AutoRT

In proteogenomics, a spectrum matched to a variant peptide is a claim that a tumour expresses a mutated protein, and it is hard to audit: decoy-based false discovery rate estimates rest on a handful of variant sequences in a database dominated by reference ones, so three defensible FDR schemes return very different neoantigen lists from the same raw files. Liquid chromatography, meanwhile, records an orthogonal fact about every peptide it separates — when it elutes. If the assignment is correct, the peptide should come off the column when its sequence says it should, and the gap between predicted and observed retention time is a check that never saw the search engine's score.

AutoRT makes that check usable by predicting elution time from sequence alone accurately enough that a residual of half a minute carries information. Instead of hand-designing the network, its authors ran a genetic algorithm over convolutional architectures paired with a bidirectional gated recurrent unit layer, scoring each candidate by validation mean squared error and evolving the population through selection, mutation and crossover. The ten best architectures were then trained out and retained as an ensemble. AutoRT was developed by Bo Wen, Kai Li, Yun Zhang and Bing Zhang at Baylor College of Medicine and published in Nature Communications in April 2020, alongside NeoFlow, the neoantigen prioritization workflow it supplies with a quality metric.

The release ships the searched-and-trained ensembles, not only the search code: prediction runs directly against the committed general base ensemble with no preceding training step, and separate ensembles cover phosphorylated, acetylated and ubiquitinated peptides. Fine-tuning on a laboratory's own runs is an optional adaptation to a new gradient or column, not a prerequisite.

#Key Features

• Architecture search in place of architecture design: A genetic algorithm evolved 50 candidate networks per generation across 20 generations, varying convolutional layer count, feature maps, activations, dropout, batch normalization and pooling around a fixed 50-unit bidirectional GRU.
• Ten-model ensemble with outlier trimming: Predictions from the ten retained architectures are filtered by the interquartile-range rule and averaged, lowering median absolute error by 25%, 18% and 18% relative to the constituent models on three benchmark sets.
• Modification-aware encoding: Modified residues are written as distinct single-character codes, so they pass through the same network as unmodified peptides. Dedicated base ensembles ship for S/T/Y phosphorylation, lysine acetylation and lysine ubiquitination.
• Transfer learning for a specific experiment: The base ensemble is fine-tuned with no layers frozen on 700–10,000 peptides from the target LC-MS/MS run, in under 15 minutes on a single GPU.

#Technical Details

Peptides are mapped to integer codes over the 20 amino acids plus up to five modification symbols and embedded; the shipped general ensemble supports lengths up to 60 residues and is min–max scaled to a 0–101.3 minute gradient. Architecture search ran on three NVIDIA Titan Xp GPUs over a public label-free dataset of 174,182 peptides (PRIDE PXD006109), after which the ten best architectures were trained for up to 100 epochs at batch size 64 in Keras/TensorFlow. On held-out data AutoRT reached a median absolute error of 0.47 minutes on PXD006109, against 0.57 for GuanMCP2019, 0.66 for DeepMass, 0.72 for Prosit and 2.37 for the feature-based GPTime, and led on two SWATHAtlas datasets at 1.41 and 1.44 minutes. Fine-tuned per run across CPTAC cohorts, median error was 0.57 minutes over 540 label-free runs, 0.68 over 252 TMT runs and 0.63 over 900 iTRAQ runs.

#Applications

The source paper uses AutoRT to arbitrate between quality control strategies across 287 tumour samples and some 57 million spectra from three CPTAC cohorts, comparing global, separate and two-stage FDR estimation under three search engines. The same residual serves immunopeptidomics, where MHC-bound peptides lack tryptic regularity, and modification studies, where a misassigned site changes predicted elution. Groups building spectral libraries or rescoring peptide-spectrum matches use predicted retention time as an added feature; the DeepRescore2 pipeline starts from AutoRT's phosphorylation base ensemble.

#Impact

AutoRT arrived as mass spectrometry proteomics was absorbing deep learning, and stands out for treating the network as something searched rather than specified — a choice that let one codebase cover tryptic, MHC-bound and modified peptides without separate hand-tuning. Code and trained ensembles are distributed together under GPL-3.0 with a Docker image and Colab notebooks, making it a common building block in downstream proteogenomics and rescoring tools. Its limits are stated plainly by its authors: peptides longer than 60 residues fall outside the shipped ensembles, and because retention time is scaled to a particular chromatographic gradient, a base ensemble moved to an unfamiliar setup generally needs the fine-tuning step to reach the reported accuracy. Later catalog entries such as Prosit-PTM and DDA-BERT carry the same idea — predict a peptide property, then treat the residual as evidence — into fragment intensities and rescoring.