5 November 2023 in bioRxiv
A bstract: In silico transcription factor and DNA (TF-DNA) binding affinity prediction plays a vital role in examining TF binding preferences and understanding gene regulation. The existing tools employ TF-DNA binding profiles from in vitro high-throughput technologies to predict TF-DNA binding affinity. However, TFs tend to bind to sequences in open chromatin regions in vivo , such TF binding preference is seldomly considered by these existing tools. In this study, we developed TRAFICA, an open chromatin language model to predict TF-DNA binding affinity by integrating the characteristics of sequences from open chromatin regions in ATAC-seq experiments and in vitro TF-DNA binding profiles from high-throughput technologies. We applied self-supervised learning to pre-train TRAFICA on over 13 million nucleotide sequences from the peaks in ATAC-seq experiments to learn the TF binding preference in vivo . TRAFICA was further fine-tuned using the TF-DNA binding profiles from PBM and HT-SELEX technologies to learn the association between TFs and their target DNA sequences. We observed that TRAFICA significantly outperformed both machine learning-based and deep learning-based tools in predicting in vitro and in vivo TF-DNA binding affinity. These findings indicate that considering the characteristics of sequences from open chromatin regions could significantly improve TF-DNA binding affinity prediction, particularly when limited TF-DNA binding profiles from high-throughput technologies are available for specific TFs.
29 October 2026 in bioRxiv
Many phage proteins remain functionally uncharacterized despite advances in structural-similarity-based annotation. We introduce PhageFlare, a deep learning framework for predicting phage protein functions lacking significant structural matches to annotated references. Across three structure-disjoint benchmarks, where no significant Foldseek match was detected between training and test proteins (E-value thresholds of
10
−
3
,
10
−
4
, and
10
−
2
), PhageFlare achieved macro-averaged
F
1
scores of 0.871, 0.889, and 0.738 outperforming the strongest baseline (0.789, 0.814, and 0.732). Systematic evaluations demonstrated that protein language model selection critically influences performance, with ESM-C achieving stronger overall performance than ESM-2 and ProtBert. Embedding-guided case studies revealed local structural correspondences among proteins with limited global similarity. Applied to the efam and Metagenomic Gut Virus datasets, PhageFlare annotated 38.0% and 51.8% of proteins unresolved by Fold-seek, respectively. Furthermore, PhageFlare identified lysogenic potential in 1,866 prophage-designated genomes misclassified as lytic by BACPHLIP. These results show that PhageFlare complements structural-similarity-based annotation and expands functional discovery across viral dark matter.