References

It's amazing to see you guys improve so quickly in refining these genotype-phenotype mapping methodologies! For someone less familiar with the cutting edge, it seems like the novelty stems from integrating more recent machine learning techniques in an opportunistic fashion. This must be a fun challenge. I have a more general question regarding plans for applying these new methodologies. Given the vast amount of existing GWAS and QTL mapping data analyzed with traditional linear regression, are there specific datasets you're excited about trying out these enhanced attention-based models with the goal of potentially uncovering novel insights missed by those earlier methods like GWAS?

This is a fantastic demonstration of how canonical architectures can outperform bespoke design—even in complex biological domains. I’m particularly intrigued by your decision to model all phenotypes jointly using a shared output head, which clearly capitalizes on cross-trait correlations.

Two follow-up questions:

Epistasis interpretability: Given the motivation around capturing nonlinear interactions, would you consider incorporating synthetic datasets with known epistatic structure to more directly assess interpretability of the attention weights or learned representations?

Positional encoding: While you omitted positional information to reflect the unordered nature of loci, have you considered evaluating models that incorporate chromosomal positions or LD block structure as soft ordering? It might reveal latent spatial patterns not captured by purely attention-based inference.

Overall, this work offers a strong foundation for the field and an exciting invitation for future exploration. Thank you for sharing it!

This is a fascinating application of transformer architectures to genotype-phenotype mapping. I’m particularly intrigued by the performance of the canonical transformer across traits without relying on custom attention mechanisms. Could you elaborate on how the model balances trait-specific and shared genetic signals in the multitask setting? Also, are there plans to open-source the fine-tuning scripts or datasets used in cross-trait evaluation to promote reproducibility and further community development?

Could the authors elaborate on how the model handles potential confounding factors in cross-trait learning? Additionally, how does the model's performance vary with traits of differing heritability? Understanding these aspects could provide deeper insights into the model's robustness and applicability across diverse phenotypic traits.

did you ever consider a covariance-aware loss function into the mix? Like, maybe using a multivariate Gaussian negative log-likelihood? I also inquired whether they observed phenotype-specific gains in traits with known pleiotropic linkages, such as temperature-sensitive growth rates.