calibrated pathogenicity score that achieves superior performance. Furthermore, recognizing that clinical utility extends beyond simple classification, we implemented a ranking simulation framework to evaluate MobiDeep's performance in prioritizing causal variants within large, clinically realistic backgrounds of benign variations. Our aim is to provide not only a new computational tool but also a transparent framework for its responsible implementation, adhering to community-driven best practices for reproducibility and clinical utility\u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eGlobal and stratified benchmarking reveals variable performance of existing VEPs\u003c/p\u003e \u003cp\u003eWe were able to obtain the scores for all 448 pathogenic variants for the 3 ReMM, phyloP_Primates and Cactus241way VEPs, however, as this study is based on pre-computed datasets, we only scored 400/448 variants for GPN-MSA and 416/448 for CADD v1.7. In the global analysis, which combined all non-coding regions, the integrative models clearly outperformed standalone conservation scores (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eA, \u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eB; Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, Supplementary Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e). ReMM v0.4 demonstrated the highest overall discriminative ability, achieving a mean AUPRC of 0.814 and an AUROC of 0.953 (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e), these results were supported by strong F1-score of 0.660 and the highest MCC at 0.623. This was followed by CADD v1.7 (AUPRC\u0026thinsp;=\u0026thinsp;0.744, AUROC\u0026thinsp;=\u0026thinsp;0.913) and GPN-MSA (AUPRC\u0026thinsp;=\u0026thinsp;0.735, AUROC\u0026thinsp;=\u0026thinsp;0.921), which also performed robustly. The conservation-based tools, Cactus241way (AUPRC\u0026thinsp;=\u0026thinsp;0.686) and phyloP_Primates (AUPRC\u0026thinsp;=\u0026thinsp;0.615), provided a solid baseline but showed lower overall efficacy.\u003c/p\u003e \u003cp\u003eThe mean optimal global thresholds were found to be 0.800 \u0026plusmn; (0.029) for ReMM, 10.373 \u0026plusmn; (0.184) for CADD, -2.549 \u0026plusmn; (0.312) for GPN-MSA, 0.304 \u0026plusmn; (0.081) for PhyloP_Primates and 1.345 \u0026plusmn; (0.059) for Cactus241way Vertebrates. However, score distributions for all predictors show significant overlap between pathogenic and benign classes (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eC-G), illustrating the challenge of relying on a single threshold for perfect classification. We note that while our data-driven thresholds for GPN-MSA (-2.3) and ReMM (0.81) are optimal, more stringent thresholds of -7.0 and 0.961, respectively, are commonly used to isolate only high-confidence pathogenic variants for PP3 criteria, a stringency supported by our score distribution plot (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eE).\u003c/p\u003e \u003cp\u003eA context-dependent analysis revealed significant performance heterogeneity across five harmonized genomic regions, indicating that a tool's global performance does not guarantee its utility in a specific locus (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, Supplementary Figure \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe specialized nature of ReMM was evident in its superior performance in Intronic (AUROC\u0026thinsp;=\u0026thinsp;0.942) and Non-coding exon (AUROC\u0026thinsp;=\u0026thinsp;0.987) regions, where it most effectively separated pathogenic from benign variants (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e and Supplementary Figure \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eGPN-MSA showed its key strength in the 3' UTR region (AUROC\u0026thinsp;=\u0026thinsp;0.901, AUPRC\u0026thinsp;=\u0026thinsp;0.604), outperforming all other tools in this context (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). This makes GPN-MSA the tool of choice for variants located within 3\u0026rsquo;UTR regions but may necessitate using the 3\u0026rsquo;UTR specific threshold at -3.59 \u0026plusmn; (0.465) (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eE, right).\u003c/p\u003e \u003cp\u003eCADD v1.7 demonstrated the most consistency and balance across the genomic regions. While it excelled in Non-coding exon regions with an AUROC of 0.959 and an AUPRC of 0.819, its efficacy was markedly reduced in the 5' UTR (AUROC\u0026thinsp;=\u0026thinsp;0.783, AUPRC\u0026thinsp;=\u0026thinsp;0.641), where its score distributions for pathogenic and benign variants showed considerable overlap (Supplementary Figure \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe performance of pure conservation scores was also context-dependent; for instance, PhyloP_Primates, which was respectable globally (AUROC\u0026thinsp;=\u0026thinsp;0.865), showed sharply reduced efficacy in the 5' UTR (AUROC\u0026thinsp;=\u0026thinsp;0.738, AUPRC\u0026thinsp;=\u0026thinsp;0.566), suggesting its evolutionary signal is less informative for variants in this specific region.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eClassification Performance of Predictors at Optimal Thresholds in Global and Regional Analyses.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"8\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAnalysis\u003c/p\u003e \u003cp\u003eGroup\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTool\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAUROC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAUPRC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eF1-Score\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eTP (FN)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eFP (TN)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGlobal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReMM v0.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.953\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.814\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.660\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e371 (77)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e310 (4170)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eGPN-MSA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.921\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.735\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.548\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e318 (82)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e448 (4032)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCADD v1.7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.913\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.744\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.648\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e335 (81)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e283 (4197)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCactus241way_V\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.863\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.686\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.617\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e322 (126)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u003cb\u003e275 (4196)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePhyloP_P\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.865\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.615\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.504\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e337 (111)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e558 (3913)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e3' UTR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eGPN-MSA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.901\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.604\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.572\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e32 (10)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e38 (422)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReMM v0.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.893\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.599\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.568\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e33 (13)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e38 (422)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCADD v1.7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.854\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.654\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.596\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e31 (11)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e31 (429)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCactus241way_V\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.847\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.628\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.674\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e32 (14)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u003cb\u003e17 (443)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePhyloP_P\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.794\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.394\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.363\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e30 (16)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e94 (366)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e5' UTR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReMM v0.