SmartSepsis-Oph: Multimodal AMR variants for ophthalmology research

Curated multimodal dataset of 43 antimicrobial resistance (AMR) gene variants relevant to ocular bacterial pathogens (endophthalmitis, keratitis, perioperative prophylaxis), each annotated with DNA sequence, protein sequence, ESM-2 + ProtT5 embeddings, predicted 3D structure (PDB), structural descriptors, drug class labels and resistance mechanism.

Built as the public dataset accompanying the SmartSepsis-Oph research line at HC-FMUSP × Mass Eye and Ear (Harvard Medical School), led by Dr. Gustavo Sakuno (oculomics & multi-omics biomarkers).

Why this dataset

Existing AMR resources (CARD, AMRFinderPlus, NDARO) provide reference sequences but no aligned multimodal annotations. To train and benchmark protein-language-model classifiers, structural property predictors, and AI-driven CRISPR diagnostics targeting ocular pathogens specifically, we curated a compact set of 43 clinically prevalent variants and pre-computed every modality used in our pipeline. Inspired by the NanoFold-public distribution pattern (Hayduk 2026).

Coverage

Family	Variants	Drug class	Source
mecA	mecA1, mecA2	penam, cephalosporin, methicillin (MRSA)	RefSeq
blaKPC	KPC-2, 3, 4, 5, 11, 30, 31	carbapenem, cephalosporin, penam	RefSeq
blaNDM	NDM-1, 2, 5, 7	carbapenem, cephalosporin, penam	RefSeq
blaOXA-48	OXA-48, 181, 232	carbapenem, cephalosporin, penam	RefSeq
blaVIM	VIM-1, 2, 4	carbapenem, cephalosporin, penam	RefSeq
blaIMP	IMP-1, 6	carbapenem, cephalosporin, penam	RefSeq
blaGES	GES-1, 5	carbapenem, cephalosporin	RefSeq
blaCTX-M	CTX-M-2, 8, 9, 14, 27	cephalosporin, penam (ESBL)	RefSeq
vanA	vanA	glycopeptide (vancomycin)	RefSeq
mcr	mcr-1, mcr-1.1, mcr-5	polymyxin, peptide	RefSeq
qnrS	qnrS1, qnrS2	fluoroquinolone	RefSeq
armA	armA	aminoglycoside	RefSeq

Schema

{
    "variant_id": str,                 # "blaKPC-3"
    "gene_family": str,                # "blaKPC"
    "dna_accession": str,              # "NG_049257.1"
    "dna_sequence": str,
    "protein_sequence": str,           # longest ORF, table 11 (bacterial)
    "protein_length": int,             # AA
    "drug_classes": list[str],         # ["carbapenem", "cephalosporin", "penam"]
    "resistance_mechanism": str,       # "antibiotic inactivation"
    "esm2_embedding": list[float],     # 640d, mean-pooled
    "esm2_model": str,                 # "esm2_t30_150M_UR50D"
    "prott5_embedding": list[float],   # 1024d, mean-pooled
    "prott5_model": str,               # "Rostlab/prot_t5_xl_uniref50"
    "structure_pdb": str,              # full PDB text
    "structure_source": str,           # "ColabFold/ESMFold/AlphaFoldServer"
    "struct_length": int,              # CA atoms
    "struct_rg": float,                # radius of gyration (A)
    "struct_compactness": float,       # Rg / L^0.6
    "struct_contact_density": float,   # fraction of CA-CA pairs <8 A
    "struct_mean_plddt": float         # 0-100, prediction confidence
}

Two configs

Config	Rows	Size	What's inside
panel	45 (43 com tudo)	3.2 MB	Multimodal completo: DNA + protein + ESM-2 + ProtT5 + PDB + struct features + drug class
extended	9.034	34 MB	AMRFinderPlus catalog (8.991) + panel (43): variant_id + source + drug_classes + ESM-2 embedding

How to use

from datasets import load_dataset
import numpy as np

# Multimodal panel (curated, full pipeline)
ds_panel = load_dataset("jvlegend/smartsepsis-oph", "panel", split="train")
row = ds_panel[0]
print(row["variant_id"], row["gene_family"], row["protein_length"])
emb = np.array(row["esm2_embedding"])           # 640d
ensemble = np.concatenate([emb, np.array(row["prott5_embedding"])])  # 1664d
print("ensemble shape:", ensemble.shape)

