Team Delta @ Phagos x AWS Hackdays 2025
This repository contains the codebase, notebooks, and models developed during the Phagos x AWS Hackdays 2025 by Team Delta (Train Tune Deploy). Our goal was to identify bacteriophage proteins with strong lytic activity, leveraging LLMs, transfer learning, and protein embeddings, with a focus on Klebsiella-targeting phages.
📄 Click here to read our slides presenting the full story behind our project.
Get a concise overview of our methodology, data insights, model architecture, saliency maps, and biological validation — all in one place.
Given a collection of phages isolated to target an unknown bacterial infection (e.g., Klebsiella), how can we:
- Annotate phage proteins with functional labels (lysis, replication, etc.)?
- Select candidates with strong lytic activity?
- Transfer knowledge from annotated datasets (like ESKAPE) to less-curated sets (like PHL-Klebsiella)?
- Use ESM-C, a protein language model with 6B parameters, to embed amino acid sequences.
- Train a feed-forward classifier to predict 14 functional protein classes, including Holins and Endolysins (key lysis proteins).
- Visualize important residues via saliency maps.
- Apply the model to the PHL-Klebsiella dataset (Boeckaerts et al., 2024) and validate against biological assays.
- Use entropy filtering to identify strong candidate proteins per phage.
explore-eskape.ipynb— Overview of the ESKAPE dataset (14k annotated proteins).explore-phl-klebsiella.ipynb— Exploration of the Boeckaerts Klebsiella-phage interaction dataset.explore-inphared.ipynb— Optional exploration of the INPHARED dataset (for future extension).
Classifier_trained__over_Eskapet.ipynb— Classifier training on ESKAPE embeddings.t-SNE.ipynb— 2D visualization of protein embeddings by functional class.transfer_learning_over_Boeckaerts.ipynb— Entropy-based ranking of predicted proteins for PHL-Klebsiella phages.
- ESM-C (Meta AI, 2024) — Pretrained transformer model for protein sequence embeddings.
- PyTorch + scikit-learn — For model training and evaluation.
- AWS SageMaker — Infrastructure for scalable training and deployment (with GPU support).
- Jupyter Notebooks — For interactive analysis and reproducibility.
- Classifier trained on ESKAPE achieved ~99% accuracy across 14 functional classes.
- Saliency maps revealed meaningful amino acid positions related to lysis.
- Transfer predictions over PHL-Klebsiella revealed 4/8 top pipeline scorers overlapping with top biological assays.
- Add phage-host affinity prediction.
- Extend functional labels using protein ontologies.
- Evaluate null model for strength-of-signal validation.
- Boeckaerts et al. (2024) — PHL-Klebsiella Dataset.
- ESM-C (Meta AI, 2024) — 6B parameter protein language model.
- Phagos x AWS Hackdays 2025 — Event Info
- Antoine Aragon
- Emmanuel de Bézenac
- Francesco Camaglia
- Mathieu Crilout
For questions or collaboration: hackathon2025@phagos.org