On the Unfounded Enthusiasm for Soft Selective Sweeps III: The Supervised Machine Learning Algorithm That Isn't.

Area of Science:

• Evolutionary Biology
• Genomics
• Computational Biology

Background:

• Soft selective sweeps are increasingly proposed as a major driver of adaptive evolution.
• Previous studies, notably by Schrider and Kern, claimed soft sweeps dominate human genome adaptation and challenged the Neutral Theory of Molecular Evolution.
• These claims relied on methods presented as artificial intelligence, specifically supervised machine learning (SML).

Purpose of the Study:

• To critically evaluate the methodology used in Schrider and Kern's work, particularly their application of supervised machine learning (SML).
• To address the validity of their conclusions regarding soft selective sweeps and their challenge to the Neutral Theory of Molecular Evolution.

Main Methods:

• Analysis of the methodological claims in Schrider and Kern's publications.
• Identification of the absence of a valid training dataset for supervised machine learning (SML).
• Examination of the reliance on simulations with adjustable parameters and selective data-picking.

Main Results:

• Schrider and Kern's claim of using supervised machine learning (SML) is misleading, as they lacked an empirically validated training dataset.
• Their methodology involved unvalidated simulations and data cherry-picking, compromising the integrity of their findings.
• Results from their SML-based algorithms (e.g., S/HIC) lack robust validation and reproducibility.

Conclusions:

• The evolutionary inferences drawn from Schrider and Kern's purported SML analyses are unreliable.
• The study highlights critical flaws in the application of machine learning techniques in evolutionary genomics.
• Findings suggest that claims of soft sweeps dominating adaptation and challenging neutral theory, based on these methods, require significant re-evaluation.