Ihor Kendiukhov
A systematic evaluation of single-cell foundation models reveals that attention mechanisms capture co-expression patterns rather than unique regulatory signals, with gene-level baselines outperforming attention-based predictions.
In this study, researchers evaluated how well single-cell foundation models interpret biological data. They discovered that the models' attention mechanisms, which are supposed to highlight important features, actually capture patterns of genes being expressed together rather than unique regulatory signals. This means that simpler methods, which look at genes individually, are often better at predicting biological changes. The researchers also developed a new approach to improve model interpretability, setting a standard for future studies in this field.
