DBbun/oligotox-phase2-dataset

OligoTox Phase 2 dataset

An open dataset for modeling oligonucleotide toxicity, combining measured human outcomes and formula-verified chemistry with a clearly separated simulation layer.

Prepared by DBbun LLC for the NIH/NCATS Oligonucleotide Toxicity (OligoTox) Open Data Challenge.

Version 2.0 supersedes version 1.0. Version 1.0 contained only computationally generated data and carried chemistry errors that are documented and corrected below. It remains reachable at commit 18dcb004b27f973b5ff594a064be72099f1b1efe for anyone who needs it, but should not be used.

Two layers, never mixed

Every row carries a data_layer field. Nothing requires you to guess which kind of value you are looking at.

measured_real — retrieved, not generated

Table	Rows	What it is
measured/clinical_adverse_events.csv	70,385	Observed adverse-event counts per trial arm and MedDRA term, from 154 registered trials covering 31 products
measured/oligo_chemistry_verified.csv	73	Real sequences with per-position sugar and backbone chemistry
measured/oligo_modification_positions_verified.csv	2441	Position-level chemistry for those structures
measured/oligo_characterization.csv	19	Molecular formulas, masses and derived linkage counts from FDA labelling
measured/drug_annotation_verification.csv	31	Chemistry annotations checked against published formulas
metadata/inn_corpus_index.csv	149	Every oligonucleotide INN located in the WHO lists scanned, with extraction status; products named only in regulatory labelling, such as imetelstat, are in the chemistry tables rather than here

computationally_generated — a simulation, not evidence

1,120 oligo records, 5,600 assay instances and 16,800 replicate-level readouts across five time points. These are not measurements and must not be cited as evidence of toxicity for any molecule.

How the measured chemistry is verified

Every structure is validated arithmetically rather than by assertion. The parsed sequence, per-position sugar assignment and linkage pattern are used to recompute the molecular formula, which must equal the formula published independently in the same source. Structures that fail are excluded, not corrected.

Worked example — inotersen, a 5-10-5 2'-MOE gapmer:

sequence                  TCTTGGTTACATGAAATCCC
computed, free acid       C230H318N69O121P19S19
stated on the FDA label   C230H318N69O121P19S19   ✓
computed, sodium salt     C230H299N69Na19O121P19S19
stated on the FDA label   C230H299N69Na19O121P19S19   ✓

You can re-run this yourself: python code/formula_checksum.py.

Toxicity coverage in the measured layer

Toxicity	Records	Trials	Drugs
immunotoxicity	1,617	101	29
chronic neurotoxicity	1,356	129	31
kidney toxicity	1,245	94	27
coagulopathy	1,166	96	27
hepatotoxicity	1,080	87	28
thrombocytopenia	288	47	19
hydrocephalus	41	9	6
complement activation	15	4	3

Generated layer: chemistry classes

Class	Records	Length window
ASO	312	18–25 nt
CpG_ODN	5	20–22 nt
aptamer	5	27–40 nt
gapmer	146	16–20 nt
miRNA_mimic	11	21–23 nt
other_or_context_record	77	18–25 nt
sgRNA	4	20–20 nt
siRNA	560	19–23 nt

Quick start

import pandas as pd

ae = pd.read_csv("measured/clinical_adverse_events.csv")

# Do NOT sum counts across MedDRA terms. An arm's at-risk count is repeated on
# every one of its term rows, so summing inflates the denominator once per term
# and double-counts anyone with more than one qualifying event. Use code/ae_rates.py.
import sys; sys.path.insert(0, "code")
from ae_rates import hepatic, grouped_lower_bound

terms = hepatic(ae)                      # one exact rate per arm per term
by_drug = grouped_lower_bound(terms, ["drug"])   # lower bound per drug
print(by_drug.sort_values("pct_lower_bound", ascending=False).head(10))

OligoTox_Phase2_Optional_Analysis_Notebook.ipynb loads every table, runs the validation checks and trains a baseline classifier.

Read this before drawing conclusions

• Pooled event rates across trials are a blunt instrument. Pooling weights by trial size, and two very large cardiovascular programs (inclisiran, zilebesiran) dominate the siRNA class. Within that same class fitusiran reaches 18.1% and cemdisiran 5.7%, above most antisense drugs. Variation within a chemistry class exceeds the difference between classes, and indication is confounded with chemistry throughout. Per-trial analysis is the correct use.
• The arm_is_comparator flag is derived from free-text arm titles and is approximate.
• Only 73 structures are formula-verified, of 149 nonproprietary names located. GalNAc-conjugated siRNAs are among them: the tri-antennary cluster is modelled explicitly, so 33 of them reconstruct element for element. What still resists the check is a conjugate the parser does not model - eplontersen is the one such molecule here, and it carries the narrower phosphorothioate-count check instead - together with entries whose naming notation is not yet implemented. metadata/inn_corpus_index.csv records what blocks each remaining molecule.
• Complement activation is thinly covered because trials rarely report it as a discrete term.

Corrections made in version 2.0

Issue in v1.0	Fix
482 of 1,120 DNA-based oligos written with uracil instead of thymine	Correct alphabet, recorded per record in sequence_alphabet
Lengths drawn uniformly 13–30 nt regardless of modality	Modality-appropriate windows
375 generated sequences labeled as literature-derived	Relabeled to state that a source informed the design, not the sequence
65 free-text modality strings, 154 free-text control labels	Explicit taxonomies, original strings retained
A patisiran annotation contradicted by its own molecular formula	Corrected, caught by the formula cross-check
Calibration, position table and packaging existed in no script	Reconstructed; the dataset now rebuilds from one command

Sources

Source	Used for	Basis
ClinicalTrials.gov	Adverse-event counts	US Government work, public domain
FDA labelling (DailyMed)	Chemistry, formulas, characterization	US Government publication; extracted facts only
WHO INN lists	Sequences, position-level chemistry, CAS	Chemical composition of named substances is fact; no WHO text or document reproduced

No source PDF, label text or list document is redistributed. Each record carries a resolvable identifier pointing back to its source.

License

• Data and documentation — Creative Commons Attribution 4.0 International (CC BY 4.0). See LICENSE.
• Code in code/ — MIT. See LICENSE-CODE.

Both layers are reproducible. code/build_v2_dataset.py rebuilds the whole package from one command; code/OligoTox_Master_Dataset_Generator_v0_62.py produces the generated layer under the seeds recorded in reproducibility_manifest.json and toxicity_readouts_calibration_audit.json.

Archival copy and citation

A fixed copy of this release is deposited at Zenodo under a persistent identifier, so it remains citable and retrievable independently of this platform:

DOI: 10.5281/zenodo.23551008

The Zenodo deposit and this repository hold the same v2.0 release. Use the DOI when citing; use this repository to browse or load the tables directly.

@dataset{dbbun_oligotox_phase2_2026,
  title     = {OligoTox Phase 2: formula-verified oligonucleotide chemistry
               joined to clinical adverse-event outcomes, with a labeled
               simulation layer},
  author    = {Kartoun, Uri},
  year      = {2026},
  publisher = {Zenodo},
  version   = {2.0},
  doi       = {10.5281/zenodo.23551008},
  url       = {https://doi.org/10.5281/zenodo.23551008},
  note      = {DBbun LLC. Prepared for the NIH/NCATS OligoTox Open Data
               Challenge. Also at
               https://huggingface.co/datasets/DBbun/oligotox-phase2-dataset}
}

Contact: Uri Kartoun, PhD — uri.kartoun@dbbun.com — ORCID 0000-0003-0988-8037