論文紹介：Transfer learning in DeepLC

Nature Co mmunications | 17 : 2601 (2026) Transfer learning in DeepLC improves LC retention time prediction across substantially different modifications and setups Bouwmeester R., Nameni A., Declercq A. et al., Martens L. K ey messa ge One pre-trained DeepLC, fine-tuned by transfer learning, adapts to almost any LC setup or PTM — with far less data than training from scratch. E-Journal | 2026-04-27 | Shunichi Ito

INTRODUCT ION 1/ 5 DeepLC predicts retention time from peptide sequence DeepLC: learns Peptides / Retention Time relationship Peptide ID confidence Peptide sequence / Modifications (e.g. AGK[ac]PEPTIDER) DeepLC Retention Time (RT) Identification validation (Deep Learning Model) DIA spectral library The problem • RT shifts with: pH · stationary phase · pressure · temperature • Models don't transfer between labs / setups • New post-translational modifications (PTMs) break existing models Bouwmeester et al., Nat. Commun. 17:2601 (2026) | EJ 2026-04-27 Shunichi Ito

APPROAC H 2/5 Three ways to adapt DeepLC to a new LC setup DeepLC can be adapted to new data in three different ways: (A) Calibration (B) New model (C) Transfer learning traditional approach train from scratch fine-tune Pre-trained DeepLC DeepLC with RANDOM parameters Pre-trained DeepLC (FROZEN – no update) (No Peptide / RT knowledge) (still trainable) Fit GAM / spline Train fully on NEW data Transfer Learning on NEW data simple curve fit on output Train from scratch Fine tune prior knowledge Adjusts OUTPUT only Setup-specific model Adjusts MODEL itself Cannot handle large setup changes Needs lots of training data Best of both worlds Bouwmeester et al., Nat. Commun. 17:2601 (2026) | EJ 2026-04-27 Shunichi Ito

RESULTS ① 3/ 5 Standard datasets: transfer learning nearly always wins Learning curves across 3 PRIDE datasets (Fig. 1 d–f) Take-home < 100 peptides 1k – 10k peptides 473 / 474 datasets Calibration is fine Transfer learning wins by 1–2.5 % RMAE Transfer learning ≥ from-scratch Bouwmeester et al., Nat. Commun. 17:2601 (2026) | EJ 2026-04-27 Shunichi Ito

RESULTS ② 4/5 Extreme cases: calibration breaks, transfer learning holds Case A · TMPP (+572 Da modification) Case B · Basic (high-pH) LC Calibration (10-fold CV) New Model (10-fold CV) Transfer Learning (10-fold CV) Calibration (10-fold CV) New Model (10-fold CV) Transfer Learning (10-fold CV) 𝑅 = 0.821 𝑅 = 0.974 𝑅 = 𝟎. 𝟗𝟗𝟒 𝑅 = 𝟎. 𝟔𝟎𝟏 𝑅 = 0.946 𝑅 = 𝟎. 𝟗𝟕𝟎 Transfer learning adapts even when pre-training assumptions no longer hold. Bouwmeester et al., Nat. Commun. 17:2601 (2026) | EJ 2026-04-27 Shunichi Ito

RESULTS ③ · CONCL USION Unseen PTMs: transfer learning beats training from scratch & take-home Unseen PTMs (Fig. 5) CO NC LU SIO N One pre-trained model, many LC realities. Beats calibration on new LC conditions Beats from-scratch with far less data Generalizes to unseen PTMs (13 / 14) Prac tic al r ecipe <100 shared peptides → calibration. Otherwise → fine-tune DeepLC. 13 / 14 PTMs: transfer learning beats training from scratch. Bouwmeester et al., Nat. Commun. 17:2601 (2026) | EJ 2026-04-27 Shunichi Ito 5/5