B-256 DAEDALUS: a de novo artificial intelligence co-designed bioinformatics pipeline for infectious disease next-generation sequencing applications

Abstract

One of the primary barriers to implementing next-generation sequencing (NGS) in clinical laboratories is the perceived need for a dedicated bioinformatician and/or costly commercial analysis software. As a result, many laboratories lack the flexibility to design, validate, and adapt NGS analysis workflows tailored to their specific diagnostic needs. This proof-of-concept study describes a de novo artificial intelligence co-designed bioinformatics pipeline for infectious disease next-generation sequencing applications (DAEDALUS). This workflow addresses a critical rate-limiting step in clinical NGS adoption by enabling autonomous data analysis without reliance on specialized bioinformatics software, personnel, or high-performance computing infrastructure. A total of eight contrived clinical specimens representing bacterial, mycobacterial, and fungal infections were prepared. DNA was extracted using the Maxwell RSC instrument and Genomic DNA kit, followed by amplification with universal degenerate primers targeting 16S rRNA, hsp65, ITS, and D1/D2 regions. Amplicons were sequenced on an Illumina MiSeq platform (2×250 bp). DAEDALUS was developed through more than 100 iterative conversational prompts with a custom generative pre-trained transformer (GPT) configured to assist with bioinformatic scripting. Domain expertise guided all analytical design decisions, while the GPT provided support for Python code generation, debugging, refactoring, parameterization, and documentation. All scripts were executed within a Spyder integrated development environment on standard consumer computing hardware. The GPT-assisted, human-directed development process produced a fully functional, standalone bioinformatics pipeline capable of processing paired-end infectious disease NGS data. The pipeline underwent more than twenty major revisions, including eight complete script rewrites. DAEDALUS generated three primary outputs: (1) on-board reference-based mapping against publicly available databases, (2) representative FASTA sequences for NCBI BLAST, and (3) the most abundant single sequence for reference-independent BLAST analysis. Consolidated quality-control outputs included read counts, Q-score summaries, read-length distributions, and mapping summaries. Across eight contrived specimens, on-board mapping yielded the expected genus- and species-level identifications in 87.5% and 50% of cases, respectively, whereas BLAST analysis of the representative or most abundant sequence yielded expected species-level identification in 87.5% and 100% of cases, respectively. Taken together, these results demonstrate that complex bioinformatics workflows for infectious disease NGS analysis can be developed through iterative human-AI collaboration without reliance on specialized bioinformatics personnel or high-performance computing infrastructure. DAEDALUS illustrates the feasibility of using generative artificial intelligence as a development partner to lower technical barriers that have historically limited clinical NGS adoption. By integrating adaptive analysis logic, transparent outputs, and comprehensive documentation, this proof-of-concept pipeline supports both on-board reference-based and reference-independent sequence interpretation in a clinically relevant framework. Future refinement of DAEDALUS will focus on establishing standardized interpretive criteria with the ultimate goal of subsequent clinical validation.