PhageAMR-Finder

Mawhiba Ibdaa 2027 · Samer Ali Alghamdi
Al-Andalus International School, Jeddah, Saudi Arabia
samerjahran@gmail.com

Most ocean phage proteins return zero BLAST hits — no known relatives, no functional annotation. Standard tools skip them entirely. This project uses ESM-2 protein language model embeddings to annotate those proteins without needing sequence similarity to anything previously described.

The pipeline is geNomad-first: viral contig classification before any functional analysis. This is a hard requirement, not a preference. Without it, 56% of the top-ranked candidates originate from non-viral contigs (measured; see Ablation below). ESM-2 classification alone is insufficient.

Try the web tool →

Raw reads → MEGAHIT assembly (92,500 contigs)
         → geNomad viral classification (3,610 phage contigs, score >0.9)
         → Protein extraction from confirmed viral contigs
         → ESM-2 v6b classification (8 classes)
         → BLAST novelty screen (full nr, PAM30 short-peptide parameters)
         → Foldseek structural search (PDB + AlphaFold DB)
         → Candidate identification and assembly verification

Classifying proteins from all 92,500 contigs without viral pre-filtering and then running Prodigal gene calling:

The 0% result is not circular. An independent flanking-gene audit of Pipeline B's top candidates confirmed that 10/10 carry viral hallmark genes or Caudoviricetes taxonomy — evidence not derived from the geNomad viral score.

The retired candidate (k99_19554_1, see below) was caught by this same flanking-gene check. Its discovery was what prompted the ablation measurement.

A 54 aa cationic membrane-disrupting peptide from a confirmed Red Sea phage contig.

Contig context: k99_98199, 26,607 bp, geNomad virus score 0.9815
Hallmark genes: terminase large subunit (gene 52), head-to-tail connecting protein (gene 53)
Partner endolysin: k99_98199_51 (SLT domain, 171 aa; BLAST matches to Citrobacter and Burkholderia phage proteins, consistent with gram-negative host range — not biochemical confirmation of substrate specificity)

Candidate protein evidence:

Assembly verification:
9,560,317 read pairs mapped back to contig k99_98199. Breadth 100%, mean depth 13.01×, median 13.0×, zero-coverage positions 0. Gene k99_98199_25 at positions 13,870–14,034 averages 14.57× with a minimum of 10×. No zero-coverage gap exists anywhere on the 26,607 bp sequence.

Genomic context:
53 genes on the contig. Zero bacterial universal single-copy marker genes (USCGs). Three viral hallmark genes. The candidate sits between two Caudoviricetes-assigned genes with 16 bp and 0 bp intergenic gaps.

Important: The candidate is a computational lead. No biological activity has been demonstrated. Novelty means no database homolog was found; it does not establish antimicrobial function. Wet-lab validation is planned pending regulatory approval.

k99_19554_1 (206 aa) was identified in an earlier pipeline version before geNomad-first validation was implemented. It returned zero BLAST hits against phage databases and showed high classification confidence. Flanking ORF analysis resolved it as bacterial CcmB: CcmA (96% identity) and CcmC (100% identity) flank it in the same Paracoccaceae genome in nr. Current BLASTp returns CcmB at 98% identity, e = 5e-134.

k99_19554_1 is bacterial CcmB, not a phage protein. It is retained in this repository as a documented false-positive case study. Its structure and associated files remain for reproducibility.

Cluster-aware leakage sweep (MMseqs2 GroupKFold, whole clusters held out):

The 30% decline is reported openly. It represents genuine generalization difficulty at extreme divergence — the regime where viral dark matter proteins actually live. It is not hidden.

Candidate-vs-training leakage: Maximum pairwise similarity between k99_98199_25 and any training sequence is 35.8%, against a synthetic repeating-pentamer spacer with no biological relevance. No biologically related training sequence was found.

8 classes: host_binding, membrane_disruption, iron_acquisition, structural, replication, regulatory, metabolic_amg, non_phage

The lower v6b F1 relative to v4 reflects harder evaluation conditions, not degraded performance.

Note: GOPhage and Empathi are serious published systems using overlapping methods. A quantitative head-to-head benchmark against these tools on a shared homology-controlled labeled set is the most important planned upgrade. The current comparison is qualitative.

No significant depth effect was detected on the genuinely novel dark-matter tier (confidence 0.85–0.99) across four ocean metagenomes (two-way ANOVA, p > 0.05). The depth gradients reported in prior virome studies appear to originate from the near-relative tier rather than from genuinely novel proteins.

Note: A non-significant ANOVA does not establish depth-invariance. It establishes that a depth effect was not detected. This result is presented as a secondary finding pending a formal equivalence analysis.

/
├── embed_and_predict.py         CLI tool — annotate proteins using v6b
├── methods_overview.md          Full methodology documentation
├── README.md
├── structures/
│   ├── k99_98199_25.pdb         Primary candidate (54 aa, pLDDT 88.225)
│   ├── k99_19554_1.pdb          Retired false positive (bacterial CcmB)
│   └── [other structures]
└── results/
    ├── dark_matter_candidates.tsv   Membrane-disruption candidates
    └── [other result files]

Model files (classifier_v6b.pkl, label_encoder_v6b.pkl) are archived at Zenodo.

pip install fair-esm scikit-learn torch joblib

Download model files from Zenodo (link above), then:

python embed_and_predict.py --fasta your_sequences.faa \
    --classifier classifier_v6b.pkl \
    --label_enc label_encoder_v6b.pkl \
    --threshold 0.50

Optional flags: --batch_size 4 (reduce if GPU OOM), --verbose (show all class probabilities)

Alghamdi, S. A. (2026). PhageAMR-Finder: ESM-2 MLP Classifier for Ocean Phage Dark Matter Functional Annotation. Zenodo. https://doi.org/10.5281/zenodo.20935067

Grade 9 · Al-Andalus International School · Jeddah, Saudi Arabia