Reuse existing DIANovo predictions

DIANovo inference can be computationally expensive. Reuse a completed, checked prediction file when possible. FastaLake reads its CSV and curates the candidate database; it does not run DIANovo or reproduce its model inference.

Take the pred_seq and pred_prob columns from DIANovo, developed for Disentangling the Complex Multiplexed DIA Spectra in De Novo Peptide Sequencing, and convert them to a two-column sequence,score CSV (pred_seq becomes sequence, pred_prob becomes score). The study runner accepts only that generic CSV or AlphaNovo's peptide_prediction_detokenized_unmodified,score CSV and rejects other headers. Give the converted CSV to the prediction input in the study tutorial. Preserve the original prediction file, model/preprocessing settings and acquisition identity.

Two formats, explicit confidence rules

Input	Example	FastaLake interpretation
Sequence generation	AcD and [0, 0.9, 0.9]	Three residue probabilities; mean 0.6
Complete path output	A c D and 0 0 0.9 0.9	One extra leading path marker, then three residue probabilities; mean 0.6
Incomplete path	A 113.084 D	Excluded: unresolved mass remains between the residues

Only canonical amino-acid letters are accepted after converting case and removing ASCII whitespace. Lowercase modified residues retain their underlying amino acid. Numeric gaps and modification strings are not erased or inferred. For a path vector, the extra leading zero is removed only when it is space-delimited and has exactly one more entry than the complete peptide. Bracket/comma vectors need exactly one entry per residue. Values must be finite and between zero and one. Real zero probabilities participate in the mean; an all-zero mean is excluded by the normal positive-confidence selection rule.

Extraction and razor selection log excluded unresolved/invalid prediction rows. Rust reads the complete file during selection. The same parsing contract is used by the experimental Python helper.

Zero probabilities are kept and numeric gaps are preserved.

DIA scope

A smaller database can feed spectral-library prediction and DIA searching. Reuse of one acquisition for de novo selection and searching does not establish whole-workflow FDR. Validate acquisition windows, modifications and calibration before interpreting results. See the method and its scope.

Find the code

• Python parser: complete sequence and probability checks.
• Shared Rust parser: extraction and parsimony input.
• Regression tests: zero-aware target ranking, path markers and rejection of unresolved or malformed input.

The format contract was checked against upstream commit cc413f25140e47a5f6a4234bae15da0ef525ac3a, specifically sequence_generation_inference_utils.py and optimal_path_inference_classic.py. The parser is a FastaLake adapter; DIANovo's model implementation is not copied.