README.md

---
license: cc-by-4.0
task_categories:
  - tabular-classification
  - tabular-regression
language:
  - en
tags:
  - oligonucleotide
  - hepatotoxicity
  - antisense
  - siRNA
  - GalNAc
  - in-silico
  - synthetic
  - toxicology
  - liver
size_categories:
  - 10K<n<100K
---

# OligoTox Phase 2 Dataset

A computationally generated, AI-ready, literature-informed dataset for
modeling oligonucleotide-associated hepatotoxicity. Released by DBbun
LLC as part of a submission to the NIH/NCATS OligoTox Open Data Challenge
Phase 2.

## Summary

The dataset connects oligonucleotide sequence, chemical modification
pattern, delivery platform, dose/exposure context, in vitro or
translational liver-relevant assay context, controls, and toxicity
readouts in a structured set of tables. The final dataset contains:

- **1,120 oligo records** (1,000 generated non-control oligos + 120 controls)
- **5,600 assay instances** across multiple time points and replicates
- **16,800 replicate-level toxicity readouts**
- **127 tables** comprising 8 core modeling tables (oligo metadata, aggregate chemistry, position-level modifications, biophysical, dose, assays, readouts, controls) plus per-source
  evidence modules and supporting metadata files

The dataset is intended for use in developing, training, benchmarking,
and stress-testing in silico predictive models of oligonucleotide
toxicity. It is not presented as experimentally measured wet-lab data;
all values are computationally generated, with row-level provenance
metadata distinguishing literature-grounded generated values from
inferred and reported values.

## Repository structure

```
/
├── README.md                            ← This file
├── LICENSE                              ← CC-BY-4.0 license text
│
├── oligos.csv                           ┐
├── chemical_modifications.csv           │
├── oligo_modification_positions.csv     │ 8 core modeling tables
├── biophysical_properties.csv           │ Use these directly for predictive
├── dose_exposure.csv                    │ modeling tasks
├── assays.csv                           │
├── toxicity_readouts.csv                │
├── controls.csv                         ┘
│
├── schema.json                          ┐
├── data_dictionary.csv                  │
├── data_dictionary_full.csv             │
├── dataset_manifest.json                │ Top-level documentation
├── reproducibility_manifest.json        │ and validation
├── validation_summary.csv               │
├── toxicity_readouts_calibration_audit.json  ┘
│
├── metadata/                            (11 governance/provenance tables:
│                                         assumptions, provenance, sources,
│                                         triage log, gap map, etc.)
│
├── evidence/                            (105 per-source and broad-context
│                                         evidence tables documenting
│                                         literature-grounded generation
│                                         logic for each source paper)
│
└── figures/                             (5 reference figures)
```

## How to use

```python
import pandas as pd

# Core modeling tables (root level)
oligos    = pd.read_csv("oligos.csv")
mods      = pd.read_csv("chemical_modifications.csv")
positions = pd.read_csv("oligo_modification_positions.csv")
biophys   = pd.read_csv("biophysical_properties.csv")
dose      = pd.read_csv("dose_exposure.csv")
assays    = pd.read_csv("assays.csv")
readouts  = pd.read_csv("toxicity_readouts.csv")
controls  = pd.read_csv("controls.csv")

# Governance / provenance metadata (in metadata/ subfolder)
sources    = pd.read_csv("metadata/sources.csv")
provenance = pd.read_csv("metadata/provenance.csv")

# Several tables share metadata columns (oligo_id, source_id, value_origin,
# etc.). When joining for modeling, select only the predictor columns you
# need from each secondary table to avoid duplicate-column collisions.
model_ready = (
    readouts
    .merge(assays[["assay_id", "model_system", "cell_model", "species",
                   "organoid_or_MPS", "assay_type", "replicate_count"]],
           on="assay_id", how="left")
    .merge(oligos[["oligo_id", "oligo_modality", "target_gene",
                   "sequence_5to3", "sequence_length",
                   "is_administered_oligo"]],
           on="oligo_id", how="left")
    .merge(mods[["oligo_id", "sugar_chemistry", "backbone_class",
                 "phosphorothioate_fraction", "GalNAc_conjugated"]],
           on="oligo_id", how="left")
    .merge(biophys[["oligo_id", "GC_content", "predicted_Tm_celsius",
                    "predicted_delta_G_kcal_mol",
                    "predicted_offtarget_hybridization_burden"]],
           on="oligo_id", how="left")
    .merge(dose[["oligo_id", "dose_nM", "exposure_duration_days"]],
           on="oligo_id", how="left")
)

# Restrict to administered oligonucleotides for predictive modeling
model_ready = model_ready[model_ready["is_administered_oligo"] == True]
```

A baseline RandomForest classifier on standard predictors achieves
overall accuracy of approximately 0.84 on the calibrated dataset, with
per-class F1 scores of 0.95 (low-risk), 0.74 (moderate), and 0.49
(high-risk).


