barneyhill/aso-atlas-2

ASO Atlas 2.0

A dataset of patent-reported antisense oligonucleotide (ASO) experiments for studying in vitro activity, dose response, hepatotoxicity and neurotoxicity across the preclinical pipeline. It is designed for research benchmarking, exploratory modelling and analysis of the published ASO assay landscape.

At a glance

• 295,007 assay readouts from 606 USPTO patent filings
• 165,782 unique, chemistry-resolved ASOs represented in HELM notation
• 430 named target genes
• Four ready-to-load configurations covering in vitro efficacy and in vivo safety
• Patent-grouped cross-validation folds for leakage-controlled benchmarking

Quick start

from datasets import load_dataset

dataset = load_dataset(
    "barneyhill/aso-atlas-2",
    "in_vitro_inhibition",
    split="full",
)
print(dataset[0])

Replace in_vitro_inhibition with dose_response, hepatotoxicity or neurotoxicity to load another assay type.

What's included

Each configuration is a separate table because the experimental endpoints and schemas differ. Together they cover four stages commonly encountered in preclinical ASO evaluation:

Config	Readouts	Model-eligible readouts	Unique HELM compounds	Named target genes
in_vitro_inhibition	174,459	173,286	161,434	389
dose_response	99,821	97,236	16,375	130
hepatotoxicity	15,831	15,427	1,881	42
neurotoxicity	4,896	4,892	3,312	15
Total	295,007	290,841	165,782	430

Of the 295,007 readouts, 290,841 have a resolved HELM annotation. The remaining 4,166 retain a reported assay outcome but cannot be used directly by chemistry-aware models.

Benchmark folds

The folds/ directory provides frozen, endpoint-specific five-fold assignments for reproducible evaluation with OligoGym and OligoAI. The splits group observations by source patent to reduce train-test leakage.

Each model-ready endpoint was constructed separately before patent-level GroupKFold was applied. A patent's fold number can therefore differ between endpoints and between OligoGym and OligoAI. In each file, rows assigned to fold k are the test set for run k; the other four folds form the training set. See folds/README.md for schemas and fold sizes.

Versioning and reproducibility

This card describes ASO Atlas 2.0. The data tables, generated counts and machine-readable metadata are released together. The source and release-building code is available in the ASO Atlas 2 repository, and croissant.json provides Croissant 1.1 core + RAI metadata. The dataset has a persistent Hugging Face DOI: 10.57967/hf/8687.

Data Fields

Common fields (all configs)

Field	Type	Description
USPTO ID	string	Source patent identifier (e.g., US20150011212)
Table Number	int	Table number within the patent
Compound ID	int	Compound identifier within the patent
HELM Annotation	string, nullable	Full chemical structure in HELM notation; null where chemistry could not be resolved
model_eligible	bool	True when a resolved HELM annotation permits chemistry-aware model input
target_RNA	string	Target gene/transcript name

in_vitro_inhibition / dose_response

Field	Type	Description
cell_line	string	Cell line used (e.g., HepG2, A549)
dosage_nm	float	Dose in nanomolar
Inhibition_pct	float	mRNA inhibition percentage (-1000 to 100)
cells_per_well	float	Cells seeded per well
cell_line_species	string	Species of origin (human, mouse, etc.)
transfection_method	string	Delivery method (electroporation, lipofection, etc.)
treatment_period_hrs	float	Treatment duration in hours
ccle_cell_line_name	string	Matched CCLE cell line name
ccle_model_id	string	DepMap model ID
ccle_oncotree_lineage	string	Oncotree tissue lineage
ccle_oncotree_disease	string	Oncotree disease classification
cell_line_mapping_source	string	How the CCLE match was determined

hepatotoxicity

Field	Type	Description
species	string	Test species (mouse, rat, monkey, dog)
species_strain	string	Strain (e.g., CD-1 mice, Sprague-Dawley rats)
dosage_mg_per_kg	float	Dose per administration (mg/kg)
num_doses	float	Number of doses administered
dosing_period_days	float	Duration of dosing period
adminstration_method	string	Route (subcutaneous, intraperitoneal)
measurement_source	string	Sample type (plasma, urine)
ALT	list[float]	Alanine aminotransferase values (IU/L)
AST	list[float]	Aspartate aminotransferase values (IU/L)
ALB	list[float]	Albumin values (g/dL)
BUN	list[float]	Blood urea nitrogen values (mg/dL)
CREA	list[float]	Creatinine values (mg/dL)
TBIL	list[float]	Total bilirubin values (mg/dL)
PC_ratio	list[float]	Protein/creatinine ratio

neurotoxicity

Field	Type	Description
species	string	Test species
species_strain	string	Strain
dosage_ug	float	Dose in micrograms
FOB_score	list[float]	Functional observation battery scores (0-7)
latency_time_hours	float	Time from dosing to observation
administration_method	string	Route (typically intracerebroventricular)
tolerability_score_type	string	Scoring system used

HELM Chemistry Encoding

HELM (Hierarchical Editing Language for Macromolecules) encodes full oligonucleotide chemistry including sugar modifications, backbone linkages, and base modifications.

