9 October 2023 in Nature Communications
Designing promoters with desirable properties is essential in synthetic biology. Human experts are skilled at identifying strong explicit patterns in small samples, while deep learning models excel at detecting implicit weak patterns in large datasets. Biologists have described the sequence patterns of promoters via transcription factor binding sites (TFBSs). However, the flanking sequences of cis-regulatory elements, have long been overlooked and often arbitrarily decided in promoter design. To address this limitation, we introduce DeepSEED, an AI-aided framework that efficiently designs synthetic promoters by combining expert knowledge with deep learning techniques. DeepSEED has demonstrated success in improving the properties of Escherichia coli constitutive, IPTG-inducible, and mammalian cell doxycycline (Dox)-inducible promoters. Furthermore, our results show that DeepSEED captures the implicit features in flanking sequences, such as k-mer frequencies and DNA shape features, which are crucial for determining promoter properties.
26 September 2024 in National Science Review
Recent advancements in artificial intelligence (AI) have revolutionized our ability to model biological sequences. A new AI-driven paradigm for next-generation biological sequence design is emerging, providing new directions for sequence generation and optimization. This article introduced the application of AI in conducting digital experiments to explore sequence space and map the relationships between sequence and function. We also discussed integrating digital experiments and biological experiments with active learning strategies to improve the efficiency of AI models.
8 February 2025 in Nature Communications
Enhancers serve as pivotal regulators of gene expression throughout various biological processes by interacting with transcription factors (TFs). While transcription factor binding sites (TFBSs) are widely acknowledged as key determinants of TF binding and enhancer activity, the significant role of their surrounding context sequences remains to be quantitatively characterized. Here we propose the concept of transcription factor binding unit (TFBU) to modularly model enhancers by quantifying the impact of context sequences surrounding TFBSs using deep learning models. Based on this concept, we develop DeepTFBU, a comprehensive toolkit for enhancer design. We demonstrate that designing TFBS context sequences can significantly modulate enhancer activities and produce cell type-specific responses. DeepTFBU is also highly efficient in the de novo design of enhancers containing multiple TFBSs. Furthermore, DeepTFBU enables flexible decoupling and optimization of generalized enhancers. We prove that TFBU is a crucial concept, and DeepTFBU is highly effective for rational enhancer design.
19 February 2025 in Nature Communications
Regulatory sequences encode crucial gene expression signals, yet the sequence characteristics that determine their functionality across species remain obscure. Deep generative models have demonstrated considerable potential in various inverse design applications, especially in engineering genetic elements. Here, we introduce DeepCROSS, a generative artificial intelligence framework for the inverse design of cross-species and species-preferred 5’ regulatory sequences in bacteria. DeepCROSS constructs a meta-representation using 1.8 million regulatory sequences from thousands of bacterial genomes to depict the general constraints of regulatory sequences, employs artificial intelligence-guided massively parallel reporter assay experiments in E. coli and P. aeruginosa to explore the potential sequence space, and performs multi-task optimization to obtain de novo regulatory sequences. The optimized regulatory sequences achieve similar or better performance to functional natural regulatory sequences, with high success rates and low sequence similarities with the natural genome. Collectively, DeepCROSS efficiently navigates the sequence-function landscape and enables the inverse design of cross-species and species-preferred 5’ regulatory sequences.
21 May 2025 in Bioinformatics
Motivation: Massively Parallel Reporter Assays (MPRAs) have emerged as pivotal tools for systematically profiling cis-regulatory element activity, playing critical roles in deciphering gene regulation mechanisms and synthetic regulatory element engineering. However, MPRA experiments involve multi-step library processing procedures coupled with high-throughput sequencing. Operational errors during these complex workflows can lead to substantial resource depletion and experimental delays. Thus robust and user-friendly quality control methods are essential to minimize experimental failures and ensure reproducibility between replicates. Results: Here, we present esMPRA, an integrated quality control and analysis pipeline designed for MPRA experiments. Building on our experience in MPRA and its derivative techniques, coupled with systematic analysis of public MPRA datasets, we established standardized quality control metrics and developed a stepwise quality monitoring framework. esMPRA generates stage-specific diagnostic reports and provides experimental recommendations to avoid potential risks throughout the workflow. Designed for maximal accessibility, esMPRA features a one-line command-line interface and requires minimal bioinformatics expertise. Beyond quality assessment, the pipeline delivers processed data outputs, comprehensive analysis reports, and interface files compatible with downstream analyses, establishing an end-to-end solution for MPRA experimentation. Availability and implementation: esMPRA is released as an open-source software under the MIT license. The source code for esMPRA is available on Zenodo (DOI: 10.5281/zenodo.15362711) and GitHub (https://github.com/WangLabTHU/esMPRA/) for Linux, macOS, and Windows and is available via PyPI as esMPRA. Data for testing and reference is available via Zenodo repository at https://zenodo.org/records/15034449.