Jakub M. Bartoszewicz

This study evaluates data leakage effects in protein language model pretraining on thermostability prediction by comparing dataset split strategies.

The paper examines how motif interactions affect post-hoc interpretability of genomic CNNs and shows that interpretability methods can miss motifs when interactions are present.

This paper evaluates prompt tuning as a cost-efficient alternative to fine-tuning protein language models for sequence generation, finding it more efficient in resource usage but less effective than fine-tuning in sequence quality.

The paper curates a fungal host range dataset and uses deep learning to detect novel fungal pathogens from DNA sequences.

This paper develops a deep learning method for real-time detection of novel pathogens during sequencing, showing improved sensitivity and accuracy compared to existing approaches.

CovRadar provides a platform for tracking and filtering SARS-CoV-2 mutations in the spike gene, enabling users to analyze and visualize mutational changes and spatio-temporal distributions.

This paper develops a deep learning method to detect novel human viruses from genome sequencing data, achieving lower error rates than existing methods and providing interpretability tools for analysis.

DeePaC predicts pathogenic potential of novel DNA sequences using reverse-complement neural networks, reducing error rates compared to existing methods.