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.853\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.747\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.667\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e91 (38)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e54 (342)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eGPN-MSA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.800\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.646\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.566\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e61 (51)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u003cb\u003e44 (352)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCADD v1.7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.783\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.641\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.564\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e80 (36)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e88 (308)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCactus241way_V\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.746\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.623\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.553\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e69 (60)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e53 (343)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePhyloP_P\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.738\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.566\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.538\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e79 (50)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e87 (309)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIntron\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReMM v0.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.942\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.814\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.702\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e67 (17)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e40 (800)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eGPN-MSA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.922\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.804\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.711\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e65 (13)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e40 (800)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCADD v1.7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.872\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.721\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.655\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e64 (19)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e49 (791)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePhyloP_P\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.872\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.681\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.551\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e66 (18)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e89 (748)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCactus241way_V\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.826\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.704\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.659\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e60 (24)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u003cb\u003e38 (799)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNon_coding\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReMM v0.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.987\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.931\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.784\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e69 (4)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u003cb\u003e35 (695)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eexon\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCADD v1.7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.959\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.819\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.658\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e59 (7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e55 (675)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eGPN-MSA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.952\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.747\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.504\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e59 (5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e116 (614)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCactus241way_V\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.919\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.704\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.632\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e63 (10)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e64 (666)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePhyloP_P\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.900\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.579\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.457\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e63 (10)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e141 (589)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUpstream\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReMM v0.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.959\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.806\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.639\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e98 (11)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e101 (989)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eGPN-MSA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.938\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.803\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.586\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e85 (14)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e108 (982)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCADD v1.7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.936\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.797\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.719\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e90 (14)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u003cb\u003e57 (1033)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCactus241way_V\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.891\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.733\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.598\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e90 (19)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e102 (986)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePhyloP_P\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.892\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.689\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.615\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e86 (23)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e86 (1001)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eAUROC: Area Under the Receiver Operating Characteristic curve. AUPRC: Area Under the Precision-Recall curve. F1-Score: The harmonic mean of precision and recall. TP (FN): number of true positives (and false negatives) out of the total pathogenic variants for that group. FP (TN): number of false positives (and true negatives) out of the total sampled benign variants for that group. Figures are the average of the 10,000 iterations performed. VEPs are ordered per AUROC performance for each analysed group. Bold: best performing VEP for the considered metric.