Can a machine learn to recognize a real microprotein gene from its DNA alone? Across the tree of life, genomes contain thousands of short open reading frames that might encode proteins, but most annotation tools were never designed to flag sequences this small. The result is that many genuine microprotein genes have been missed entirely.

This project builds a foundation model for smORF detection: an AI system that integrates mRNA sequence, cellular context, and ribosome profiling data to predict translation at codon-level resolution. The model learns from data across the ORFeus network, including dual host-pathogen datasets from DC4, rare translation events from DC3, and genetic variant information from DC12, and is iteratively refined as experimental validation comes in from groups across the consortium. The goal is not just a better predictor, but a set of open, reusable tools that the wider community can build on.

All models and predictions will be released through the ORFeome platform and HuggingFace. X

Standard ribosome profiling gives a population-level view: it tells you that a particular ORF is translated in a sample, but not which specific RNA molecule is being read, or how different transcript forms from the same gene carry different proteins. This project goes beyond population averages to resolve translation at the level of individual RNA molecules.

The project uses RiboSTAMP, a technique that marks translating ribosomes by fusing a cytidine deaminase to a ribosomal protein, combined with Oxford Nanopore long-read sequencing to connect translation information directly to full-length transcript structure. This makes it possible to ask whether short ORFs are co-translated from polycistronic transcripts, and how RNA isoforms and RNA modifications like m6A influence which ORFs are translated. The project also tests whether these approaches can be adapted to single cells.

Results inform work across the network, from isoform-level analysis in DC7 to single-cell translation profiling in DC9, and feed into the shared ORFeome resource.

Most translation events are common enough to detect in a single ribosome profiling experiment. But many of the most interesting microproteins may be translated only under very specific conditions: in a particular tissue, during a moment of stress, or in a species no one has profiled yet. These are the hardest to find, and the most likely to be missing from current catalogs.

This project takes a data-driven approach. Rather than running new experiments one at a time, it brings together more than 15,000 publicly available ribosome profiling datasets and mines them at scale, asking which short open reading frames show evidence of translation across conditions, species, and cell states. The challenge is both computational and methodological: different labs use different protocols and organisms, and separating real signals from technical noise across such a heterogeneous collection requires new approaches.

The project builds on the Valen group’s ORFik framework and connects directly to RiboSeq.Org, the community data platform co-developed with the Baranov group at UCC. Validated rare events are cross-referenced with immunopeptidomics data from DC11 and DC15 to check whether cryptic translation products reach the cell surface.

When a bacterium infects a cell, both host and pathogen are translating proteins, and both may be producing microproteins that have never been catalogued. This project maps that dual landscape in the context of enteric infection, using Citrobacter rodentium as a model pathogen.

The project combines advanced dual proteomics (capturing host and pathogen proteins simultaneously) with ribosome profiling to build a time-resolved picture of microprotein dynamics across three stages of infection, from initial contact through to established disease. Proximity labelling (BioID) is used to identify which host and pathogen microproteins interact with each other during infection, and thermal proteome profiling (in collaboration with DC10) validates these interactions.

The result is a comprehensive ORFeome dataset of infection-associated microproteins, which also serves as a biologically relevant input for the antigen presentation studies in WP3.

Many microproteins are secreted from cells, where they may act as signalling molecules or interact with receptors on neighbouring cells. But current proteomic methods largely miss them because they are small, present at low concentrations, and quickly diluted in cell culture media. This project develops new chemical biology approaches to capture them.

The project uses bioorthogonal non-canonical amino acid labelling to introduce minimal chemical handles into microproteins, combined with crosslinking mass spectrometry (XL-MS) to capture secreted microproteins at the moment they bind their cell-surface receptors. The approach is applied to co-culture models of the tumour microenvironment, where cancer cells communicate with immune and stromal cells through signals that may include microproteins no one has seen before.

The resulting atlas of secreted microproteins feeds into the binder design work in DC6 and provides candidate targets for the immunotherapy-focused projects in DC14 and DC15.

Once a microprotein is identified, how do you find out what it does? One powerful approach is to design a synthetic molecule that binds it specifically and can switch its activity on or off. This project uses computational protein design to create such binders for microproteins identified across the ORFeus network.

