Best Practices

In this notebook, we will curate papers suggesting metrics for each domain.

Small molecules

• Fréchet ChemNet Distance: A metric for generative models for molecules in drug discovery (2018)

• MOSES: A Benchmarking Platform for Molecular Generation Models (2020)

• GuacaMol: Benchmarking Models for De Novo Molecular Design (2019)

• How Evaluation Choices Distort the Outcome of Generative Drug Discovery (2025)

DNA

• DiscDiff: Latent Diffusion Model for DNA Sequence Generation (2024) — introduces Fréchet Reconstruction Distance (FReD), a DNA analog to FID

RNA

• RNAInvBench: Benchmark for the RNA Inverse Design Problem (2024)

• Deep Generative Models Design mRNA Sequences with Enhanced Translational Capacity and Stability (2025)

Peptides

No standardized benchmark platform exists yet (analogous to MOSES for small molecules) — current papers use ad hoc combinations of perplexity, novelty, diversity, and property predictors.

• Deep Generative Models for Peptide Design (2022) — review explicitly noting the lack of standardized benchmarks

• Deep Generative Models for Therapeutic Peptide Discovery: A Comprehensive Review (2025)

Proteins

• Towards Robust Evaluation of Protein Generative Models: A Systematic Analysis of Metrics (2025)

• Protein FID: Improved Evaluation of Protein Structure Generative Models (2025)

• ProteinBench: A Holistic Evaluation of Protein Foundation Models (2024)