Artificial intelligence (AI), and particularly deep learning (DL), holds transformative potential for the biological sciences, spanning applications from genetic engineering to disease diagnosis and therapeutic development. DL models trained on DNA sequences can predict cell-type–specific regulatory activity, decipher cis-regulatory grammar, prioritise genetic variants, and design synthetic DNA sequences. However, constructing and interpreting these models remains technically challenging. In this context, gReLU provides a comprehensive software framework that enables advanced sequence-modelling workflows, including data preprocessing, model development, evaluation, interpretation, variant effect prediction, and regulatory element design.
Lal and colleagues tested DL models’ capacity to learn cis-regulatory codes from DNA sequences within biologically relevant contexts. gReLU integrates diverse architectures, including convolutional networks and transformers, to enhance modelling flexibility. Whereas traditional pipelines require additional software to compare outputs from different models, gReLU offers an automated harmonisation layer that streamlines this process. Moreover, while conventional approaches often necessitate retraining models for each new target, gReLU’s prediction-conversion layers allow the derivation of novel outputs from existing model predictions, providing a customisable and scalable solution. In addition, gReLU incorporates directed evolution and gradient-based optimisation methods into an accessible, fully integrated platform that unifies training and interpretability tools thus consolidating state-of-the-art methodologies within a single, cohesive system.
Benchmarking experiments revealed several unexpected yet compelling findings underscoring gReLU’s superiority over alternative models. In the first evaluation, a regression model was trained to predict DNase-seq signals in GM12878 cells. The resulting dsQTL (DNase-seq Quantitative Trait Loci) effect scores achieved an AUPRC of 0.27, outperforming the widely used gkmSVM model. In the second evaluation, the authors examined the possibility of inferring dsQTL effects de novo. dsQTLs are known to have bad effects on transcription factors (TFs). They do their job by overlapping with TF motifs, which weakens the integrity of the motifs and lowers the TF binding affinity. This reduction subsequently alters chromatin accessibility, producing a quantifiable dsQTL effect. gReLU went beyond mathematically modelling this mechanism; it empirically identified that the rs10804244 variant weakens a motif for an Interferon Regulatory Factor (IRF). In a third evaluation, using the Borzoi model, gReLU successfully visualised the function of an enhancer regulating the PPIF gene and accurately predicted mutational effects with high agreement with experimental data. Furthermore, through directed-evolution functions, gReLU designed a 20-base modification within the enhancer that increased gene expression by 41.76% in monocytes and 16.75% in T cells, demonstrating the feasibility of eliciting cell-type–specific regulatory outcomes.
In summary, gReLU provides a robust and cohesive framework that streamlines the training, implementation, and analysis of genomic deep learning models, while also allowing for the creation of intricate, biologically informed regulatory components. Its fully open-source nature, receptiveness to community feedback, and accessible training materials collectively suggest that gReLU may pave the way for long-term innovation in the field of bioinformatics.
-Bioinfocodes Scientific News Service-
News articles prepared by our team members, reviewing and compiling scientific research published