Example: RNA1{{[moe](C)[sp].[moe](T)[sp].d(A)[sp].d(G)[sp].[cet](A)}}$$$$

Sugar	Chemistry	Backbone	Chemistry
d	2'-deoxy (DNA)	[sp]	Phosphorothioate
[moe]	2'-O-methoxyethyl	.	Phosphodiester
[cet]	Constrained ethyl	[am]	Phosphoramidate
[lna]	Locked nucleic acid
[fR]	2'-fluoro
[m]	2'-O-methyl

Dataset Creation

Source

All data was extracted from publicly available USPTO patent filings using an LLM-based text mining pipeline. Patents were identified via keyword search for antisense oligonucleotide-related terms, and data tables within each patent were parsed to extract compound structures, assay conditions, and measured endpoints.

Cell-line enrichment uses Model.csv from DepMap Public 25Q3, downloaded on 20 January 2026 (source SHA-256 9dbb9de8805696c1345816ab07edd23fb4fd95e117739f3c5c3b1cf062c1233b). Only cell-line names, model identifiers and OncoTree fields derived from that lookup are included here; no patient-level source fields are released.

Source terms and attribution

• USPTO patent full-text XML is governed by the USPTO website terms. Patent text and drawings are typically not copyright-restricted, subject to the exceptions stated there; source patent identifiers are retained on every record.

• DepMap Public 25Q3 is attributed to Broad DepMap and its source release terms. Broad-generated DepMap data is distributed under CC BY 4.0. ASO Atlas republishes only derived cell-line mapping fields.

• The processed ASO Atlas 2.0 release is distributed under CC BY 4.0. Dataset publication does not grant permission to practise inventions claimed by active patents.

Processing

1. HELM quality filtering: removed sequences with uncertain chemistry (?), length <11 nucleotides, missing DNA gap, homopolymers, or naked DNA
2. Gene name standardisation via manual curated mappings
3. Cell line standardisation and CCLE (Cancer Cell Line Encyclopedia) enrichment
4. Inhibition values filtered to [-1000%, 100%]
5. Hepatic biomarker values filtered to physiologically plausible ranges
6. FOB scores validated to 0-7 range
7. Deduplication by compound and measurement
8. Rows without a resolvable HELM annotation are retained and marked model_eligible=false

Intended uses

The established uses are research benchmarking and exploratory model development for patent-derived ASO in vitro efficacy and preclinical toxicity endpoints. The supplied patent-grouped folds support reproducible, leakage-controlled evaluation. The measurements represent outcomes reported in patents, not independently replicated biological truth.

This dataset is not validated for clinical decision-making, prediction of human safety, dose selection, treatment selection, prospective laboratory deployment, legal conclusions, or generalisation to chemistries and organisations absent from the source patents.

Considerations

Biases

• Patent bias: Data comes exclusively from pharmaceutical patent filings, which over-represent compound classes pursued by industry (predominantly 2'-MOE gapmers). Academic and early-stage chemistries may be underrepresented.
• Single-originator bias: All filings come from one organisation, so quantities estimated from this corpus (pass rates, enrichment factors) bound to its chemistry and target choices rather than to ASO development in general.
• Species bias: Mouse and rat data dominate; monkey and dog data are sparse.
• Reporting bias: Patents report results selectively. Negative results may be underrepresented.

Limitations

• No independent validation: All measurements originate from patent applicants' own experiments. Cross-laboratory reproducibility is unknown.
• Extraction errors: LLM-based extraction may introduce errors. A validation audit is included in the accompanying code repository.
• Heterogeneous protocols: Assay conditions (cell lines, doses, timepoints) vary across patents and are not standardised.
• Unresolved chemistry: 4,166 assay readouts lack a resolvable HELM annotation and cannot be used directly in chemistry-aware models.

Ethical considerations

No human-subject records, patient-level health information or personally identifiable information are included. The release contains public patent assay records, compound structures, animal-study measurements and public CCLE cell-line identifiers. It contains no synthetic assay records: LLMs transformed source tables but were not permitted to generate measurement values directly.

Social impact

Potential benefits include more reproducible ASO model evaluation, lower screening costs and reduced animal use. Potential harms arise if extraction errors or the source patents' narrow chemistry and target distribution are treated as universally representative; this could misprioritise candidates or research areas. Version pinning, source identifiers, leakage-controlled folds, a validation audit and the non-clinical-use limits above reduce but do not eliminate those risks.