The project starts with a consortium-prioritized list of 20 target microproteins, expresses and purifies them, and performs biophysical characterization. It then uses AI-based design tools, including AlphaFold3 and RoseTTAFold, to computationally design high-affinity synthetic binders. These are iteratively refined using structural and biophysical data from DC10 and validated using binding assays and, in collaboration with DC5, cellular assays that test whether the binder actually modulates the microprotein’s function.

The structural models, binder sequences, and biophysical data are contributed to the ORFeome platform as a shared resource for functional validation.

Long non-coding RNAs were assumed, by definition, not to encode proteins. But ribosome profiling has shown that some of them are translated, producing microproteins whose functions are almost entirely unknown. Whether this translation is widespread or limited, regulated or accidental, and what the resulting proteins do in the cell are open questions at the heart of dark proteome biology.

This project investigates the rules that govern when and how ribosomes translate long non-coding RNAs, using neuronal development in Drosophila as a model system. Fruit flies offer a powerful genetic toolkit, and many of the regulatory mechanisms they use are conserved in mammals. The project uses TCP-Seq to map ribosome scanning dynamics, ribosome profiling including Poly-Ribo-Seq to identify which lncRNAs are translated at different developmental stages, and long-read direct RNA sequencing to resolve the full-length transcript structures involved.

The collaboration with M. Mar Albà in Barcelona adds an evolutionary dimension: if a microprotein encoded by a lncRNA has been conserved across species, that is strong evidence it matters. The partnership with Oxford Nanopore Technologies contributes long-read sequencing expertise.

Evolution leaves traces. If a microprotein has been conserved across species over millions of years, it is very likely to do something important. If it appeared recently in evolutionary time, it may still be functional, but the evidence is harder to read. This project uses evolutionary analysis to predict which newly discovered microproteins matter.

The project builds computational models that integrate ribosome profiling data, phylostratigraphy (to estimate evolutionary age), and genetic variant data from population databases like gnomAD and the UK Biobank. The models predict which microproteins show signs of evolutionary constraint, and therefore which are most likely to be functional. Top candidates are then validated experimentally using CRISPR-Cas9 screens in cancer cell lines, testing whether their disruption affects cell growth, drug resistance, or DNA repair.

Evolutionary constraint data also feeds back into DC1’s prediction models, and functionally validated microproteins are prioritized for further study in DC13 and the translational projects in WP3.

Different cells in the same tissue can translate the same gene differently. A tumour is not one cell type but many, and the microproteins that matter in one subpopulation may be absent in another. Understanding this heterogeneity requires translation profiling at single-cell resolution, something that has only recently become technically feasible.

This project establishes a new “ribo-perturb-seq” framework that combines CRISPR-based editing with single-cell ribosome profiling. The approach adapts Perturb-seq vector designs so that guide RNAs and translation profiles can be read out simultaneously from the same individual cell. After validation in K562 and RPE-1 cell lines, the technology is applied to organoid models of neural development in collaboration with DC7, investigating how smORF variants affect translation regulation during differentiation.

Isoform-level translation data from DC2 informs the single-cell analysis frameworks, and top candidates from DC7’s lncRNA translation studies are functionally validated at single-cell resolution.

How does removing a single microprotein change the behaviour of a whole proteome? Most approaches to studying protein function focus on one protein at a time, but thermal proteome profiling (TPP) offers something different: a proteome-wide readout of how protein stability and interactions shift when a gene is disrupted.

This project applies TPP systematically to 20 microprotein deletion strains of Salmonella enterica across five different growth conditions, building a comprehensive functional map of the bacterial microproteome. Changes in stability point to protein-protein interactions, and the most confident hits are validated using biophysical methods and integrated with structure predictions from AlphaFold and RoseTTAFold in collaboration with DC6.

The resulting workflows are designed to be reusable: adapted for host-pathogen interaction studies with DC4, and providing biophysical data that informs the binder design work in DC6.

HPV rewrites the host cell’s RNA landscape. When the viral genome integrates into the human genome, it can create hybrid transcripts that fuse viral and human sequences, and some of these may encode proteins that have never been catalogued. If those proteins are displayed on the cell surface, the immune system could potentially recognize them as foreign, making them attractive targets for immunotherapy.

