SeqTrainer aims to become a "Keras for synthetic biology" in spirit: easy domain APIs, composable defaults, and clear extension points.
SeqTrainer should own the domain layer:
- SynBioHub/SBOL retrieval and interoperability
- SPARQL recipes for common synbio data patterns
- Dataset recipe abstractions and provenance
- DNA-aware transforms and feature extraction
- Task-focused applications (e.g., promoter regression)
SeqTrainer should not replace model ecosystems:
- Keras remains a high-level training interface
- PyTorch remains native for HF/DNABERT and graph ecosystems
- Framework-neutral core first
- Optional framework adapters
- Backbone + head composition
- Pluggable application recipes
- Strong provenance and reproducibility story over time
- Expand SynBioHub client auth + pagination support
- Curate reusable SPARQL recipe library for SBOL entities
- Add robust dataset caching/snapshot materialization
- Add production-grade torch/keras model factories
- Promote graph prototype scripts into stable graph APIs