🧬 biology

RECIPE bridges transcriptomics and proteomics with deep graph learning on Ribo-seq data

The paper introduces RECIPE, a graph neural network framework that leverages Ribo-seq or transcriptomic data integrated with protein-protein interaction networks to accurately predict proteome-wide protein abundances, effectively overcoming the limitations of incomplete proteomics measurements in both bulk and single-cell settings.

Original authors: Jun Ding, Luying Su, Bowen Zhao, Kailu Song, Dingfeng Zou, Jingtao Wang, Ziyang Wang, Wei Liu, Xiuhui Yang, Qinghui Duan, Qi Geng, Rong Huang, Heyun Guo, Yan Lu, Kai Li, David Eidelman, Wei Song

Published 2026-08-06

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jun Ding, Luying Su, Bowen Zhao, Kailu Song, Dingfeng Zou, Jingtao Wang, Ziyang Wang, Wei Liu, Xiuhui Yang, Qinghui Duan, Qi Geng, Rong Huang, Heyun Guo, Yan Lu, Kai Li, David Eidelman, Wei Song

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine your body is a bustling, high-tech city. In this city, the blueprints for every building are stored in a master library called the DNA. But blueprints alone don't build anything; they need to be copied and sent out to construction sites. These copies are called RNA, and they act like the foremen's orders. Finally, the actual buildings—proteins—are constructed based on those orders. Proteins are the real workers: they build muscles, fight infections, and keep your cells running.

For a long time, scientists had a hard time counting the workers. They could easily read the blueprints (DNA) and the foremen's orders (RNA), but counting the actual buildings (proteins) was like trying to count every single brick in a city while it was being built. The tools were slow, expensive, and often missed a huge chunk of the workforce, especially in tiny, single-cell neighborhoods. This left a big gap: we knew what the instructions said, but we didn't know exactly what was actually getting built. The big question was, could we use the foremen's orders to predict the number of buildings, even if we couldn't see the buildings themselves?

Enter RECIPE, a new digital tool created by a team of scientists that acts like a super-smart detective for this city. The researchers realized that while we can't always see the finished buildings, we can watch the construction crews in action. They used a technique called Ribo-seq, which is like taking a high-speed photo of the construction sites to see exactly where the workers are pausing and how fast they are building. But just looking at the construction speed isn't enough, because sometimes a building gets torn down or modified after it's built.

To solve this, RECIPE uses a "graph neural network," which is a fancy way of saying it maps out how all the proteins know each other. Think of it like a social network for proteins. If you know that Protein A is best friends with Protein B, and you know how many Protein Bs are being made, you can make a very good guess about how many Protein As there should be, even if you can't see Protein A directly. RECIPE combines the construction speed photos (Ribo-seq) with this social network map to predict the total population of proteins in the cell.

The team tested this tool on human and mouse cells, specifically looking at what happens when they disrupted a key part of the construction machinery. They found that RECIPE was incredibly accurate. It didn't just guess the proteins that were easy to see; it successfully predicted the abundance of proteins that were completely invisible to traditional measurement tools. In fact, when they compared RECIPE's predictions to the actual protein counts, the numbers matched up with a correlation of 0.9682 in human cells and 0.9657 in mouse cells. That is a very strong match, suggesting the tool is doing something right.

Perhaps even more impressively, the tool worked in "single-cell" mode. Usually, looking at just one cell is like trying to guess the weather by looking at a single raindrop—it's too small and messy. But RECIPE managed to predict protein levels in individual cells by learning from the "bulk" (large groups) of cells first and then zooming in. It even worked when they only had the blueprint copies (RNA) and no construction photos (Ribo-seq), though it was slightly less accurate in that mode.

The researchers also discovered something fascinating about the proteins that were hardest to predict. These weren't random errors; they were specific types of proteins, often involved in building the cell's skeleton or assembling complex machines. The tool struggled with them because their final numbers are controlled by rules after construction (like how long they last or how they are assembled), not just how fast they are built. By identifying these "hard-to-predict" proteins, RECIPE actually helps scientists find the ones that are being regulated by these hidden, post-construction rules.

In short, RECIPE bridges the gap between the instructions and the final product. It suggests that by combining construction speed data with a map of how proteins interact, we can get a complete picture of the cell's workforce, even for the workers we can't directly see. This doesn't mean we can stop measuring proteins entirely, but it offers a powerful new way to fill in the blanks, especially in situations where direct measurement is impossible or too difficult. The authors show that this approach works well across different species and conditions, providing a more complete view of how our cells function than we had before.

Technical Summary: RECIPE

Problem Statement

Direct quantification of protein abundance remains technically challenging, particularly in low-input and single-cell settings where current proteomics technologies capture only a sparse fraction of the proteome. This incompleteness limits the ability to characterize protein-level regulation and resolve cellular heterogeneity. While Ribosome profiling (Ribo-seq) offers a genome-wide readout of translation that is proximal to protein production, converting these signals into accurate protein abundance estimates—especially for proteins lacking direct proteomic measurements—remains an unresolved challenge. Existing approaches are largely limited to "closed-set" regimes, where inference is restricted to proteins observed during training, preventing the prediction of unmeasured or condition-specific targets. Furthermore, RNA-based models often fail due to transcript-protein discordance caused by translational and post-translational regulation.

Methodology: The RECIPE Framework

The authors introduce RECIPE (Ribo-seq Expression & Connectivity Integration for Proteome Estimation), a graph neural network (GNN) framework designed to bridge transcriptomics/translatomics and proteomics. The framework operates under incomplete proteomics supervision, enabling proteome-wide prediction beyond directly measured proteins.

