Deviation Error: assessing machine learning predictions for replicate measurements in genomics and beyond

The preprint operates within the paradigm of machine learning applied to genomics, specifically focusing on predictive modeling for replicate measurements. It accepts the established methods and assumptions of this field, such as the use of standard loss functions and the focus on minimizing prediction error.

This preprint introduces the "Deviation Error," a novel metric designed to improve the assessment of machine learning predictions in the context of measurement variation commonly seen in genomics data. The authors explore various established loss functions and compare their performance against the Deviation Error using simulated genomics data. While the Deviation Error itself doesn't yield significantly improved results compared to some standard loss functions (like MSE), the paper's primary contribution lies in its detailed analysis of measurement variation in genomics and its implications for model evaluation. This work falls squarely within "normal science" as defined by Kuhn. It's an attempt to refine existing tools and improve accuracy within the accepted paradigm, rather than proposing a radical shift in methodology or understanding. It's possible this work could be considered "Model Drift", should the deviation error metric gain traction. This is because the authors demonstrate that existing models and evaluation metrics might be slightly misaligned with the realities of noisy genomic data, and they propose a correction in the form of the Deviation Error. However, since the results don't demonstrate a significant improvement using this metric, the classification leans more strongly towards "Normal Science" for the time being. More research and evaluation are needed to discern if there is drift.