Casanovo v5.2.0

5.2.0 - 2026-06-02

Added

• Support timsTOF files (as .d folders) as spectra input files.
• Added --load_all_states flag to load all model states when resuming training.
• A TSV file with all candidate peptides can be exported during database searching with the --export flag.
• Track instrument-assigned scan numbers from MGF SCANS, SCAN, and SCAN ID header fields in a new opt_global_cv_MS:1003057_scan_number mzTab column.
• Modified weights loading to match based on model selectors (e.g., orbitrap and timstof are currently supported) and major/minor versions.
• Support fine-tuning a pretrained checkpoint with an extended residue vocabulary via the new_token_init config option.
• Per-file validation loss logging via valid_CELoss/<stem> keys.
• New --tracking_peak_path/-t CLI option for monitoring catastrophic forgetting on additional validation files without affecting checkpoint selection.

Changed

• Upgraded minimum Lightning version to 2.6.
• Increased minimum Python version from 3.8 to 3.10.
• Black version upgraded for Python 3.10.
• Upgraded minimum DepthCharge version to 0.4.10.
• Changed default gradient clipping to a norm of 1.0.
• Updated train_batch_size documentation to reflect per-device/effective batch computation.
• A more descriptive error message is logged for some annotated spectrum file parsing failure cases.
• The precursor mass filter is no longer applied in de novo mode, and correspondingly peptide-level scores are no longer penalized based on the precursor mass. The config options precursor_mass_tol and isotope_error_range now only apply to database search mode.
• The amino acid scores and ProForma columns in the output mzTab files have been renamed to opt_global_aa_scores and opt_global_cv_MS:1003169_proforma_peptidoform_sequence, according to the mzTab specification.
• Minor speedup during database searching through optimized candidate selection.

Fixed

• A mismatching parameter warning will now only be triggered for the tokenizer if the config and checkpoint tokenizers do not have equivalent vocabularies.
• Removed erroneous tokenizer vocabulary warning.
• Fixed an issue which led the reported peptide precision to be 0 during evaluation mode.
• Peptide predictions failing the minimum peptide length are not reported, irrespective of whether they match or exceed the precursor mass.
• Setting --output_root to a directory will no longer cause an error.
• The --force_overwrite flag now also checks whether mzTab output files would be overwritten.
• Fixed an issue where some predictions that are one residue less than the configured minimum peptide length are reported.