SpliceAI Benchmark: Performance of Deep Learning Models on Six Variant Datasets

Six datasets comprising 213,136 genetic variants were used to benchmark splice-altering variant prediction tools. The data compares the original SpliceAI algorithm against two open-source reimplementations and a legacy ensemble baseline. Authored by Nathan Fortier and last updated on 2026-05-13, the dataset is hosted on figshare under a CC-BY-4.0 license.

Use Cases

• Benchmarking splice prediction algorithms based on performance metrics across six distinct variant datasets.
• Calibrating score thresholds for deep intronic variants based on the reported optimal thresholds being an order of magnitude lower than standard recommendations.
• Comparing positional agreement between SpliceAI implementations based on exact splice-site match rates exceeding 90%.
• Analyzing correlation patterns between event types (gain vs. loss) across different algorithm implementations.

Strengths

• Benchmarks performance across six distinct datasets totaling over 213,000 variants.
• Includes a curated set of 1,316 validated variants and 242 manually curated deep intronic pathogenic variants.
• Provides direct comparisons between the original SpliceAI and two open-source reimplementations.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• The dataset is very small (5.5 KB), indicating it likely contains summary statistics rather than raw variant data.

Provenance

Source

figshare

Collection Method

Likely compiled from published variant datasets and performance evaluation results.

Freshness

Last updated 2026-05-13 17:43:06; freshness should be verified.