Core Architecture

RECIPE integrates three primary components:

Feature Extraction (FE):
- Sequence Embeddings: Transcript sequences are encoded using DNABERT (k-mer=3), generating fragment-level representations concatenated to form transcript-level embeddings.
- Translational Abundance: Ribo-seq read counts are normalized and log-transformed to quantify translation levels.
- Ribosome Pausing: A pausing score is calculated based on codon-specific ribosome footprint density to capture elongation dynamics.
- Note: In the absence of Ribo-seq, the framework can operate in an RNA-seq-only mode by substituting translational abundance with transcript abundance and omitting pausing features, albeit with reduced fidelity.
Bulk Module ( $F_B$ ):
- The model utilizes a GraphSAGE architecture on a Protein-Protein Interaction (PPI) network.
- Nodes represent protein-coding transcripts, and edges represent curated PPIs.
- Node features (sequence, abundance, pausing) are projected into a latent space and aggregated via mean aggregation to learn context-aware representations.
- A regression head predicts protein abundance. The model is trained on a subset of proteins with available proteomics labels, allowing it to generalize to unmeasured proteins through the PPI graph structure.
PPI Refinement Module ( $F_{PPI}$ ):
- A self-supervised edge classifier ( $C_{PPI}$ ) evaluates the likelihood of interactions based on learned protein embeddings.
- This module refines the prior PPI network by removing likely spurious edges and introducing high-confidence missing links, generating a dataset-specific interaction graph that captures context-aware relationships.
Single-Cell Module ( $F_{SC}$ ):
- Single-Modality Setting: When single-cell protein measurements are unavailable, the model undergoes "pseudo-bulk alignment." Single-cell profiles are aggregated to match bulk proteomics for training, followed by cell-level graph learning using a Cell-GraphSAGE on a cell-cell KNN graph.
- Paired Multi-Omics Setting: When single-cell protein data is available for a subset of observations, the model is fine-tuned using these direct measurements to improve cell-resolved predictions.

Key Contributions

Open-Set Proteome Inference: RECIPE introduces a paradigm shift from closed-set prediction to open-set inference. By leveraging translation-resolved signals and PPI topology, it predicts abundance for proteins entirely absent from the training proteomic labels.
Context-Aware PPI Refinement: The framework does not treat PPI networks as static priors but refines them based on learned representations, yielding dataset-specific interaction maps that capture condition-dependent regulatory structures.
Unified Single-Cell Framework: RECIPE provides a unified approach for single-cell protein inference, capable of operating with single-cell Ribo-seq, single-cell RNA-seq (as a fallback), or paired multi-omics data.
Identification of Translation-Protein Decoupling: The model systematically identifies a subset of "challenging-to-predict" proteins where abundance is decoupled from ribosome occupancy, pointing to post-translational regulatory mechanisms.

Results

The framework was validated across matched human (HEK293T) and mouse (C18) datasets, including DDX20 knockdown experiments and nutrient deprivation studies.

Bulk Prediction Accuracy: RECIPE achieved high correlation between predicted and real protein abundance (Pearson $r \approx 0.96$ in HEK293T and C18 cells), significantly outperforming raw Ribo-seq signals and reducing prediction error compared to baseline methods (XGBoost, Random Forest, SVM, scTranslator, etc.).
Unmeasured Protein Recovery: Crucially, RECIPE successfully predicted abundance for proteins masked during training (simulating unmeasured proteins), achieving a Pearson correlation of $r \approx 0.80$ in HEK293T cells. This demonstrates genuine extrapolation rather than interpolation.
Biological Fidelity: Predictions preserved downstream biological signals. Differential expression analysis showed strong concordance (RRHO) between real and predicted datasets. Gene Ontology (GO) enrichment analysis of predicted downregulated proteins recovered the same functional categories (e.g., protein binding, RNA binding) as real data.
Single-Cell Performance: In single-cell settings, RECIPE resolved cell-state structures (e.g., cell cycle phases) and condition-specific variations (e.g., arginine/leucine deprivation) more accurately than raw single-cell Ribo-seq signals. It successfully generalized across datasets (e.g., SVEC to C10 cell lines) in paired multi-omics settings.
PPI Refinement: The refined PPI networks showed higher functional coherence (shared pathway annotations) and stronger co-expression signals compared to deleted edges or random non-interacting pairs, validating the biological plausibility of the learned interactions.
Challenging Proteins: The model identified a reproducible subset of proteins with high prediction errors. These proteins were enriched for cytoskeletal and ribosomal functions and exhibited specific sequence features (e.g., lower GC3, higher AT3), suggesting their abundance is governed by post-translational regulation (stability, complex assembly) rather than translation rate alone.

Significance and Claims

The paper claims that RECIPE addresses a fundamental limitation in proteomics: the inability to obtain complete, condition-resolved protein measurements. By enabling proteome-wide inference directly from widely available transcriptomic and translatomic data, RECIPE shifts proteomics from a partially observed measurement modality to a foundation for comprehensive inference.

The authors emphasize that the framework provides a scalable strategy for analyzing protein abundance in data-sparse regimes, particularly single-cell settings where direct measurement is infeasible. Furthermore, the ability to identify proteins where translation and abundance are decoupled offers a systematic entry point for studying post-translational regulation. The refined PPI networks generated as a byproduct provide context-aware interaction maps that can facilitate downstream network reconstruction and pathway characterization.

The study concludes that RECIPE serves as a general framework for predicting protein abundance from translation-resolved (or transcriptomic) signals despite incomplete proteomics supervision, extending the scope of proteome-centric analysis in both basic and translational studies.