Background: Splice-altering variants (SAVs) are genetic changes that disrupt removal of introns from mature RNA precursors. SAVs are a major source of pathogenic DNA variants but most go undetected. SAV detection is facilitated by dozens of computational methods and resources. Many of these are underutilised, poorly understood or not easily accessible, and the task of SAV analysis can be time-consuming and require extensive domain knowledge. To address this, we developed “SAVViDB” and “SpliceChat”, two web-based applications for enhancing the understanding, prediction and interpretation of SAVs.
Results: SAVViDB (freely available at https://savvidb.victorchang.edu.au) integrates sixteen key SAV-related datasets, and displays these as interactive visualisation tracks, ideal for facilitating insights. Using SAVViDB, we identified pervasive issues with the leading SAV prediction tool, SpliceAI, including a mid-exonic "blind spot" resulting from distance constraints, erroneous masking of predicted losses affecting annotated splice sites, and a failure to detect SAVs due to delta score and reference score errors. The ensemble method AbSplice-DNA inherits all three of SpliceAI’s issues, and additionally, completely misses most deep intronic SAVs that are detected by SpliceAI. For each of these issues, we have identified their causes, quantified their impact, and developed solutions. SpliceChat is an LLM-powered assistant for predicting and interpreting SAVs via a conversational interface. For a queried genetic variant, SpliceChat retrieves predictions and other relevant data from many sources, including state-of-the-art resources and methods. An LLM is provided with this data, as well as knowledge and instructions designed by a human expert which allow the LLM to emulate their expert analytical process. The LLM can then summarise and interpret the variant data, provide recommendations, and respond to user questions.
Conclusion: SAVViDB and SpliceChat improve SAV discovery and interpretation by facilitating insights and automating the time-consuming and complex task of SAV analysis.