This project maps non-canonical translation in HPV-driven cancers. Using long- and short-read RNA-Seq combined with ribosome profiling in HPV-positive reference cell lines and head-and-neck squamous cell carcinoma cell lines, the project identifies hybrid viral-human transcripts and determines which of them are translated. Proteomics and immunopeptidomics then confirm whether the resulting peptides are processed and presented on HLA molecules. The immunogenicity of the most promising candidates is tested using T-cell assays.

The result is an “HPV Immune Map” of non-canonical translation and antigen presentation, contributing validated peptide candidates to DC15’s predictive modelling and informing the regulatory studies in DC14.

Common genetic variants can change which proteins a cell makes, even outside the known gene catalog. A single nucleotide change in a non-coding region might create a new start codon, eliminate one, or alter how the ribosome navigates a complex transcript. Most of these effects are invisible to standard genome-wide association studies, because the affected ORFs are not annotated.

This project investigates how genetic variants in non-coding regions alter translation, using Ribosome Decision Graphs (RDGs), a framework developed in the Baranov group for representing the complexity of eukaryotic translation. The project adapts RDGs to be genotype-specific, integrating data from population databases (UK BioBank, gnomAD) and new ribosome profiling data from DepMap cancer cell lines. With DC3, it builds an integrated map of translation diversity by identifying where variant-triggered and naturally rare translation events overlap.

The ultimate goal is a catalogue of cancer-relevant translation-altering variants, advancing the field’s ability to interpret non-coding variants identified in patient genomes.

What happens when cancer cells start reading their genome in unexpected ways? Under the metabolic stress that is typical of a tumour and its microenvironment, cells can switch on a hidden layer of protein coding, translating sequences that are normally silent and producing proteins through non-canonical mechanisms. A central question is how this process works mechanistically, and what the resulting dark proteome means for cancer biology and for the immune system’s ability to recognise tumour cells.

This project studies non-canonical translation in cancer, with particular interest in how cellular stress, including amino acid deprivation, drives the production of aberrant microproteins. Using ribosome profiling, proteomics, and immunopeptidomics, the project identifies these proteins and asks whether they are presented on HLA molecules where the immune system can find them. The work runs alongside ILLUMINE, the Cancer Grand Challenges project on the dark proteome, and will integrate and collaborate with it where the science aligns.

The precise focus will be shaped together with the successful candidate and the supervisory team.

For a microprotein to become an immune target, the cell must not only make it but also process and display it on its surface. This processing and presentation pathway has many steps, each controlled by specific cellular factors. Identifying the key regulators of this pathway for microprotein-derived antigens is essential if the dark proteome is to be exploited for immunotherapy.

This project uses pooled CRISPR screens to systematically identify the factors that control whether smORF-derived peptides are presented on HLA molecules. The approach uses a dual-reporter system that monitors both translation and surface presentation simultaneously, screening against a library of biologically and therapeutically relevant targets prioritized from across the network (including host-pathogen data from DC4 and secretome data from DC5). Top candidates are validated using immunopeptidomics and tested in patient-derived cancer models in collaboration with DC15.

The result is a regulatory map of smORF antigen presentation, with at least one validated mechanism that could serve as a therapeutic target for enhancing tumour visibility to the immune system.

Which hidden peptides actually reach the cell surface and stay there long enough for the immune system to find them? Even if a cryptic ORF is translated, the resulting peptide may be degraded before it is presented, or it may bind HLA molecules too weakly to be stable. Understanding what determines successful presentation is the key to selecting credible targets for cancer immunotherapy.

This project identifies cryptic HLA-I peptides in patient-derived primary melanoma cell lines using advanced immunopeptidomics, quantifies their stability and turnover rates using SILAC pulse-chase assays, and develops an integrative machine learning model that predicts which cryptic peptides will be stably presented. The model synthesizes diverse data streams from across the network: translation features from ribosome profiling, immunopeptidomics data, and information on secreted microproteins from DC5.

The result is a comprehensive catalogue of cryptic ORF-derived peptides from melanoma and a predictive tool for prioritizing immunogenic targets for cancer vaccines and immunotherapy.