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eGlobal performance (left panel): Mean Receiver Operating Characteristic (ROC) (A) and Area Under the ROC Curve (AUROC) (B), curves, with shaded areas representing\u0026thinsp;\u0026plusmn;\u0026thinsp;1 standard deviation. Area under ROC and PR are reported. Thresholds (left) and Region-specific analyses (right panels): Score distribution and region-specific performance for ReMM v.04 (C), CADD 1.7 (D), GPN-MSA (E), Cactus241way_v (F), and PhyloP_P (G). The distributions for pathogenic (red) and benign (blue) variants are shown. The vertical dashed line indicates the mean optimal classification threshold determined by the Youden's J Index, providing a practical reference for each tool's decision boundary when used globally. Performance was evaluated on a subset of variants located in each genomic region (5\u0026rsquo;UTR, 3\u0026rsquo;UTR, Intronic, Non-coding exon and upstream) illustrating the performance variability of tools in specific genomic contexts (violin plots).\u003c/p\u003e \u003cp\u003eMobiDeep: An ensemble learning based meta-score\u003c/p\u003e \u003cp\u003eThe final MobiDeep model is a three-hidden-layer Multilayer Perceptron (MLP) with a [115, 144, 175] neuron architecture and ReLU activation functions (Supplementary Figure S2). A permutation-based feature importance analysis revealed that the integrative models, ReMM and CADD, are the most influential features driving MobiDeep's predictions. GPN-MSA provided the next largest contribution, while the conservation scores served as complementary evidence, confirming that MobiDeep effectively weighs and synthesizes information from all five input VEPs (Supplementary Figure S3D).\u003c/p\u003e \u003cp\u003eMobiDeep achieves state-of-the-art classification performance\u003c/p\u003e \u003cp\u003eWhen evaluated on the independent benchmark dataset, MobiDeep demonstrated superior classification accuracy, outperforming all individual VEPs in both AUROC and AUPRC, all metrics are detailed in Supplementary Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e. MobiDeep achieved an AUROC of 0.974 (\u0026plusmn;\u0026thinsp;0.001) and an AUPRC of 0.889 (\u0026plusmn;\u0026thinsp;0.011) (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eA-B, Supplementary Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e). This improvement highlights the benefit of integrating multiple informative features within an optimized, non-linear framework.\u003c/p\u003e \u003cp\u003eGlobal, region-specific and clinically relevant thresholds\u003c/p\u003e \u003cp\u003eAn optimal global threshold of 0.60 was determined by maximizing the Youden's J Index, as shown by the score distributions in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eC. Our analysis also yielded the following region-specific thresholds: 0.92 for 5\u0026rsquo;UTR, 0.60 for 3\u0026rsquo;UTR, 0.40 for Intronic, 0.85 for Non-coding exonic and 0.61 for Upstream regions., highlighting the performance adaptability across genomic contexts (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eD).\u003c/p\u003e \u003cp\u003eA second generic threshold has been established, corresponding to a 95% Positive Predictive Value (95% CI: 0.941\u0026ndash;0.983), defined as a MobiDeep raw probability score of 0.968. This corresponds to a log-score of \u0026gt;\u0026thinsp;1.5 and captures approximately 55% of true pathogenic variants in our independent test set.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eComparative (A) Receiver Operating Characteristic (ROC) and (B) Precision-Recall (PR) curves for MobiDeep and the five other predictors achieving a state-of-the-art AUROC of 0.973 and an AUPRC of 0.888. (C) score distributions for benign and pathogenic variants confirm the model's high discriminative power with a mean Youden threshold of 0.6 in global (all regions). (D) Youden thresholds for global (0.6) and harmonized regions highlights the specific thresholding needed for each genomic context with specific thresholds of 0.6, 0.92, 0.4, 0.85 and 0.61 for 3\u0026rsquo;UTR, 5\u0026rsquo;UTR, Intronic, Non-coding exonic and Upstream regions respectively.\u003c/p\u003e \u003cp\u003eMobiDeep robustly prioritizes pathogenic variants in large-scale ranking simulations\u003c/p\u003e \u003cp\u003eThe primary test of a VEP's clinical utility is its ability to prioritize causal variants within the large lists generated by GS. MobiDeep demonstrated highly scalable performance in these simulations (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eSummary of MobiDeep's absolute ranking performance for 137 independent VOIs against increasing noise backgrounds (5 replicates).\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"9\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBackground (N)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMedian Rank\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eMean Rank\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eTop 1%\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eTop 1\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eTop 5\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eTop 10\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c9\"\u003e \u003cp\u003eTop 20\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1,000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e17.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e88.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e54.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e72.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e80.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e91.4%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e10,000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e5.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e171.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e89.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e21.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e62.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e71.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e74.9%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e100,000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e45.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1,697.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e89.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e5.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e15.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e24.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e34.2%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eWhen ranked against a background of 1,000 neutral variants, the median rank for a pathogenic VOI was just 1 (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). The prioritization performance was maintained even as the search space expanded: the median rank increased to only 5 against 10,000 variants and 45 against 100,000 variants. The model's ability to enrich for causal variants is further highlighted by its success in placing them at the absolute top of the ranked list, a critical feature for manual review in a clinical setting. Across all 5 replicates, 72.3% of VOIs achieved a median rank in the Top 5 when tested against 1,000 benign variants. This success rate remained high at 62.0% against a 10,000-variant background and 40.1% against a 100,000-variant background.\u003c/p\u003e \u003cp\u003eThe model's scalability is illustrated in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e and Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eA. The median rank increases sub-linearly with the background size, demonstrating that the model's performance remains robust even as the number of candidate variants grows substantially. Furthermore, when viewed as a percentage of the total pool, the model's enrichment power is remarkably stable. Nearly 90% of pathogenic VOIs were consistently ranked within the top 1% of their respective pools, a figure that remained nearly constant whether the pool contained 1,000 or 100,000 variants. The cumulative distribution of ranks (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eA) illustrates this scalability, with the curves showing only modest rightward shifts as background complexity increases.\u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eB further emphasizes this prioritization power by displaying the cumulative percentage of VOIs by rank for N\u0026thinsp;=\u0026thinsp;10,000. The steep initial rise of this curve demonstrates that MobiDeep efficiently concentrates pathogenic variants at the very top of ranked lists\u0026mdash;approximately 62% of VOIs achieve a rank within the top 5 positions, and over 70% are ranked within the top 10.\u003c/p\u003e \u003cp\u003eAnalysis of individual variants across all replicates for N\u0026thinsp;=\u0026thinsp;10,000 confirms this overall consistency. The vast majority of VOIs are consistently ranked highly, with only a small subset of variants ranking poorly (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eC). Notably, specific variants such as NM_000044.6:c.