# Extended (9034 entries from AMRFinderPlus + panel, ESM-2 + drug class)
ds_ext = load_dataset("jvlegend/smartsepsis-oph", "extended", split="train")
print(f"Extended: {len(ds_ext)} entries, {len(set(ds_ext['source']))} sources")
# -> Extended: 9034 entries, 2 sources

Dataset construction

1. Source acquisition — variants pulled from NCBI RefSeq via Entrez API (NG_* accessions curated against AMRFinderPlus / CARD ontology).
2. Translation — longest ORF using bacterial genetic code (table 11) via Biopython.
3. ESM-2 embeddings — esm2_t30_150M_UR50D, mean-pooled across residues.
4. ProtT5 embeddings — Rostlab/prot_t5_xl_uniref50, mean-pooled.
5. 3D structure prediction — combination of:
   • ESMFold via HuggingFace transformers (proteins ≤400 aa)
   • ColabFold AF2 (mcr-1, mcr-5 ~540 aa)
   • AlphaFold Server AF3 (mecA1, mecA2 ~665 aa) PDB rank_1 selected per variant.
6. Structure descriptors — Rg, compactness ratio, contact density, mean Cα-Cα, aspect ratio, mean pLDDT computed from CA coordinates.
7. Drug class / mechanism labels — derived from CARD ontology (CC BY 4.0) harmonized to clinically meaningful classes.

Companion code

Pipeline source at https://github.com/JVLegend/smartsepsis. Includes:

• CRISPR-Cas12a guide design (design_guides.py)
• Multi-label OvR classifier with NanoFold-augmented negative calibration (phenotype_probe_v2.py)
• Structure-aware ensemble (structure_features_v3.py)
• Pangenome of 21 K. pneumoniae + E. coli isolates (pangenome.sh)

Personal & sensitive information

None. All data is derived from public NCBI RefSeq sequences and predicted structures. No patient data, biological samples, or PII. Embeddings are deterministic functions of the public sequences.

Considerations for use

• Predicted structures carry inherent uncertainty (mean pLDDT ~85-95 across the panel, but local regions can be lower). Use the per-variant struct_mean_plddt for downstream weighting.
• AlphaFold Server (AF3) terms apply to the mecA1/mecA2 PDBs — research use only, no commercial redistribution. For commercial use, regenerate via ColabFold AF2 or AlphaFold 2 OpenFold.
• Class imbalance — drug class distribution mirrors clinical relevance (heavy on β-lactams, lighter on glycopeptide/polymyxin). For balanced training, augment with AMRFinderPlus catalog (8,991 sequences) referenced in the companion paper.

Citation

@dataset{smartsepsis_oph_2026,
  title  = {{SmartSepsis-Oph}: Multimodal AMR variants for ophthalmology research},
  author = {Dias, Jo{\~a}o Victor and Sakuno, Gustavo and Primo, Raul},
  year   = {2026},
  url    = {https://huggingface.co/datasets/JVLegend/smartsepsis-oph},
  doi    = {tbd},
  note   = {Dataset accompanying the SmartSepsis-Oph research line, HC-FMUSP × Mass Eye and Ear (Harvard).}
}

Authors & affiliations

• João Victor Dias — CTO & AI Architect, IA para Médicos; PhD candidate HC-FMUSP
• Dr. Gustavo Sakuno — Clinical & Scientific Lead, postdoc Harvard Medical School / Mass Eye and Ear; PhD USP — Ophthalmology & Oculomics
• Raul Primo — Software Engineer, IA para Médicos

Conflict of interest

Authors declare affiliation with IA para Médicos (project sponsor). No financial conflicts.

License

CC BY 4.0.

You are free to share and adapt for any purpose, including commercial — provided you give appropriate credit (cite as above) and indicate if changes were made.

Contact

Issues, corrections, additions: open an issue on https://github.com/JVLegend/smartsepsis or reach out via https://www.iaparamedicos.com.br/.