oncosplice

Given two (or more) mutations in the same gene, classify how their joint effect on splicing differs from the additive prediction — into one of four mutually-exclusive mechanism classes: rescue, cryptic rescue, deletion synergy, or cryptic synergy.

oncosplice is a sequence-level pipeline for splicing-epistasis analysis of single-, double-, and N-variant constructs. It runs a splice-site predictor (SpliceAI, OpenSpliceAI, Pangolin, or Spliceformer) under each variant context, computes per-site residuals against the additive expectation, and applies a crisp 4-class mechanistic classifier.

Implements the algorithms from:

1. Detecting and understanding meaningful cancerous mutations based on computational models of mRNA splicing — Lynn & Tuller, npj Systems Biology 2024.
2. Large-scale insight into missplicing, intra-gene epistasis and its relevance to human cancer — in preparation.

pip install oncosplice[spliceai_pytorch]

What it does

Given two (or more) genomic variants in the same gene, oncosplice answers:

• Single-variant impact. For each mutation alone, how much does it perturb every splice site in the gene? analyze_single().
• Joint behavior. What does splicing look like when both mutations co-occur, and how does that compare to the additive prediction? analyze_pair() / analyze_multi().
• Mechanism. Is the joint effect a synergistic gain (joint > additive), a rescue (single disrupts, joint restores WT), a compounding sub-additive stack, or just dominance / noise? Per-site and pair-level classification.
• Bulk classification. Run the same analysis over a DataFrame of hundreds of thousands of pairs with per-gene scheduling, batched inference, and resumable checkpointing. scan() / classify_dataframe().

Install

# Recommended — original SpliceAI weights, PyTorch backbone (no TF dependency)
pip install oncosplice[spliceai_pytorch]

# Or pick another engine
pip install oncosplice[openspliceai]   # OpenSpliceAI (MANE-trained, retrained)
pip install oncosplice[pangolin]       # Pangolin (40-model multi-tissue)
pip install oncosplice[spliceformer]   # Spliceformer (40k transformer)
pip install oncosplice[all]            # all 4 production engines

# Optional add-ons
pip install oncosplice[protein]        # protein-divergence score (Lynn & Tuller 2024)

Core requires numpy, pandas, matplotlib, biopython, seqmat. The classification core (analyze_pair, scan, classify_dataframe) has no geney dependency — geney is only needed for the protein-divergence score path ([protein] extra).

Model weights download automatically from the Hugging Face Hub on first use and are cached in ~/.oncosplice/weights/ — no manual step. Set ONCOSPLICE_AUTO_DOWNLOAD=0 to require an explicit oncosplice-download-weights instead (useful offline / in CI).

Highlights

• Four production engines under one interface — SpliceAI (PyTorch port, numerically identical to Keras), OpenSpliceAI, Pangolin, Spliceformer. Swap with one string. Cross-engine ensembling via ensemble:a,b,c.
• Four-class mechanistic classifier — rescue / cryptic rescue / deletion synergy / cryptic synergy, defined on probability bands with a hard WT-vs-annotation prerequisite that filters predictor noise.
• TCGA-scale runner — classify_dataframe() does per-gene grouping + batched inference + resumable checkpointing. ~23× faster than per-pair after the 3.2.0 vectorization; 800k pairs in ~22 hours on an L40S.
• Numerical parity tests between Keras SpliceAI and the PyTorch port so the migration is auditable.
• Pure-python scoring core (oncosplice.scoring) with no model dependencies — usable as a library in other splicing-prediction stacks.

Quickstart

from oncosplice import OncospliceEngine

eng = OncospliceEngine(splicing_engine="spliceai_pytorch")

# Single variant — does this mutation cause missplicing?
single = eng.analyze_single("KRAS:12:25227344:A:T")
print(single.summary())
print(single.missplicing.to_dataframe())     # missed + discovered sites
single.plot_missplicing()

# Pair — what happens when both mutations co-occur?
pair = eng.analyze_pair("KRAS:12:25227343:G:T", "KRAS:12:25227344:A:T")
print(pair.pair_classification)               # → "rescue"
print(pair.epistatic_sites())                 # only the syn/rescue/comp sites
pair.plot_case_study()                        # the bar figure

# N-variant (higher-order)
multi = eng.analyze_multi([
    "KRAS:12:25227343:G:T",
    "KRAS:12:25227344:A:T",
    "KRAS:12:25227345:G:C",
])