-547C\u0026thinsp;\u0026gt;\u0026thinsp;T, which disrupts an upstream open reading frame (uORF), consistently rank poorly across all simulation conditions. This demonstrates that while certain pathogenic mechanisms, particularly uORF disruption in 5' UTRs, remain challenging for the model, MobiDeep is nonetheless a robust and reliable tool for drastically reducing the search space for the vast majority of NCVs in a clinical setting.\u003c/p\u003e \u003cp\u003eAn analysis of rank distribution stratified by VEP-annotated variant effect provides additional mechanistic insight (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eD). While most variants in effect categories such as introns and upstream regions were ranked very highly (median rank near 1), variants in the 5_prime_UTR_variant category showed a wider distribution with a notable tail extending to higher ranks. This finding corroborates our earlier observation that 5' UTR variants are more challenging to prioritize and suggests that specific regulatory mechanisms active in this region\u0026mdash;such as uORF disruption, ribosome binding site alterations, or other translational regulatory elements\u0026mdash;are not fully captured by the current feature set. This represents a clear direction for future model enhancement, potentially through integration of specialized predictors like MORFEE\u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e,\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e that explicitly model uORF effects.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e(A) Cumulative distribution function (CDF) of ranks for 137 pathogenic VOIs against increasing backgrounds (N\u0026thinsp;=\u0026thinsp;1,000, 10,000, and 100,000 neutral variants). Over 88% of VOIs ranked within the top 1% across all background sizes. X-axis: log scale. (B) Cumulative percentage of VOIs by rank (N\u0026thinsp;=\u0026thinsp;10,000 background). 89.0% of VOIs ranked within top 1,000; 62.7% within top 10. X-axis: log scale. (C) Individual rank trajectories for 137 VOIs (5 replicates, N\u0026thinsp;=\u0026thinsp;10,000 background). Most VOIs (X-axis) show consistent high ranking (low rank numbers). Outliers include 5' UTR variants with uORF disruption. Y-axis: log scale. (D) VOI rank distributions by VEP-annotated effect (N\u0026thinsp;=\u0026thinsp;10,000 background). Most categories show tight distributions near top ranks. The 5_prime_UTR_variant category displays wider distribution with poorer-ranked outliers, reflecting limited capture of uORF-related mechanisms.\u003c/p\u003e \u003cp\u003eComplementary test on pathogenic RNU4-2 variants\u003c/p\u003e \u003cp\u003eAs a final validation step, we scored a set of 30 variants in the \u003cem\u003eRNU4-2\u003c/em\u003e non-coding RNA retrieved from Clinvar VCF at 2025/05/30, which were not present in our first extraction of ClinVar (Supplementary Table S3). The model performed well. Of the 19 variants in this set classified as P/LP in ClinVar, MobiDeep assigned a score exceeding our high-confidence 95% PPV threshold (0.968, log_score\u0026thinsp;\u0026gt;\u0026thinsp;1.5) and Non-coding Exon threshold (0.85) for all the 19 variants. In contrast, all three \"Likely benign\" variants were correctly assigned scores below these specific pathogenicity thresholds. Remaining variants were annotated in Clinvar as of unknown clinical significance, and most of them were scored pathogenic by MobiDeep.\u003c/p\u003e"},{"header":"Discussion and Conclusions","content":" \u003cp\u003eThe interpretation of NCVs remains a primary bottleneck, limiting the full diagnostic potential of GS. Our study systematically addresses this challenge by first rigorously benchmarking state-of-the-art VEPs and then developing MobiDeep, a meta-score that demonstrates superior performance in NCV prioritization. This work responds directly to the challenges highlighted by community-wide efforts like the CAGI challenge, which reveal the need for more accurate and context-aware tools for variant prioritization in rare disease diagnostics\u003csup\u003e\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eOur initial benchmarking of existing VEPs revealed two critical findings. First, no single tool is universally optimal; performance is highly context dependent. While integrative tools like ReMM and CADD showed strong global performance, stratified analysis uncovered specific strengths and weaknesses. For instance, ReMM demonstrated clear dominance in intronic (AUROC\u0026thinsp;=\u0026thinsp;0.942) and non-coding exon regions (AUROC\u0026thinsp;=\u0026thinsp;0.987), whereas GPN-MSA showed superiority in 3' UTRs (AUROC\u0026thinsp;=\u0026thinsp;0.901) (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). Conversely, the performance of all predictors was notably diminished in 5' UTRs, suggesting that key pathogenic mechanisms in this region, such as uORF disruption, may not be adequately captured by these models. Second, our derivation of optimal thresholds confirms the necessity of region-specific calibration. Our optimal global CADD 1.7 threshold of 10.37 closely replicates the CADD\u0026thinsp;\u0026ge;\u0026thinsp;10 threshold proposed by Villani et al.\u003csup\u003e9\u003c/sup\u003e for assigning moderate evidence of pathogenicity (PP3) to regulatory variants. Similarly, our optimal ReMM threshold of 0.80 aligns with their REMM\u0026thinsp;\u0026ge;\u0026thinsp;0.86 threshold. Recent work by Tenywa et al.\u003csup\u003e27\u003c/sup\u003e provides further independent validation of this region-specific approach, deriving CADD v1.7 thresholds for distinct noncoding regions using ClinVar data. Our independently derived region-specific CADD thresholds (Supplementary Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e, Supplementary Figure \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e) show notable concordance with their findings: both studies identify 5'UTR regions as requiring substantially higher thresholds (17.37 in our dataset vs. 16.79 in theirs), while thresholds for 3'UTR and non-coding exonic regions cluster in the 11\u0026ndash;13 range. The substantial threshold variation observed across genomic contexts\u0026mdash;ranging from 8.39 for intronic regions to 17.37 for 5'UTRs in our analysis, and from 11.08 to 25.5 (for splicing variants, excluded in our study) across different noncoding categories in Tenywa et al\u003csup\u003e27\u003c/sup\u003e. This demonstrates, empirically, the inadequacy of applying a single genome-wide cutoff. This convergence of independently derived, region-specific thresholds across multiple independent studies confirms that missense-variant thresholds are not applicable to NCVs and establishes genomic context-aware calibration as an essential principle for accurate noncoding variant interpretation. (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe significant overlap in score distributions between pathogenic and benign variants (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eC-G), even for the best-performing VEPs, illustrates the inherent difficulty in achieving perfect classification. This presents a major clinical challenge. In a diagnostic setting, a stringent threshold chosen to minimize false positives (i.e., increase specificity) will inevitably increase the rate of false negatives, causing true disease-causing variants to be missed. The clinical and personal cost of a missed diagnosis is high, therefore, the primary goal for a clinically useful tool must be to manage this trade-off, with a strong emphasis on maximizing sensitivity to avoid discarding potential causal candidates.\u003c/p\u003e \u003cp\u003eThe fact that no single VEP excels across all genomic regions argues strongly for an integrative strategy. We hypothesized that by combining the distinct, and at times complementary, predictive signals from different tools, a meta-score, could achieve a more robust and accurate classification. Our development of MobiDeep was a direct test of this hypothesis. This approach proved effective, as MobiDeep outperforms each individual predictor on our independent test set (AUROC\u0026thinsp;=\u0026thinsp;0.973, AUPRC\u0026thinsp;=\u0026thinsp;0.888; Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eA-B). The feature importance analysis (Supplementary Figure S3D) provides further insight: while it confirms the strong predictive power of integrative tools like ReMM and CADD, it also shows that conservation-based scores, though weaker on their own, still contribute meaningfully to the final prediction. This implies that the most accurate classification arises from a synergistic combination of all features, which MobiDeep is designed to leverage. The model was carefully optimized to avoid overfitting (Supplementary Figure S3), ensuring that this performance reflects a genuine ability to generalize.