variant_id large_stringlengths 4 14 | source large_stringclasses 2
values | drug_classes listlengths 1 12 | esm2_embedding listlengths 640 640 | esm2_model large_stringclasses 1
value |
|---|---|---|---|---|
AAA16360.1 | AMRFinderPlus | [
"stx2",
"stxa2b"
] | [
0.01722903922200203,
-0.09252014011144638,
-0.07549464702606201,
0.059299152344465256,
-0.042918842285871506,
-0.068768210709095,
0.03019370324909687,
-0.13185401260852814,
-0.08644857257604599,
0.1332593560218811,
0.05099514126777649,
-0.021565251052379608,
0.008430372923612595,
0.0869032... | esm2_t30_150M_UR50D |
AAA16361.1 | AMRFinderPlus | [
"stx2",
"stxb2b"
] | [
0.019734010100364685,
-0.06910474598407745,
0.03391232341527939,
-0.02611152082681656,
0.08625944703817368,
-0.017418572679162025,
0.05089410021901131,
-0.041286665946245193,
-0.08087063580751419,
-0.0005570403300225735,
0.05017893761396408,
-0.07959882915019989,
0.0033753805328160524,
-0.... | esm2_t30_150M_UR50D |
AAA16362.1 | AMRFinderPlus | [
"stx2",
"stxa2c"
] | [
-0.05587009713053703,
-0.0734146386384964,
-0.003063344396650791,
0.007302338723093271,
-0.02557326667010784,
-0.0993405357003212,
-0.04480775445699692,
-0.10157107561826706,
-0.12204278260469437,
0.14171960949897766,
0.05503581091761589,
0.04586479440331459,
-0.022356316447257996,
0.13243... | esm2_t30_150M_UR50D |
AAA19189.1 | AMRFinderPlus | [
"stx2",
"stxa2e"
] | [
-0.07340193539857864,
-0.024153249338269234,
0.011484120972454548,
0.02947269193828106,
-0.042672984302043915,
-0.11891087144613266,
-0.0353582501411438,
-0.14217311143875122,
-0.13003107905387878,
0.17758770287036896,
0.10515487194061279,
0.08033274859189987,
-0.005810432136058807,
0.1677... | esm2_t30_150M_UR50D |
AAA19190.1 | AMRFinderPlus | [
"stx2",
"stxb2e"
] | [
0.01203994546085596,
-0.05248701572418213,
0.027401963248848915,
-0.07396135479211807,
0.10458164662122726,
-0.02562444470822811,
0.053799256682395935,
-0.04223468527197838,
-0.08169803023338318,
-0.022478343918919563,
0.0483599491417408,
-0.10651002079248428,
0.01983652077615261,
-0.04728... | esm2_t30_150M_UR50D |
AAA19623.1 | AMRFinderPlus | [
"stx2",
"stxa2"
] | [
-0.0517878383398056,
-0.06990651786327362,
0.0027774276677519083,
0.004556042142212391,
-0.02698397822678089,
-0.10339666157960892,
-0.0459853857755661,
-0.11709664016962051,
-0.11646304279565811,
0.1385488510131836,
0.05034836381673813,
0.04641200602054596,
-0.02132129669189453,
0.1491189... | esm2_t30_150M_UR50D |
AAA21094.1 | AMRFinderPlus | [
"arsenic",
"arsenite"
] | [
-0.05045432224869728,
-0.09815092384815216,
-0.04958828538656235,
-0.06403839588165283,
0.09926477819681168,
-0.00020558388496283442,
0.0008873174665495753,
-0.028984973207116127,
-0.13889092206954956,
0.045846644788980484,
0.047852348536252975,
0.01844688318669796,
-0.048404011875391006,
... | esm2_t30_150M_UR50D |
AAA21095.1 | AMRFinderPlus | [
"arsenic",
"arsenite"
] | [
-0.015300950966775417,
-0.08012813329696655,
0.01925826631486416,
-0.05524183064699173,
0.13841506838798523,
0.031522661447525024,
0.027081826701760292,
0.06405065953731537,
-0.03654347360134125,
0.1653694063425064,
0.16235829889774323,
