Skip to content

roadmap.md

Latest commit

 

History

History
63 lines (49 loc) · 3.51 KB

roadmap.md

File metadata and controls

63 lines (49 loc) · 3.51 KB
Raw

Roadmap and design ideas

The most valuable next step is not adding more aliases; it is expanding chemical scope while keeping every transformation explicit and testable.

Near term: make the current core production-ready

  • Add golden structures for every supported PTM from curated public records.
  • Add per-definition provenance fields and checksums to the custom modification registry.
  • Emit a provenance manifest containing package/RDKit/SELFIES versions and command arguments.
  • Add gzip input, iterator-based million-row processing, chunked output, and optional multiprocessing.
  • Add configurable salt stripping, fragment selection, normalization, tautomer, and charge policies.
  • Add SVG/PNG depictions and similarity/Tanimoto commands.
  • Publish wheels, API documentation, a changelog policy, and a PyPI release workflow.

Biomolecule representation

  • Implement the ProForma 2.0 grammar separately from structure resolution. A valid ProForma mass annotation is not automatically enough information to construct a molecular graph.
  • Map additional PSI-MOD/Unimod terms only after site-aware structures and regression fixtures are available; keep the JSON registry as the escape hatch for laboratory-specific chemistry.
  • Integrate the public OpenHELM monomer library and record its version/checksum. This enables non-natural residues and linkers without hard-coding hundreds of templates.
  • Represent polymers as an internal residue/atom graph before rendering SMILES, HELM, or SDF. This is required for cyclic peptides, disulfides, branched peptides, stapling, and cross-links.
  • Add explicit ambiguity objects rather than selecting a localization silently.
  • Add DNA/RNA FASTA flavors only after terminal cap and phosphate conventions are exposed in the API.
  • Support PEFF for protein collections carrying sequence variants and modifications.

Reverse recognition

  • Add an optional peptide-backbone recognizer that returns candidate residue paths plus confidence, rather than pretending an arbitrary molecular graph has one unique sequence.
  • Preserve atom-to-residue maps in SDF properties/CXSMILES so toolkit round trips remain auditable.
  • Distinguish exact reconstruction, metadata-assisted reconstruction, and heuristic recognition in result status fields.

Small molecules and reactions

  • Offer local systematic-name parsing through OPSIN for supported organic nomenclature, while keeping database synonym lookup through PubChem as a separate mode.
  • Add reaction SMILES/RXN support, atom mapping validation, and reactant/product fingerprints.
  • Add SDF property preservation, multi-conformer files, 2D coordinate generation, and optional ETKDG 3D conformers with energy minimization.
  • Add SMARTS queries as a distinct type; never auto-detect SMARTS as ordinary SMILES.

Machine learning outputs

  • Add count-based and sparse Morgan fingerprints, pharmacophore fingerprints, and MinHash for long sequences.
  • Add similarity matrices and approximate-nearest-neighbor export, with explicit metric metadata.
  • Treat learned embeddings as model-specific plugins pinned by model name, weights, revision, pooling, and preprocessing—not as generic “protein fingerprints.”
  • Support Parquet/Arrow and NumPy outputs for large feature matrices.

Proposed internal architecture

parser -> typed source model -> normalization policy -> molecular/polymer graph
       -> validator -> renderer (SMILES/InChI/HELM/SDF) or featurizer

This separates syntax parsing from chemical interpretation and makes unsupported cases visible.