Module rt 

Retention-time prediction: shared featurization + a gbdt model distilled from DeepLC.

Both uno-rt-train (offline distillation) and the search scorer use rt_features so the model sees identical inputs at train and inference time. rt_features emits the 99-d order-aware set described on RT_N_FEATURES below. It REPLACED an earlier composition-only set in the style of Klammer et al. (2007) (20 AA counts, 20 N-terminal-window counts, 20 C-terminal-window counts, length, ln(1+mass), intercept = 63 features), which is no longer built here.

Structs

RtModel

A loaded RT model plus a content fingerprint (a stable hash of the model file bytes). The fingerprint keys the on-disk prediction cache (crate::rt_cache) so cached predictions are reused only for the exact model that produced them.

Constants

DEFAULT_RT_CNN

The supported pretrained CNN, embedded so installed binaries work anywhere.

RT_N_FEATURES

Number of features produced by rt_features. Order-aware “rich” set: 20 AA composition + 6 terminal-residue identities (first 3, last 3) + N_POS_BINS*20 position-binned composition + 13 scalar descriptors. Validated to match DeepLC’s MS1 protein recovery (vs the old 63-feature composition-only set above, which trailed it), via order-aware features the composition set lacked.

Functions

read_label_tsv

Read a distillation label file: SEQUENCE\tRT per line (no header), as produced by sherlock’s DeepLC scripts. Returns (sequences, rts).

rt_features

Order-aware “rich” RT feature vector for a stripped peptide sequence. Layout: AA composition (20), terminal residue identities (6), position-binned composition (N_POS_BINS*20), then 13 scalar descriptors.