Reproduce “A measurement noise scaling law for cellular representation learning” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Gokul Gowri, Igor Sadalski, Dan Raviv, Peng Yin, Jonathan Rosenfeld, Allon M. Klein

The paper identifies a quantitative scaling law describing how measurement noise impacts the performance of cellular representation learning models across diverse biological modalities and architectures. This work establishes a framework for benchmarking models by their information capacity and noise robustness, offering practical guidance for experimental design in inherently noisy biological data generation.

Key insights

  • Model performance in biological representation learning follows a predictable sigmoidal scaling curve with increasing measurement noise, precisely captured by the derived logarithmic scaling law.
  • The noise-scaling law provides two interpretable benchmarking metrics: I_max (maximal information capacity in noiseless conditions) and \u03b7 (noise robustness).
  • Different models exhibit trade-offs between noise robustness and information capacity; for example, Geneformer is highly noise-robust, while scVI often achieves higher maximal information capacity under low-noise conditions.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.