Type 1 polyketides are a major class of natural products used as antiviral, antibiotic, antifungal, antiparasitic, immunosuppressive, and antitumor drugs. Analysis of public microbial genomes leads to the discovery of over sixty thousand type 1 polyketide gene clusters. However, the molecular products of only about a hundred of these clusters are characterized, leaving most metabolites unknown. Characterizing polyketides relies on bioactivity-guided purification, which is expensive and time-consuming. To address this, we present Seq2PKS, a machine learning algorithm that predicts chemical structures derived from Type 1 polyketide synthases. Seq2PKS predicts numerous putative structures for each gene cluster to enhance accuracy. The correct structure is identified using a variable mass spectral database search. Benchmarks show that Seq2PKS outperforms existing methods. Applying Seq2PKS to Actinobacteria datasets, we discover biosynthetic gene clusters for monazomycin, oasomycin A, and 2-aminobenzamide-actiphenol.
25 August 2025 in Nature Structural & Molecular Biology
Visualization plays an important role in extracting insights from complex and large-scale genomics data. Traditional graphical user interfaces (GUIs) offer limited flexibility for custom visualizations. Our prior work, Gosling, enables expressive visualization creation using a grammar-based approach, but beginners may face challenges in constructing complex visualizations. To address this, we explore multimodal interactions, including sketches, example images, and natural language inputs, to streamline visualization creation. Specifically, we customize two deep learning models (YOLO v7 and GPT3.5) to interpret user interactions and convert them into Gosling specifications. A workflow is proposed to progressively introduce and integrate multimodal interactions. We then present use cases demonstrating their effectiveness and identify challenges and opportunities for future research.