4CPS-136 DeepPL guided in silico mutagenesis for precision Phage therapy in hospital pharmacy

<h3>Background and Importance</h3> Multidrug–resistant (MDR) infections are rising while antibiotic options narrow. Phage therapy offers pathogen–specific killing, yet hospital pharmacy (HP) lacks sequence–level tools to verify lytic lifestyle and to anticipate whether single–nucleotide variants (SNVs) could shift predictions, with consequences for lot release and patient–tailored cocktails. We integrated DeepPL, a nucleotide–level classifier of phage lifestyle, with an in–silico mutagenesis module to bring genome–informed, auditable decisions into routine HP practice. <h3>Aim and Objectives</h3> Aims: Classify phage lifestyle on curated GenBank datasets; Quantify robustness to SNVs via in silico mutagenesis; Deliver HP ready outputs (eligibility sheet, mutagenicity/robustness indices, and a lot release LIMS compliant dossiers); To support precision phage therapy against MDR pathogens. <h3>Material and Methods</h3> DeepPL was used for phage lifestyle classification on GenBank sequences, integrating DNABERT embeddings within a PyTorch environment. Performance metrics (AUC, accuracy, F1) were computed at thresholds 0.015–0.016. In silico mutagenesis (±5 nt windows, 2–128 SNVs/genome) and a greedy KL-filtered approach quantified Δp changes, mapping genomic regions influencing classification stability. <h3>Results</h3> DeepPL achieved AUC ~0.96 and accuracy ~0.87–0.90; the tuned threshold (~0.01508) closely matched the default 0.016. Across full predictions we observed 1,347 lytic (69%) and 598 lysogenic (31%) classes, with silhouette &gt;0.5 and intra–cluster variance &lt;1×10⁻⁴. On a subset (n=584), macro–F1 was 0.88 and accuracy 0.90. In mutagenesis tests, high–confidence genomes were stable (only one class switch across runs). Borderline genomes showed a monotonic rise in switches as SNV counts increased. A case genome (57,677 nt) yielded ~173,000 possible single–base mutations; exhaustive scanning of the first 6kb required &gt;30h and revealed localised Δp spikes, which the greedy approach recovered at a fraction of the cost. A Machine Learning–driven baseline reached AUC ~0.96, accuracy ~0.90, with GC content, genome length and number of proteins as leading features. <h3>Conclusion and Relevance</h3> Embedding DeepPL plus in–silico mutagenesis within HP enables reliable lifestyle assignment, quantitative robustness profiling, and traceable documentation for lot release and precision cocktail design against MDR pathogens–advancing antimicrobial stewardship with actionable, auditable evidence. <h3>References and/or Acknowledgements</h3> 1. Zhang Y, <i>et al</i>. DeepPL: A deep–learning–based tool for predicting phage lifecycles. <i>PLoS Comput Biol.</i> 2024. 2. Ji Y, <i>et al</i>. DNABERT: pre–trained bidirectional encoder representations for DNA sequences. <i>Bioinformatics</i> 2021. 3. WHO/Europe. Building evidence for the use of bacteriophages against antimicrobial resistance. 2024–2025. <h3>Conflict of Interest</h3> No conflict of interest