0.013107983395457268,
-0.00422894861549139,
0.0769883... | esm2_t30_150M_UR50D |
AAA21096.1 | AMRFinderPlus | [
"arsenic",
"arsenate"
] | [
0.025078918784856796,
-0.0915963351726532,
-0.1141354963183403,
0.06740514934062958,
0.16894759237766266,
-0.011682865209877491,
0.01881820149719715,
-0.04658827185630798,
-0.07139265537261963,
0.09539720416069031,
0.07312611490488052,
0.0637715682387352,
0.039820097386837006,
-0.050606533... | esm2_t30_150M_UR50D |
AAA24632.1 | AMRFinderPlus | [
"stx2",
"stxa2f"
] | [
-0.013581555336713791,
-0.059965427964925766,
0.01439567469060421,
-0.06786220520734787,
-0.02411963976919651,
-0.061510175466537476,
-0.08049574494361877,
-0.04027355834841728,
-0.0024659393820911646,
0.18476444482803345,
-0.002669238718226552,
0.031020041555166245,
-0.03364850953221321,
... | esm2_t30_150M_UR50D |
AAA24633.1 | AMRFinderPlus | [
"stx2",
"stxb2f"
] | [
0.021590903401374817,
-0.047164518386125565,
0.02436019666492939,
-0.0741935670375824,
0.09768541902303696,
-0.042484574019908905,
0.0335003025829792,
-0.054831791669130325,
-0.07142899185419083,
-0.046316929161548615,
0.07593265920877457,
-0.07999462634325027,
0.04665587469935417,
-0.0419... | esm2_t30_150M_UR50D |
AAA25276.1 | AMRFinderPlus | [
"cadmium"
] | [
-0.03911568596959114,
-0.008377430960536003,
-0.05601844564080238,
-0.07876842468976974,
0.10477006435394287,
0.09338221698999405,
-0.029942385852336884,
-0.015400922857224941,
-0.07187719643115997,
0.015733959153294563,
0.025326009839773178,
0.07321316003799438,
-0.021180009469389915,
-0.... | esm2_t30_150M_UR50D |
AAA25636.1 | AMRFinderPlus | [
"arsenic"
] | [
-0.06029549613595009,
-0.060961220413446426,
-0.1431986689567566,
-0.06938546895980835,
0.23007620871067047,
0.02589522674679756,
-0.04323003068566322,
0.017489802092313766,
-0.026190945878624916,
-0.02826480194926262,
0.019108852371573448,
0.02743617445230484,
0.12048082053661346,
-0.0244... | esm2_t30_150M_UR50D |
AAA25637.1 | AMRFinderPlus | [
"arsenic",
"arsenite"
] | [
-0.033662330359220505,
-0.0326121523976326,
0.0020357288885861635,
-0.07369742542505264,
0.16787436604499817,
0.06178465113043785,
-0.00920663308352232,
0.00518846558406949,
0.025162845849990845,
0.1807299107313156,
0.14728213846683502,
0.0009768352610990405,
0.023589204996824265,
0.017100... | esm2_t30_150M_UR50D |
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
- Source acquisition — variants pulled from NCBI RefSeq via Entrez API (NG_* accessions curated against AMRFinderPlus / CARD ontology).
- Translation — longest ORF using bacterial genetic code (table 11) via Biopython.
- ESM-2 embeddings —
esm2_t30_150M_UR50D, mean-pooled across residues. - ProtT5 embeddings —
Rostlab/prot_t5_xl_uniref50, mean-pooled. - 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.
- Structure descriptors — Rg, compactness ratio, contact density, mean Cα-Cα, aspect ratio, mean pLDDT computed from CA coordinates.
- 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_plddtfor 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
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/.
- Downloads last month
- 25