## Dataset structure

### Core modeling tables

| Table | Description |
|-------|-------------|
| `oligos.csv` | Oligonucleotide identifiers, target genes, modality, 5'-to-3' sequence, sequence length, `is_administered_oligo` flag, control assignment, source role |
| `chemical_modifications.csv` | Aggregate chemistry per oligo: sugar chemistry, backbone class, phosphorothioate fraction, modification pattern, GalNAc conjugation, purity, characterization-method metadata |
| `oligo_modification_positions.csv` | Position-level chemical modifications: per-position base, sugar modification, backbone linkage, region (e.g., gapmer wing/gap, siRNA seed region), terminal conjugate, and provenance |
| `biophysical_properties.csv` | Predicted Tm, ΔG, GC content, off-target hybridization burden, sequence-derived risk fields |
| `dose_exposure.csv` | Dose/concentration, exposure duration, treatment frequency, exposure normalization |
| `assays.csv` | Model system, cell model, species/human-proxy flag, organoid/MPS status, replicate count, assay type |
| `toxicity_readouts.csv` | Replicate-level readouts: viability, ALT/AST proxy fold change, apoptosis, stress response, transcriptomic perturbation, immune activation, inflammatory context, overall risk score, hepatotoxicity label |
| `controls.csv` | Positive, negative, vehicle, and platform-specific control oligos with rationale, derived from public literature |

### Supporting documentation tables

| File | Description |
|------|-------------|
| `schema.json` | Programmatic schema describing every table, column types, value ranges, and categorical values |
| `data_dictionary.csv` | Curated variable-level definitions for the core modeling tables |
| `data_dictionary_full.csv` | Auto-generated comprehensive column-level documentation across all dataset tables |
| `dataset_manifest.json` | Full inventory of dataset files with sizes |
| `reproducibility_manifest.json` | Reproducibility configuration: pipeline parameters, seeds, calibration step description |
| `validation_summary.csv` | Automated validation checks: row counts, identifier mapping, range checks, schema consistency |
| `assumptions.csv` | Explicit list of biological and modeling assumptions |
| `provenance.csv` | Provenance category definitions |
| `sources.csv` | Literature source metadata: paper titles, source IDs, relevance, license |
| `source_triage_log.csv` | Documentation of source-level inclusion, deferral, exclusion decisions |
| `information_gap_map.csv` | Identification of fields where future experimental data would have highest value |
| `model_readiness_report.csv` | Modeling-readiness summary across required predictor and outcome fields |
| `challenge_alignment_scorecard.csv` | Mapping of dataset components to challenge judging factors |
| `toxicity_readouts_calibration_audit.json` | Full specification of the stochastic readout calibration step (random seed, per-readout sigmas, score weights, label fractions) |
| `figures/` | Five reference figures: label distribution before/after calibration, classifier performance, risk score distribution, predictor distributions, top source contributors |

### Per-source evidence modules

The dataset includes per-paper or per-context evidence tables (e.g.,
`hsd17b13_genetics_aso_translation.csv`, `pnpla3_azd2693_precision_mash.csv`,
`galnac_sirna_offtarget_rat_hepatotoxicity.csv`) that record the
literature-grounded generation logic applied for each source. These
tables document which design rationale, control logic, or readout
constraint was contributed by each source.

## Provenance

Every generated value carries a provenance label, which can take one of
five values:

- `reported_in_paper` — value reported directly in a source paper
- `extracted_from_paper` — value extracted from a source paper's tables/figures
- `inferred_from_paper` — value inferred from source-level reasoning
- `literature_grounded_generated` — value generated within
  literature-informed bounds
- `not_reported` — value not available

Toxicity readouts are stochastically calibrated with assay-specific
noise to produce realistic predictive structure for in silico modeling.
The full calibration specification is recorded in
`toxicity_readouts_calibration_audit.json`.

## Limitations

- This dataset is computationally generated; no physical oligonucleotides
  were synthesized, purified, administered, or assayed.
- The dataset is intended for in silico model development, schema
  evaluation, and benchmarking. It is not intended as direct evidence
  for regulatory decision-making.
- The hepatotoxicity label distribution (60% low / 30% moderate / 10%
  high) is a deliberate benchmark stratification chosen to produce a
  non-trivial classification problem with both common and rare strata;
  it is not a claim about wet-lab oligonucleotide toxicity prevalence.
- Of the 1,120 rows in `oligos.csv`, 1,089 represent oligos with
  explicit sequences and 31 represent context records without an
  administered oligo: vehicle/mock controls, clinical risk-score
  contexts, extracellular-vesicle delivery contexts, small-molecule
  rescue contexts, and non-administered injury-model contexts. Each
  row carries an `is_administered_oligo` boolean column for
  unambiguous filtering. Users restricting analysis to sequence-bearing
  oligos should filter on `is_administered_oligo == True` (equivalent
  to `sequence_length > 0`). Context records additionally carry an
  `oligo_modality` value that explicitly identifies their non-oligo
  nature (e.g., `vehicle_mock`, `context_not_administered_oligo`,
  `extracellular_vesicle_context_not_oligo`,
  `clinical-risk-score-context`).
- Original source PDFs are not redistributed; sources are identified
  through title-level metadata only.

## Citation

If you use this dataset, please cite:

```
DBbun LLC. OligoTox Phase 2 Dataset: A computationally generated,
literature-informed dataset for oligonucleotide-associated hepatotoxicity
modeling. Hugging Face Datasets, 2026.
```

## Related work

DBbun LLC has previously released open synthetic biomedical datasets
including the [MELD-Plus 1M cohort](https://huggingface.co/datasets/DBbun/1M_MELD_Plus_v1.0), the [UK Biobank ASCVD 10M cohort](https://huggingface.co/datasets/DBbun/1M_CIRCULATIONAHA.120.052430_v1.0), and
the [Insomnia 1M cohort](https://huggingface.co/datasets/DBbun/1M_Insomnia_Nature_SR_2018_v1.0), all available under the DBbun namespace on
Hugging Face.