\u003c/p\u003e \u003cp\u003eThe clinical utility of a VEP is ultimately determined by its performance in prioritizing variants from large candidate lists. In the simulation results, the low median rank, which scales sub-linearly with the background size (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eA), suggests that the typical pathogenic variant will be placed near the top of a ranked list. However, the stark divergence between the median and the much higher mean rank (e.g., 45.0 vs. 1,697.4 at N\u0026thinsp;=\u0026thinsp;100,000) points to a significant issue: a tail of pathogenic variants that are ranked very poorly. This suggests that while MobiDeep is reliable for the majority of variants, it can fail substantially in a minority of cases.\u003c/p\u003e \u003cp\u003eThis trade-off is further highlighted by the model's performance on the most stringent success criterion: placing the causal variant at rank 1. This metric diminished from 54.7% to 5.4% as the background set grew from 1,000 to 100,000 variants. Therefore, while MobiDeep consistently places nearly 90% of pathogenic variants in the top 1% of the list, clinical biologists can use the top-ranked candidate as a strong hypothesis requiring verification. The tool is best understood as a method for generating a highly enriched shortlist (e.g., the top 20\u0026ndash;50 candidates) that facilitate focused manual review. The outlier analysis in our results (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eC, \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eD) directly links the poorly ranked variants to specific biological contexts, such as 5' UTRs, underscoring that these performance characteristics reflect specific biological contexts (e.g. 5\u0026rsquo;UTRs) that benefit from complementary approaches. Importantly, our simulations represent a deliberately challenging 'filter-free' scenario, that does not incorporate the standard filtering steps used in clinical practice (e.g. restricting analysis to a gene panel, filtering by allele frequency, or leveraging patient phenotype terms).\u003c/p\u003e \u003cp\u003eThis context is essential for understanding how MobiDeep is designed to be implemented: not as a replacement for the established diagnostic pipeline, but as a flexible and powerful component within it. Its utility is best realized through a dual-threshold strategy tailored to different stages of the workflow. For initial, high-throughput screening where maximizing sensitivity is paramount, the global threshold of 0.60 (or the even more granular region-specific thresholds) can be applied to ensure potential candidates are not prematurely discarded. Conversely, for the final step of prioritizing pre-filtered list of challenging VUS, the high-confidence threshold (calibrated to a 95% PPV) can help to pinpoint the most likely candidate for manual review and functional validation.\u003c/p\u003e \u003cp\u003eHowever, this flexibility embodies a critical trade-off. While the high-confidence threshold provides very strong evidence for the variants it flags, it only captures approximately 55% of the pathogenic variants in our independent test set. Careful review of candidates that fall below this stringent cutoff remains essential. Ultimately, our simulations serve as a proxy for clinical reality, and real-world implementation is the necessary next step to confirm the usefulness of this integrated strategy.\u003c/p\u003e \u003cp\u003eDespite its overall strong performance, MobiDeep has areas for future enhancement that define our roadmap. Its current scope focuses on single nucleotide variants (SNVs) and its predictive power is constrained by its five input features. For instance, our ranking analysis revealed that a subset of variants, particularly in 5' UTRs, were consistently ranked poorly (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eB). Post-hoc analysis of these variants, like NM_000044.6:c.-547C\u0026thinsp;\u0026gt;\u0026thinsp;T, confirmed they disrupt upstream Open Reading Frames, a mechanism not explicitly modeled by the input VEPs. As shown by Meguerditchian et al.\u003csup\u003e25\u003c/sup\u003e, uORF-altering variants represent a significant fraction of functional NCVs. The future integration of features from specialized tools like MORFEE\u003csup\u003e\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e is a clear priority to improve MobiDeep's mechanistic coverage and address this specific weakness.\u003c/p\u003e \u003cp\u003eSimilarly, deep intronic variants without predicted splicing effects represent another mechanistic frontier requiring enhanced interpretation frameworks. The example of NM_003024.3(ITSN1):c.3661\u0026thinsp;+\u0026thinsp;1376T\u0026thinsp;\u0026gt;\u0026thinsp;C (Supplementary Figure S2), located\u0026thinsp;+\u0026thinsp;1376 nucleotides from the canonical splice donor, illustrates this challenge. At such distances, conventional splicing mechanisms cannot account for pathogenicity. Several non-mutually-exclusive mechanisms may explain the functional impact of these variants. First, some annotated \"intronic\" variants may in fact affect regulatory elements within alternative 3' UTRs of minor transcript isoforms that terminate within what is annotated as an intron of the canonical, longer transcript. This phenomenon of overlapping isoforms with differential 3' UTR usage could be particularly relevant given the critical role of 3' UTR composition in determining mRNA stability, localization, and translational efficiency. Alternative mechanisms include disruption of intronic enhancers, polyadenylation signals, or non-coding RNA elements that regulate expression in a tissue or developmental-stage-specific manner. Distinguishing among these mechanisms for individual variants requires systematic functional characterization combined with improved transcript-level annotation that captures tissue-specific and cell-type-specific isoform expression. As with uORF-disrupting variants, the future integration of specialized predictive features that explicitly model these regulatory mechanisms would enhance interpretability for this important class of variants.\u003c/p\u003e \u003cp\u003eCollectively, these examples underscore that while MobiDeep achieves strong overall performance through integration of existing tools, further improvements will require expanding its mechanistic scope through incorporation of specialized features that capture variant effects not adequately represented in current VEPs. Furthermore, this work represents a comprehensive retrospective validation and serves as proof of concept; prospective validation on large, diverse cohorts of undiagnosed patients is the necessary next step to confirm the model's real-world clinical impact.\u003c/p\u003e "},{"header":"Methods","content":"\u003cp\u003eDataset curation\u003c/p\u003e \u003cp\u003eWe established a rigorous data curation pipeline to generate distinct, high-confidence datasets for model training, testing and independent evaluation. All variant coordinates were based on the GRCh38/hg38 human reference genome.\u003c/p\u003e \u003cp\u003eWe aggregated a set of pathogenic non-coding variants (NCVs) from several sources. The primary source was the ClinVar VCF file (release 2025-02-09)\u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e. This was supplemented with variants from the Human Gene Mutation Database (HGMD Professional v.2024.1), from which we extracted \"Disease Causing Mutations\" (DM class) located in regulatory regions such as non-coding RNA, miRNA, and TFBS (queried March 10, 2024)\u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e. The dataset was further enriched with manually curated NCVs from published literature and functionally validated variants provided by collaborating French diagnostic genetics laboratories.\u003c/p\u003e \u003cp\u003eAll variants (SNVs), regardless of their source, were processed through a uniform quality control pipeline. To ensure variants were strictly non-coding and to minimize confounding effects from strong splicing alterations, we used Ensembl Variant Effect Predictor (VEP v103)\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e to annotate them against all available transcripts, without applying transcript selection filters (i.e., --flag_pick_allele was not used) or GENCODE Basic restrictions. A variant was excluded from our dataset if any of its annotated consequences across any transcript matched protein-coding categories (missense_variant, stop_gained, stop_lost, start_lost, frameshift_variant, inframe_insertion, inframe_deletion, or protein_altering_variant) or if it was located at canonical splice sites (\u0026plusmn;\u0026thinsp;1, \u0026plusmn;\u0026thinsp;2). This stringent filtering approach ensured that variants with mixed consequences, for instance, annotated as 5_prime_UTR_variant on multiple transcripts but as missense_variant on a single transcript, were conservatively excluded to maintain a strictly non-coding dataset. Additionally, variants with a predicted SpliceAI\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e delta score\u0026thinsp;\u0026gt;\u0026thinsp;0.2 (maximum across acceptor gain, acceptor loss, donor gain, and donor loss) were removed from the collection (based on SpliceAI pre-computed dataset v1.3) to further minimize splicing-related confounding effects..