Skip to content

Release 4.9

Latest

Choose a tag to compare

@halx halx released this 06 Nov 19:42
  • Modular diversity filter architecture: pluggable filter types (IdenticalMurckoScaffold,
    IdenticalTopologicalScaffold, ScaffoldSimilarity, PenalizeSameSmiles, BitBirchDiversityFilter)
    selectable via [diversity_filter] type
  • BitBirch clustering-based diversity filter (optional: pip install reinvent[bitbirch])
  • Configurable penalty shapes for diversity filters: Step, Linear, Tanh, Sigmoid, Erf
  • Intrinsic reward support: RND and Information-based rewards via [intrinsic_reward]
  • Importance-sampling weighted inception (is_weighted_sampling, is_weighted_scores)
  • Dynamic filter components: any scoring component can act as a filter using filter = true
    and a step-type transform
  • New scoring component PathFPSimilarity: Tanimoto similarity using RDKit path fingerprints
  • New scoring component FingerprintPrecedence (analogous to RingPrecedence)
  • New transform symmetric_exponential_decay: bell-shaped reward centred on a target value
  • pumas is now an optional dependency (pip install reinvent[pumas])
  • Stricter config validation: run_type, device, file paths, and key parameters are now checked
    at parse time with clear error messages rather than failing mid-run
  • Cleaner terminal error output: validation errors are formatted; tracebacks go to the log file
  • Various robustness fixes: NaN handling in scoring components, scoring run mode validity mask,
    CSV header stability, stereochemistry in diversity filter, checkpoint backward compatibility
  • Migration note: deprecated config parameters use_cuda and unique_sequences have been
    removed; configs still using them will fail validation on startup