DeePEn: New Benchmark Measures Protein Engineering Models' Performance on Distant Mutations
Researchers introduced DeePEn, a benchmark that evaluates how well protein engineering models predict fitness across increasing numbers of simultaneous mutations. The benchmark addresses a gap in existing evaluation methods by testing models' generalization capabilities at varying mutational distances from wild-type proteins. This matters because robust benchmarking is essential for advancing practical protein engineering applications.
DeePEn is a depth-sensitive benchmark designed to assess protein engineering models' ability to predict protein fitness when multiple mutations are introduced simultaneously. The researchers evaluated several predictive models—including general and biophysics-informed protein language models and non-transformer neural networks—using four deep mutational scanning datasets from ProteinGym. A key finding is that all tested models show performance degradation as mutational distance increases, meaning predictions become less accurate when proteins differ by many simultaneous point mutations from the wild-type. The benchmark uses edit distance (the number of simultaneous single amino acid variants) to quantify mutational distance. By providing a multi-metric evaluation framework, DeePEn addresses the lack of robust benchmarks that capture real-world protein engineering challenges and offers researchers a standardized resource for evaluating new models.
What's missing
The study does not discuss potential limitations of the four DMS datasets selected or whether results might generalize to other protein families or mutation types beyond point mutations.
What different sources said
- bioRxivCenter
DeePEn - A Depth sensitive benchmark for Protein Engineering
Related
Study finds cerebral blood vessel oscillations are self-generated, not driven by systemic blood pressure
Researchers observed that rhythmic oscillations in brain blood vessel diameter persist during cardiopulmonary bypass surgery when systemic blood pressure oscillations are absent, suggesting the brain generates these oscillations independently. The study involved 14 surgical patients and measured vaso-oscillations at approximately 0.1 Hz frequency. This finding challenges the understanding of how blood flow and fluid transport are regulated in the brain.
New Framework Addresses Missing Data in Space Biology Research Using NASA RR9 Mission Data
Researchers have developed a systematic four-stage imputation framework to handle incomplete datasets from space biology experiments, demonstrated using retinal imaging and omics data from NASA's RR9 mission. Space biology studies are inherently limited by small sample sizes and logistical constraints, making missing data a significant obstacle to building reliable computational models of how the human body responds to spaceflight. The framework is important because it provides practical guidance for preserving biological signals while quantifying trade-offs, though it reveals that imputation can simultaneously improve predictive performance and obscure subtle biological patterns.
OMIO: New Python Library Standardizes Microscopy Image Data Handling
Researchers have developed OMIO, a Python library that standardizes how microscopy images and their metadata are read and processed across different file formats and microscope systems. The tool addresses a longstanding problem in microscopy workflows where different file formats and reader software often introduce errors, metadata loss, or require custom workaround code. This standardization could improve reproducibility and reduce errors in microscopy-based research across biology, medicine, and materials science.