\u003c/p\u003e \u003cp\u003eThis filtered collection of NCVs was then partitioned into two subsets. The primary training and benchmarking cohort was defined as the high-confidence Pathogenic / Likely Pathogenic (P/LP) set (N\u0026thinsp;=\u0026thinsp;448), composed exclusively of variants present in ClinVar with a review status of \"criteria_provided,_multiple_submitters,_no_conflicts\", \"reviewed_by_expert_panel\", or \"practice_guideline\" (i.e., \u0026ge;\u0026thinsp;2 stars). This gold-standard set was primarily distributed across 5' UTR (n\u0026thinsp;=\u0026thinsp;129), upstream (n\u0026thinsp;=\u0026thinsp;109), intronic (n\u0026thinsp;=\u0026thinsp;84), non-coding exons (n\u0026thinsp;=\u0026thinsp;73), 3\u0026rsquo;UTR (n\u0026thinsp;=\u0026thinsp;46) and other (i.e., intergenic, n\u0026thinsp;=\u0026thinsp;7) regions, based on the VEP annotation. All other variants that passed our filtering criteria, including those sourced from HGMD, literature, collaborators, or those in ClinVar with a\u0026thinsp;\u0026lt;\u0026thinsp;2-star status, were compiled into the independent P/LP holdout test set (N\u0026thinsp;=\u0026thinsp;137). This set was held in reserve for further evaluation of the model on unseen data and for the ranking simulations.\u003c/p\u003e \u003cp\u003eFor benign variants, a pool of presumed benign variants was derived from the TraitGym mendelian_traits_full dataset\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e, a resource containing\u0026thinsp;~\u0026thinsp;4.95\u0026nbsp;million variants. From this pool, we curated a large-scale set of 38,146 unique benign NCVs that were used as the source for negative controls in all subsequent region-specific analyses. A ratio of 1\u0026ndash;10 was used at benchmarking (bootstrapping) and model training (4,480 fixed neutral variants). It is important to note that these variants are relatively frequent and therefore have a very low prior probability of being pathogenic for rare Mendelian diseases. While this does not definitively prove their functional neutrality on an individual basis, it provides a robust and conservatively selected set of negative controls for a rare disease context.\u003c/p\u003e \u003cp\u003eWhile most of our pathogenic variants were sourced from the public ClinVar database (release 2025-02-09), ensuring full transparency, the dataset was also supplemented with variants from the HGMD Professional database (v.2024.1) and collaborating French diagnostic laboratories. Due to commercial licensing restrictions for HGMD Pro and regulatory constraints for patient data from clinical laboratories, these specific variant lists cannot be shared publicly, with the exception of those from Clinvar. However, to ensure maximum reproducibility, we have meticulously documented the version numbers of the databases. This allows for the precise reconstruction of our methodology.\u003c/p\u003e \u003cp\u003eModel training\u003c/p\u003e \u003cp\u003eOur model was trained and evaluated on a fixed dataset constructed from the core variant sets described previously. We combined the entire set of 448 high-confidence P/LP variants with 4,480 benign variants, which were randomly sampled once from our benign pool to create a static cohort. This final dataset of 4,928 variants, maintaining a strict 1:10 ratio, was then deterministically partitioned into a 90% training set (403 P/LP and 4,032 benign variants) and 10% hold-out test set (45 P/LP and 448 benign variants).\u003c/p\u003e \u003cp\u003eIndependent dataset\u003c/p\u003e \u003cp\u003eFor a final, completely unbiased assessment of model performance, we utilized the independent P/LP test set (N\u0026thinsp;=\u0026thinsp;137), which had no overlap with the training data. This set was used in two validation experiments: The first is an independent benchmarking where the 137 P/LP variants were paired with 1,370 newly randomly sampled benign variants (maintaining the 1:10 ratio) to directly compare the performance of the final MobiDeep model against existing VEPs on unseen data. The second is an \u003cem\u003ein-silico\u003c/em\u003e ranking simulation where they served as the Variants of Interest (VOIs) in our large-scale ranking simulations to evaluate MobiDeep's ability to prioritize causal variants within large, clinically realistic backgrounds of neutral variants.\u003c/p\u003e \u003cp\u003eBenchmarking the VEPs\u003c/p\u003e \u003cp\u003eChoice of VEPS\u003c/p\u003e \u003cp\u003eFive VEPs were selected for benchmarking (CADD\u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e (v1.7), ReMM\u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e (v0.4), GPN-MSA\u003csup\u003e\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e, PhyloP_Primates\u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e, and Cactus241way\u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e), based on their published performance, resolution (1 base pair), diverse underlying methodologies and availability of pre-computed datasets covering the whole human genome. The Regulatory Mendelian Mutation score (ReMM v.0.4)\u003csup\u003e17\u003c/sup\u003e, is a random forest model supervised on clinically annotated variants. It was specifically trained to distinguish known pathogenic non-coding variants from HGMD against a background of common variants, using a curated feature set optimized for regulatory function, including epigenomic data such as DNase I hypersensitivity, key histone modifications (H3K4me1/3, H3K27ac), and eQTL data. In contrast, Combined Annotation Dependent Depletion (CADD 1.7)\u003csup\u003e16\u003c/sup\u003e is an integrative model trained to distinguish observed human variation from a simulated genomic background, independent of clinical labels. Its power derives from the comprehensive integration of over 90 diverse annotations, combining foundational conservation metrics with a vast catalogue of ENCODE\u003csup\u003e\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e and Roadmap Epigenomics data\u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e, as well as other genomic and transcript-level features. Representing a newer paradigm, GPN-MSA (genomic pretrained network with multiple-sequence alignment)\u003csup\u003e\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e is a sequence-based deep learning model that leverages a transformer architecture to learn complex evolutionary constraints directly from the DNA context within large-scale multiple sequence alignments of 100 vertebrate species. Finally, to provide a foundational and orthogonal measure of evolutionary constraint, we included phyloP conservation scores, which quantify nucleotide-level selection. We utilized scores derived from two distinct alignments: a primate-specific alignment (phyloP_Primates)\u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e and a broad 241-mammals alignment (Cactus241way)\u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e, thereby capturing constraint across different evolutionary timescales. Pre-computed, genome-wide scores for GRCh38 were downloaded from their official sources. Scores for each variant were extracted using tabix and home-made python scripts.\u003c/p\u003e \u003cp\u003eBenchmarking VEPS\u003c/p\u003e \u003cp\u003eThe benchmarking analysis was performed over 10,000 bootstrap iterations. In each iteration, a test dataset was constructed by combining the fixed set of 448 high-confidence P/LP variants with 4,480 benign variants selected via sampling with replacement from the global benign pool, maintaining a strict 1:10 pathogenic-to-benign ratio. For region-stratified analysis, a similar bootstrap was performed within each harmonized genomic region (5' UTR, 3' UTR, Intron, Upstream, Non-coding exon), sampling benign variants only from the corresponding regional pool.\u003c/p\u003e \u003cp\u003eFor each of the 10,000 datasets, multiple performance metrics (Confusion matrix, AUROC, AUPRC, F1-score, MCC) were calculated, and an optimal classification threshold was determined by maximizing the Youden's J Index (Sensitivity\u0026thinsp;+\u0026thinsp;Specificity\u0026thinsp;\u0026minus;\u0026thinsp;1). This process generated mean and standard deviation for performance estimates for each VEP, both globally and across specific genomic contexts.\u003c/p\u003e \u003cp\u003eModel development\u003c/p\u003e \u003cp\u003eModel selection\u003c/p\u003e \u003cp\u003eWe developed MobiDeep, a deep learning classifier based on a Multilayer Perceptron (MLP) architecture implemented using scikit-learn v1.6.1 library. The MLP architecture was selected for its well-established capability in handling complex, non-linear classification tasks and its effectiveness in learning hierarchical feature representations from high-dimensional biological data. The model was optimized using Optuna, maximizing Area Under the Precision-Recall Curve (AUPRC) on 5-fold cross-validation to ensure robust performance.