OpenAI has introduced Genebench-Pro, a new benchmark designed to rigorously test AI models on complex genomics problems. Unveiled on June 30, 2026, this benchmark features 10 detailed case studies, each replicating real-world challenges in genetic and clinical research.
These case studies span critical areas, including structural variant-guided tumor therapy, CRISPR target validation, and Mendelian randomization for drug target prioritization. Each scenario provides a specific prompt, relevant datasets, and supporting materials, requiring AI models to produce precise JSON-formatted answers and analytical reasoning.
Benchmark Depth and Diversity
The Genebench-Pro benchmark's strength lies in its diverse and intricate problems. For instance, a somatic oncology case demands an AI to estimate the net clinical utility of a synthetic inhibitor for tumors driven by structural variants.
Another challenge involves functional genomics, where models must discern if an lncRNA dependency is transcript-specific or influenced by neighboring genes. This requires accounting for various confounding factors like local DNA perturbation and guide swaps.
Statistical genetics scenarios push models to prioritize protein drug targets in linked genetic loci, navigating issues like assay scale and linkage disequilibrium. Clinical genomics tests involve estimating ancestry-specific carrier frequencies and residual risks for conditions like DRX1, under complex pseudogene and CNV calibrations.
Single-cell genomics tasks require models to estimate genotype effects on gene expression after meticulous correction for ambient RNA contamination. Structural genetics cases evaluate the ability to assess clinical associations and expression support for nested structural subhaplotypes within inversion-like loci.
Each case study includes extensive datasets, such as clinical registries, expression summaries, and genomic metadata. Models are graded not only on numerical accuracy but also on the quality of their analytical reasoning, documented in a dedicated 'reasoning' field within the JSON output.
