A unified framework for the systematic detection of microproteins in standard proteomics workflows using a Ribo-seq informed transcriptomic language model

Technical Summary: A Unified Framework for Microprotein Detection

Problem Statement
Despite the widespread translation of non-canonical open reading frames (ncORFs) revealed by ribosome profiling (Ribo-seq), integrating this knowledge into routine mass spectrometry (MS) workflows remains difficult. Current genomic annotations rely on arbitrary length thresholds (e.g., ~100 codons), systematically excluding thousands of functional microproteins. While specialized proteogenomic approaches exist, they face significant barriers:

1. Biochemical Constraints: Microproteins are often low-abundance and generate few tryptic peptides, making them difficult to detect via standard shotgun MS.
2. Computational Limitations: Searching against exhaustive databases containing all predicted microproteins (e.g., OpenProt) drastically increases search space, reducing sensitivity for canonical proteins and requiring excessive computational resources.
3. Workflow Incompatibility: Combining Ribo-seq and proteomics on the same sample is often impractical due to differing sample preparation requirements, costs, and technical complexities, preventing routine integration in standard laboratories.

Existing deep learning models for predicting translation initiation sites (TIS), such as the original TIS Transformer, are limited by their training data, which relies heavily on canonical Ensembl annotations. Consequently, they fail to accurately predict non-AUG starts and ncORFs, which are prevalent in Ribo-seq data.

Methodology
The authors propose a unified framework that bridges the gap between canonical and non-canonical proteome analysis without requiring specialized experimental protocols for every sample. The methodology consists of three primary components:

1. Ribo-seq Meta-Analysis and Data Curation:

   • The authors processed 108 Ribo-seq experiments (771 FASTQ files) from 74 studies across 50 human cell lines and tissues.
   • Using the RiboTIE tool, they identified 87,110 unique ORFs, extracting 40,832 non-canonical TISs (ncTISs) associated with 19,585 distinct ncORFs. This dataset included various biotypes (uORFs, uoORFs, intORFs, etc.) and non-AUG initiation codons (e.g., CUG, GUG).

2. Retraining of TIS Transformer (TISnc Transformer):

   • The authors retrained the TIS Transformer deep learning model by incorporating the 40,832 experimentally derived ncTISs as positive labels alongside the original 96,555 canonical TISs from Ensembl.
   • This retraining aimed to teach the model sequence features unique to ncORFs while preserving its ability to predict canonical TISs.
   • The resulting model, termed TISnc Transformer, was calibrated to maintain a 91–92% recall rate for canonical TISs.

3. Construction of Swiss-Prot/MicroProt Database:

   • A size-controlled database was created by combining the canonical Swiss-Prot sequences with the 48,265 ncORF-derived microproteins predicted by the TISnc Transformer.
   • This database was designed to be compatible with standard search engines (specifically MaxQuant/Andromeda) to allow routine identification of microproteins without the computational penalty of exhaustive, unfiltered databases.

4. Validation and Reanalysis:

   • The framework was tested by reanalyzing a comprehensive HeLa cell line MS dataset (PXD004452) using both the standard Swiss-Prot database and the new Swiss-Prot/MicroProt database.
   • Identified microproteins were cross-referenced with orthogonal evidence from Ribo-seq (RiboCRYPT, TransCODE v35, RiboTIE) and independent MS studies (Wacholder et al., Deutsch et al., OpenProt).

Key Results

• Model Performance: Retraining with ncTISs significantly improved the model's ability to detect ncORFs. TISnc Transformer predicted 48,265 ncORFs compared to 18,031 by the original model, while retaining 98% of the original canonical ORF predictions. Statistical analysis confirmed a highly significant enrichment of curated TransCODE ncORFs in the new predictions (p < 10⁻⁴⁶), indicating improved sensitivity rather than just increased search volume.
• Proteomic Concordance: Searching the HeLa dataset with Swiss-Prot/MicroProt maintained a 98.83% overlap in peptide-spectrum matches (PSMs) with the standard Swiss-Prot search. The total PSM count decreased by only 0.49%, and the core set of 13,291 canonical proteins remained largely undisturbed.
• Microprotein Discovery: The expanded database identified 539 microproteins (representing 3.87% of total protein groups). Of these, 348 were unique to the expanded search, and 270 had orthogonal evidence of existence (from Ribo-seq or other MS studies).
• Database Stability: The Jaccard index for leading proteins between the two searches was 94.1%, demonstrating that the inclusion of microproteins did not destabilize the identification of the canonical proteome.

Significance and Claims
The paper claims to provide a unified framework that facilitates the systematic detection of microproteins within conventional proteomics workflows. By leveraging Ribo-seq data to inform a language model, the authors created a practical database (Swiss-Prot/MicroProt) that:

• Bridges the Gap: It allows for the simultaneous analysis of canonical and non-canonical proteomes without requiring specialized, sample-specific Ribo-seq experiments for every proteomic study.
• Preserves Sensitivity: It avoids the sensitivity loss associated with exhaustive databases by using a size-controlled, high-confidence set of predicted microproteins.
• Enables Routine Integration: It offers an accessible intermediary between exploratory microproteome resources and widely used reference databases, potentially reducing the methodological bias that currently underestimates the regulatory impact of microproteins.

The authors explicitly note that this approach is not designed for cutting-edge, specialized microprotein research requiring low-molecular-weight enrichment or highly specialized pipelines. Instead, it is intended to recover a significant fraction of biological information in standard shotgun proteomics experiments, thereby improving the completeness of functional proteomic studies.