\u003c/p\u003e \u003cp\u003eFeature selection and final model optimization\u003c/p\u003e \u003cp\u003eFive VEPs, CADD 1.7, GPN-MSA, ReMM v.0.4, Cactus241way, and Phylop_Primates were chosen as inputs for MobiDeep: to prepare these features and address the 1:10 class imbalance, we constructed a multi-step pipeline using the scikit-learn v1.6.1 library. This structure is critical as it ensures that all preprocessing steps, especially resampling, are applied only to the training data within each fold of our cross-validation, preventing data leakage. The pipeline consisted of three sequential steps: 1) Iterative Imputation where missing values were imputed using an IterativeImputer with a RandomForestRegressor estimator; 2) Feature Scaling where all features were subsequently normalized to a [0, 1] range using a MinMaxScaler; 3) Class Imbalance Resampling: The Synthetic Minority Over-sampling Technique (SMOTE) was applied to the training data to generate synthetic examples of the minority (pathogenic) class.\u003c/p\u003e \u003cp\u003eThe final MobiDeep architecture and its hyperparameters were determined via a 200-trial Optuna study that optimized this entire pipeline for AUPRC. This entire workflow was encapsulated in a single scikit-learn pipeline object to ensure that all data transformations are consistently and reproducibly applied during any prediction task.\u003c/p\u003e \u003cp\u003ePerformance evaluation framework\u003c/p\u003e \u003cp\u003eAfter training was complete, the final, optimized pipeline was evaluated a single time on the unseen 10% hold-out test set. We calculated standard classification metrics (AUROC, AUPRC, confusion matrix, F-scores) on this set to obtain an unbiased estimate of the model's generalization performance.\u003c/p\u003e \u003cp\u003eClinical thresholds\u003c/p\u003e \u003cp\u003eFor clinical applicability, we determined optimal probability thresholds for global and genomic region-specific analyses using bootstrap resampling (10,000 iterations with replacement) on 448 pathogenic and 38,146 benign variants. Thresholds were selected by maximizing Youden's J index on the ROC curve to optimize sensitivity and specificity.\u003c/p\u003e \u003cp\u003eWe also establish a more stringent threshold to achieve a Positive Predictive Value (PPV) of 95% for high-confidence predictions for manual curation.\u003c/p\u003e \u003cp\u003e \u003cem\u003eIn silico\u003c/em\u003e ranking simulation\u003c/p\u003e \u003cp\u003eTo assess MobiDeep's utility in a realistic filtering scenario, we conducted a large-scale ranking simulation using the independent test set (n\u0026thinsp;=\u0026thinsp;137). For each VOI, we created test pools by mixing it with N neutral variants bootstrapped from the whole TraitGym dataset, where N was 1,000, 10,000, and 100,000. To ensure robustness, this process was replicated 5 times for each VOI at each background size using different random seeds. In every simulation, all N\u0026thinsp;+\u0026thinsp;1 variants were scored by the MobiDeep pipeline and ranked. We recorded the absolute rank of the VOI and calculated the mean/median rank across the 5 replicates. We also calculated the percentage of VOIs that ranked in the Top 1, 5, 10, and 20 absolute positions, and within the Top 1%, 5%, and 10% of the total pool size.\u003c/p\u003e \u003cp\u003eComplementary test on pathogenic RNU4-2 variants\u003c/p\u003e \u003cp\u003eTo test the model's generalization capabilities on a specific gene and disease mechanism not represented in our training data, we curated a set of 30 variants in the RNU4-2 gene from ClinVar (accessed 2025-05-30). We applied the trained MobiDeep model to these variants and evaluated its ability to assign high, pathogenic-range scores to the known P/LP variants using the region specific (non-coding exon) and the stringent PPV 95% thresholds.\u003c/p\u003e \u003cp\u003eOpen-source availability and adherence to FAIR principles\u003c/p\u003e \u003cp\u003eTo promote reproducibility and community use, and to ensure our framework is Findable, Accessible, Interoperable, and Reusable (FAIR), the complete MobiDeep framework, including the trained model pipeline is available on GitHub at (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/mobidic/MobiDeep\u003c/span\u003e\u003cspan address=\"https://github.com/mobidic/MobiDeep\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). MobiDeep is provided for academic and non-commercial use only, under a GNU GPL (GPLv3) license. An Apptainer container is also provided to ensure a fully reproducible environment. For users without a computational background, MobiDeep is accessible via a user-friendly web application, MobiDetails\u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e. Finally, to maximize utility and interoperability, we are also releasing a pre-computed dataset of MobiDeep scores for 8,773\u0026nbsp;billion single nucleotide variants (SNVs) covering 94.7% of all genomic positions across the GRCh38p14 reference genome. The dataset includes the raw score and a log-transformed score for easier interpretation.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eAcknowledgements\u003c/p\u003e\n\u003cp\u003eWe would like to thank the ANPGM (Association Nationale des Praticiens en G\u0026eacute;n\u0026eacute;tique Mol\u0026eacute;culaire), Fondation Groupama \u0026laquo;\u0026nbsp;Vaincre les Maladies Rares\u0026nbsp;\u0026raquo; and the FHU GenOMedS for their support.\u003c/p\u003e\n\u003cp\u003eWe would also like to express our gratitude to the MoBiDiC platform (Montpellier Bioinformatics for Clinical Diagnosis). Our thanks also go to the Montpellier Laboratory of Computer Science, Robotics, and Microelectronics (LIRMM), especially to Charles Lecellier, Laurent Br\u0026eacute;h\u0026eacute;lin, and Elliot Butz for their valuable expertise. Finally, we thank Roselyne Vallo (Biostatistician / Data manager, Pathogenesis \u0026amp; Control of Chronic \u0026amp; Emerging Infections unit, Inserm) for her expertise in biostatistics.\u003c/p\u003e\n\u003cp\u003e7. Funding\u003c/p\u003e\n\u003cp\u003eThis work was supported by Fondation Groupama \u0026laquo;\u0026nbsp;Vaincre les Maladie Rares\u0026nbsp;\u0026raquo;, the French Association of Molecular Genetics Practitioners ANPGM (Association Nationale des Praticiens en G\u0026eacute;n\u0026eacute;tique Mol\u0026eacute;culaire) and FHU GenOMedS.\u003c/p\u003e\n\u003cp\u003e8. Author information\u003c/p\u003e\n\u003cp\u003eContributions\u003c/p\u003e\n\u003cp\u003eStudy design: AB, DB, SCA\u003c/p\u003e\n\u003cp\u003eData collection and management: AB, DB\u003c/p\u003e\n\u003cp\u003eMobidetails web app implementation: DB\u003c/p\u003e\n\u003cp\u003eExpertise: JM-SA, SCA, VK, A-FR\u003c/p\u003e\n\u003cp\u003eData sharing: OE, JN, MK, LQ, TC, SC, PB, YJ, CR, MC\u003c/p\u003e\n\u003cp\u003eAnalysis and results interpretation: AB, DB, SCA, JM-SA\u003c/p\u003e\n\u003cp\u003eBioinformatics infrastructure and support: DB, CVG, OA\u003c/p\u003e\n\u003cp\u003eWriting: AB, DB, all other authors contributed to the writing and review of the manuscript\u003c/p\u003e\n\u003cp\u003eThe authors read and approved the final manuscript\u003c/p\u003e\n\u003cp\u003eCorresponding authors\u003c/p\u003e\n\u003cp\u003eCorrespondence to David BAUX at [email protected] or Abdelhakim BOUAZZAOUI at [email protected]\u003c/p\u003e\n\u003cp\u003e9. Ethics declarations\u003c/p\u003e\n\u003cp\u003eEthics approval and consent to participate\u003c/p\u003e\n\u003cp\u003eThis research only used publicly available data. The remaining patients\u0026rsquo; variants underwent a full anonymization process.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eConsent for publication\u003c/p\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\u003cp\u003eCompeting interests\u003c/p\u003e\n\u003cp\u003eThe authors declare no competing interests\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003ePandey R, et al. A meta-analysis of diagnostic yield and clinical utility of genome and exome sequencing in pediatric rare and undiagnosed genetic diseases. Genet Med. 2025;27:101398.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEllingford JM, et al. Recommendations for clinical interpretation of variants found in non-coding regions of the genome. Genome Med. 2022;14:73.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAbadie C, et al. \u003cem\u003ePFMG2025\u003c/em\u003e\u0026ndash;integrating genomic medicine into the national healthcare system in France. Lancet Reg Health - Eur. 2025;50:101183.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTurnbull C, et al. The 100 000 Genomes Project: bringing whole genome sequencing to the NHS. BMJ. 2018;361:k1687.