Research protocol · experiments not run · documentation and paper scaffold
Can a pretrained genomic model learn transferable splice recognition by predicting a mature transcript from an intron-containing genomic sequence, without explicit splice-site labels?
This repository specifies that experiment. It contains the hypothesis, study protocol, related work, a LaTeX manuscript skeleton, and contracts for a future implementation. It does not contain a trained splicing model, a training pipeline, or scientific results.
- Input: a continuous genomic sequence spanning one reference transcript, in transcription orientation, with introns retained.
- Target: the reference spliced transcript, including annotated untranslated regions (UTRs). The model uses the DNA alphabet internally; exported RNA uses
Uinstead ofT. - Supervision: transcript sequence only. Exon coordinates construct targets and support evaluation; they are not model inputs or auxiliary labels in the primary experiment.
- Cohort: a target of 120 species, with 40 animals, 40 plants, and 40 fungi. A balanced 12-species subset provides the first compute pilot.
- Backbone: NVIDIA's BF16-compatible Evo 2 1B checkpoint, adapted by conditional autoregressive fine-tuning. Its 8,192-token context must accommodate both genomic input and transcript output.
- Evaluation: exact transcript and splice-junction recovery, family-grouped splits, species and clade holdouts, and controlled splice-signal perturbations.
The supplied transcript endpoints make this a test of sequence processing within a known span, not discovery of transcription boundaries. A selected reference transcript is not necessarily the dominant transcript in every tissue. Pretraining exposure, reference annotation errors, and the short-gene restriction are explicit limitations.
| Document | Purpose |
|---|---|
| Hypothesis | Research question, predictions, and limits of the claim |
| Study protocol | Cohort, adaptation, controls, splits, metrics, and decision rules |
| Related work | Verified precedents and differences from this experiment |
| Source register | URLs, retrieval dates, and evidence notes |
| Interfaces | Proposed dataset, prediction, and report contracts |
| Decision log | Reasons for the initial design choices |
| Implementation roadmap | Staged work required before any substantive experiment |
| Paper | Compilable research-protocol manuscript; results are marked not run |
The format checks use Python's standard library only:
python3 scripts/check_scaffold.pyBuild the paper using a TeX installation with pdflatex, bibtex, natbib, booktabs, geometry, and hyperref:
make paperOpen paper/main.tex in a LaTeX editor for live preview. Generated PDFs, auxiliary files, data, and model weights are excluded from Git. CI checks document links and example consistency, compiles the paper, and retains the PDF as a build artifact.
examples/contracts.json contains two artificial, short sequences, including a reverse-strand example. They demonstrate coordinate conventions, retained introns, spliced targets, and UTR preservation. They are not biological training data and their example predictions are manually supplied, not model outputs.
The public development home is rewire-bio/implicit-splicing. Research design discussion and corrections are welcome. Changes to cohort selection, split rules, or outcome definitions must be logged before test-set evaluation. Do not interpret a manuscript placeholder or a software check as experimental evidence.
MIT applies to original code and configuration; CC BY 4.0 applies to original documentation and manuscript text. Data, models, and external material retain their own licences. See LICENSE.