Bulk classification of a DataFrame

import pandas as pd
df = pd.read_csv("pairs.csv")   # column: epistasis_id (e.g. "GENE:CHR:POS:REF:ALT|GENE:CHR:POS:REF:ALT")

out = eng.classify_dataframe(
    df, epistasis_id_col="epistasis_id",
    checkpoint_path="results.csv",
)
# adds: pair_classification, max_abs_residual, max_abs_event_delta,
#       n_del_syn, n_cryp_syn, n_rescue, n_cryp_rescue, engine, error

Per-gene grouping + batched scan() underneath — typically 10–40× faster than the per-pair path on TCGA-shaped datasets. The runner is resume-safe (re-running with the same checkpoint path skips already-done pairs) and emits both per-pair and per-single CSVs.

Engine-only API (no geney needed)

from oncosplice.engines import get_predictor, list_available_engines
print(list_available_engines())

p = get_predictor("spliceai_pytorch")
pred = p.predict(padded_sequence)  # → SplicingPrediction(acceptor, donor)

The classifier — 4 mechanism classes

At every splice site, given four predicted probabilities ref, mut1, mut2, event (all in [0, 1]) and the annotation flag, we test four mutually-exclusive rules. The residual expected − event (or event − expected, depending on direction) plus the band-membership of ref, mut1, mut2, event decide the class. expected = mut1 + mut2 − ref is the additive null.

Thresholds (one set, used everywhere):

Hard prerequisite — WT prediction must agree with annotation. Every rule first checks that the engine's wild-type prediction is consistent with the annotation: annotated == True ⇒ ref ≥ HIGH, annotated == False ⇒ ref ≤ LOW. Sites where the engine disagrees with the annotation are dropped as non-epistatic without consulting the mutations. This is the noise filter.

The four rules

Anything else → non-epistatic.

Numeric example

# annotated acceptor in INPP5J — spliceai_pytorch
ref = 0.972   annotated = True
m1  = 0.658   (m1 alone preserves: 0.658 ≥ 0.50)
m2  = 0.841   (m2 alone preserves: 0.841 ≥ 0.50)
event   = 0.339
expected = m1 + m2 - ref = 0.527

# ref ≥ HIGH ✓ and annotated ✓                   → annotated branch
# min(m1, m2) = 0.658 ≥ HIGH ✓                    → not rescue (singles preserve)
# ref - event = 0.633 ≥ RES (0.10) ✓
# expected - event = 0.188 ≥ RES (0.10) ✓         → deletion_synergy
# synergy_residual = 0.188

Pair-level aggregation

A pair's overall label is the class of the splice site with the largest mechanism residual (rescue or synergy). Ties break by class priority: deletion_synergy > cryptic_synergy > rescue > cryptic_rescue > non-epistatic. The full per-site breakdown is always retained in pair.site_residuals.

Available splicing engines

Package layout

oncosplice/
├── engine.py             # OncospliceEngine — orchestrator (analyze_single/pair/multi, scan, classify_dataframe)
├── results.py            # typed dataclasses: SingleVariantResult, DoubleVariantResult, MultiVariantResult
├── variants.py           # Variant + VariantPair (no geney dependency)
├── viz.py                # plot_case_study + supporting bar figures
├── engines/              # standalone splice-site predictor adapters (uniform interface)
│   ├── base.py
│   ├── spliceai_pytorch.py
│   ├── openspliceai.py
│   ├── pangolin.py
│   ├── spliceformer.py
│   ├── spliceai_keras.py
│   └── ensemble.py
├── scoring/              # pure-Python scoring primitives
│   ├── splicing.py
│   ├── epistasis.py      # the 3-bucket classifier + vectorized residual computation
│   ├── oncosplice.py     # protein-divergence Oncosplice score
│   └── fingerprint.py    # splicing-outcome hashing
└── weights/              # weight-resolution + downloader CLI

Examples

See examples/:

• KRAS_rescue.ipynb — a canonical KRAS donor disrupted by mut1 alone, restored by the joint event. The mechanism the classifier surfaces as rescue.
• CREBBP_synergistic.ipynb — the joint event activates one cryptic acceptor (synergy) while rescuing another from each single's activation (rescue). The classifier reports the dominant synergistic call with the rescue site preserved in the per-site table.

Testing

pytest tests/                              # full suite
pytest tests/test_scoring.py               # classifier + residual rules
pytest tests/test_spliceai_equivalence.py  # Keras ↔ PyTorch numerical parity

Citing

If you use this code in a published analysis, please cite the two papers above.

License

MIT — see LICENSE.