PREreview of Functional Effect Predictions For Ion Channel Missense Variants Using a Protein Language Model

This review is the result of a virtual, collaborative Live Review organized and hosted by the PREreview team as part of an ongoing collaboration with the Rare As One Network on December 3, 2025. The discussion was joined by 4 people: 2 facilitators from the PREreview Team, 1 members of the Chan-Zuckerberg Initiative Rare As One Project team, and 1 Live Review participants from the Rare As One Network. We thank all participants who contributed to the discussion and made it possible for us to provide feedback on this preprint.

This manuscript describes the development and validation of a novel machine learning classifier (MissION) designed to predict whether missense variants in ion channel genes result in gain-of-function (GOF) or loss-of-function (LOF) effects. Recognizing the clinical importance of accurately distinguishing GOF from LOF variants for diagnosis, treatment selection, and eligibility for clinical trials, the authors assembled a large dataset of 3,176 ion channel missense variants derived from published literature and standardized sources. They trained a protein language model–based classifier and evaluated performance through cross-validation, benchmarking against existing tools, and assessment of generalizability. The model demonstrated strong predictive accuracy and outperformed existing approaches, particularly for ion channel genes lacking electrophysiological recordings, thereby enabling functional inference where experimental data are limited. The inclusion of diverse channel types, including potassium channels, and the integration of both molecular and clinical context enhanced predictive power.

Overall, the research approach is sound and aligned with the study’s goal of improving variant interpretation to support positive patient outcomes. However, as with many machine learning studies relying on curated literature data, performance depends on the accuracy and consistency of prior experimental annotations, which may not always reflect a perfect ground truth. The authors appropriately acknowledge key limitations, including imperfect ground truth labels, variants with mixed or context-dependent effects, and gene-level variability that may influence model performance. Data appear to be accessible (e.g., through the Synaptica Variant Interpreter website), supporting transparency, although full reproducibility would depend on the availability of detailed implementation materials. The conclusions are well supported by the data presented, and the figures and tables effectively communicate model performance and comparative analyses.

We thank the authors of the preprint for sharing their research openly which allowed us to learn about the work and share our feedback openly.

Competing interests

The authors declare that they have no competing interests.

Use of Artificial Intelligence (AI)

The authors declare that they did not use generative AI to come up with new ideas for their review.