Arid5a uses disordered extensions of its core ARID domain for distinct DNA- and RNA-recognition and gene regulation.

AT-rich interacting domain (ARID)-containing proteins, Arids, are a heterogeneous DNA-binding protein family involved in transcription regulation and chromatin processing. For the member Arid5a, no exact DNA-binding preference has been experimentally defined so far. Additionally, the protein binds to mRNA motifs for transcript stabilization, supposedly through the DNA-binding ARID domain. To date, however, no unbiased RNA motif definition and clear dissection of nucleic acid-binding through the ARID domain have been undertaken. Using NMR-centered biochemistry, we here define the Arid5a DNA preference. Further, high-throughput in vitro binding reveals a consensus RNA-binding motif engaged by the core ARID domain. Finally, transcriptome-wide binding (iCLIP2) reveals that Arid5a has a weak preference for (A)U-rich regions in pre-mRNA transcripts of factors related to RNA processing. We find that the intrinsically disordered regions flanking the ARID domain modulate the specificity and affinity of DNA binding, while they appear crucial for RNA interactions. Ultimately, our data suggest that Arid5a uses its extended ARID domain for bifunctional gene regulation and that the involvement of IDR extensions is a more general feature of Arids in interacting with different nucleic acids at the chromatin-mRNA interface.

Nowcasting US GDP Using Tree-Based Ensemble Models and Dynamic Factors.

In this study, we nowcast quarter-over-quarter US GDP growth rates between 2000Q2 and 2018Q4 using tree-based ensemble machine learning models, namely, bagged decision trees, random forests, and stochastic gradient tree boosting. To solve the ragged edge problem and reduce the dimension of the data set, we adopt a dynamic factor model. Dynamic factors extracted from 10 groups of financial and macroeconomic variables are fed to machine learning models for nowcasting US GDP. Our results show that tree-based ensemble models usually outperform linear dynamic factor models. Factors obtained from real variables appear to be more influential in machine learning models. The impact of factors derived from financial and price variables can only become important in predicting GDP after the great financial crisis of 2008-9, reflecting the effect extra loose monetary policies implemented in the period following the crisis.