Alex Xijie Lu

FLIP2 adds seven protein fitness datasets and 16 splits for enzymes, light-sensitive proteins, and interactions; simple ridge and zero-shot baselines often rival fine-tuned protein language models.

Vermeer generates synthetic microscopy images of protein localization using protein sequences and cell morphology, outperforming previous methods in quality and flexibility.

The paper maps the functional landscape of the human intrinsically disordered proteome using evolutionary signatures to classify and predict IDR functions.

This paper introduces EnhancAR, a deep learning model that generates functionally diverse enhancers using evolutionary homologs instead of cell type labels.

This paper introduces a pretraining task for genome language models that predicts evolutionary rate, showing improved performance over sequence-only pretraining.

This paper evaluates how background cells affect single-cell microscopy models and finds they reduce localization classification accuracy by up to 15.8%.

This study analyzes transfer learning in protein language models, finding that downstream task performance often relies on early pretraining features rather than scaling with pretraining duration.

Used without fine-tuning, Geneformer and scGPT show limited reliability for cell type clustering and batch effect correction, often landing behind simpler methods.

The paper introduces BEN, a method to correct batch effects in deep learning for microscopy images by aligning experimental and deep learning batches, improving model generalization.

The paper introduces a method to discover molecular features of intrinsically disordered regions using contrastive learning with evolutionary conservation.

This paper introduces a self-supervised CNN method to learn high-quality features for single-cell microscopy images without labeled data, showing improved performance in protein localization and analysis.