The AI drug discovery landscape has grown rapidly, with multiple academic and commercial platforms competing across overlapping domains. We believe the research community benefits from honest, transparent benchmarking. In this post, we compare Nutracie's products against leading alternatives across three core tasks: protein structure prediction, molecular docking, and de novo molecule generation.
For protein structure prediction, we compared NutraFold v3.0 against AlphaFold2, ESMFold, and RoseTTAFold on the CASP15 benchmark. NutraFold achieved a mean GDT-TS of 0.92, compared to AlphaFold2 at 0.91, RoseTTAFold at 0.87, and ESMFold at 0.84. The differences between NutraFold and AlphaFold2 are modest and within statistical noise for most target categories. Where NutraFold shows a clearer advantage is speed (3x faster than AlphaFold2 on H100) and antibody-specific prediction, where specialized training data gives NutraFold a meaningful edge.
For molecular docking, we compared NutraDiscover's docking module against DiffDock, GNINA, and AutoDock Vina on the PDBbind v2020 core set. NutraDiscover achieved a success rate of 78.3% (RMSD < 2 Angstrom), compared to DiffDock at 72.1%, GNINA at 69.8%, and Vina at 58.4%. The advantage is more pronounced on challenging targets with deep, flexible binding pockets.
For de novo molecule generation, we compared NutraDiscover's generative engine against RFdiffusion (for protein design), REINVENT, and MolGPT on the GuacaMol benchmark. NutraDiscover excels on 3D-aware metrics (shape similarity, pharmacophore overlap) while REINVENT shows advantages on certain 2D property optimization tasks. We provide detailed per-metric breakdowns in our technical report.
Our takeaway: no single platform dominates across all tasks. Each tool has strengths in specific domains. AlphaFold2 remains the gold standard for single-chain structure prediction. DiffDock offers strong performance with minimal computational requirements. RFdiffusion is unmatched for de novo protein design. NutraDiscover's advantages are in integrated workflows that combine target identification, molecular generation, and ADMET optimization in a single pipeline, and in the enterprise features (security, compliance, federated compute) that pharmaceutical customers require.
We commit to updating these benchmarks quarterly as methods and datasets evolve. All benchmark code and datasets (where legally permissible) are available on our GitHub repository. We welcome the community to reproduce and extend these comparisons.