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMohan RA, et al. A mutation in the Kozak sequence of GATA4 hampers translation in a family with atrial septal defects. Am J Med Genet A. 2014;164A:2732\u0026ndash;8.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGordon CT, et al. Identification of novel craniofacial regulatory domains located far upstream of SOX9 and disrupted in Pierre Robin sequence. Hum Mutat. 2014;35:1011\u0026ndash;20.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eStenson PD, et al. The Human Gene Mutation Database (HGMD\u0026reg;): optimizing its use in a clinical diagnostic or research setting. Hum Genet. 2020;139:1197\u0026ndash;207.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLandrum MJ, et al. ClinVar: public archive of interpretations of clinically relevant variants. Nucleic Acids Res. 2016;44:D862\u0026ndash;868.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVillani RM, McKenzie ME, Davidson AL, Spurdle AB. Regional-specific calibration enables application of computational evidence for clinical classification of 5\u0026rsquo; cis-regulatory variants in Mendelian disease. Am J Hum Genet. 2024;111:1301\u0026ndash;15.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ede Sainte Agathe J-M, et al. SpliceAI-visual: a free online tool to improve SpliceAI splicing variant interpretation. Hum Genomics. 2023;17:7.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLeman R, et al. SPiP: Splicing Prediction Pipeline, a machine learning tool for massive detection of exonic and intronic variant effects on mRNA splicing. Hum Mutat. 2022;43:2308\u0026ndash;23.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJaganathan K, et al. Predicting Splicing from Primary Sequence with Deep Learning. Cell. 2019;176:535\u0026ndash;e54824.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLivesey BJ et al. Guidelines for releasing a variant effect predictor. \u003cem\u003eArXiv\u003c/em\u003e arXiv:2404.10807v1 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMcLaren W, et al. The Ensembl Variant Effect Predictor. Genome Biol. 2016;17:122.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBenegas G, Eraslan G, Song YS, Benchmarking. DNA Sequence Models for Causal Regulatory Variant Prediction in Human Genetics. BioRxiv Prepr Serv Biol. 2025. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1101/2025.02.11.637758\u003c/span\u003e\u003cspan address=\"10.1101/2025.02.11.637758\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. 2025.02.11.637758.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchubach M, Maass T, Nazaretyan L, R\u0026ouml;ner S, Kircher M. CADD v1.7: using protein language models, regulatory CNNs and other nucleotide-level scores to improve genome-wide variant predictions. Nucleic Acids Res. 2024;52:D1143\u0026ndash;54.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchubach M, Nazaretyan L, Kircher M. The Regulatory Mendelian Mutation score for GRCh38. \u003cem\u003eGigaScience\u003c/em\u003e 12, giad024 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBenegas G, Albors C, Aw AJ, Ye C, Song YS. GPN-MSA: an alignment-based DNA language model for genome-wide variant effect prediction. \u003cem\u003eBioRxiv Prepr. Serv. Biol.\u003c/em\u003e 2023.\u003cdiv class=\"ExternalRefDOI\"\u003e10.10.561776\u003c/div\u003e (2024) doi:10.1101/2023.10.10.561776.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKuderna LFK, et al. Identification of constrained sequence elements across 239 primate genomes. Nature. 2024;625:735\u0026ndash;42.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChristmas MJ, et al. Evolutionary constraint and innovation across hundreds of placental mammals. Science. 2023;380:eabn3943.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eENCODE Project Consortium. An integrated encyclopedia of DNA elements in the human genome. Nature. 2012;489:57\u0026ndash;74.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChadwick LH. The NIH Roadmap Epigenomics Program data resource. Epigenomics. 2012;4:317\u0026ndash;24.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBaux D, et al. MobiDetails: online DNA variants interpretation. Eur J Hum Genet EJHG. 2021;29:356\u0026ndash;60.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDylan, A\u0026iuml;ssi et al. MORFEE: a new tool for detecting and annotating single nucleotide variants creating premature ATG codons from VCF files. \u003cem\u003ebioRxiv\u003c/em\u003e 2020.03.29.012054 (2020) \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1101/2020.03.29.012054\u003c/span\u003e\u003cspan address=\"10.1101/2020.03.29.012054\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMeguerditchian C, et al. Enhancing the annotation of small ORF-altering variants using MORFEE: introducing MORFEEdb, a comprehensive catalog of SNVs affecting upstream ORFs in human 5\u0026rsquo;UTRs. NAR Genomics Bioinforma. 2025;7:lqaf017.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eStenton SL, et al. Critical assessment of variant prioritization methods for rare disease diagnosis within the rare genomes project. Hum Genomics. 2024;18:44.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTenywa J-F, et al. Genome region aware CADD thresholds for noncoding variant prioritization. NAR Genomics Bioinforma. 2025;7:lqaf157.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Genome sequencing, Non-coding variants, Variant prioritization, Meta-score, Machine learning","lastPublishedDoi":"10.21203/rs.3.rs-8823759/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8823759/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003eThe interpretation of non-coding variants (NCVs) from genome sequencing represents a major bottleneck in the diagnosis of rare diseases. Existing variant effect predictors (VEPs) show variable performance across different genomic contexts, and a lack of region-specific clinical guidance hinders accurate variant prioritization. This study aimed to rigorously benchmark state-of-the-art VEPs to define region-specific thresholds and to develop MobiDeep, a novel meta-score designed to improve NCV prioritization.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eWe curated a high-confidence dataset of 448 pathogenic NCVs (ClinVar, HGMD, literature) and 38,146 presumed benign NCVs. Critically, variants affecting splicing were excluded to focus on strictly regulatory mechanisms. We benchmarked the performance of ReMM, CADD, GPN-MSA, Cactus241way, and phyloP, both globally and stratified by genomic region (e.g., 5'UTR, 3'UTR). Subsequently, we developed MobiDeep, a neural network integrating these five scores, optimized using Optuna and validated on an independent holdout set of pathogenic NCVs.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eBenchmarking confirmed that no single tool is universally optimal, with performance varying significantly by genomic context; while ReMM excelled in non-coding exons (AUROC\u0026thinsp;=\u0026thinsp;0.987), GPN-MSA demonstrated superior performance for 3'UTRs (AUROC\u0026thinsp;=\u0026thinsp;0.901). We established data-driven clinical thresholds, identifying an optimal global cutoff of 10.37 for CADD v1.7, validating previous works of CADD\u0026thinsp;\u0026ge;\u0026thinsp;10 for regulatory variants and 0.80 for ReMM. Building on these insights, MobiDeep significantly outperformed all individual predictors on an independent test set, achieving an AUROC of 0.973 and an AUPRC of 0.888. In large-scale simulations mimicking a diagnostic, MobiDeep prioritized causal variants effectively, placing 52.0% and 75% within the top 5 and top 20 ranks respectively. Furthermore, the model correctly prioritized all Clinvar pathogenic variants in the recently discovered RNU4-2 non-coding gene.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e \u003cp\u003eOur findings confirm that individual predictors and uniform thresholds are insufficient for interpreting the diverse landscape of non-coding variants. We demonstrate that region-specific calibration is essential for accurate prioritization.. Our meta-score MobiDeep improves classification performance compared to existing tools. This meta-score serves as a robust filter to streamline the identification of high-confidence variants, thereby facilitating focused manual review and subsequent biological validation in diagnostic settings.\u003c/p\u003e","manuscriptTitle":"MobiDeep: an AI-based meta-score for scoring non-coding DNA variations","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-03-11 19:06:38","doi":"10.21203/rs.3.rs-8823759/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"b2ec7d4b-fb03-4a7c-8752-3ef2112aae12","owner":[],"postedDate":"March 11th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-04-21T15:11:47+00:00","versionOfRecord":[],"versionCreatedAt":"2026-03-11 19:06:38","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8823759","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8823759","identity":"rs-8823759","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}