RiboAI

Introduction

Protein production is controlled at several regulatory layers, including mRNA abundance, mRNA stability, translation efficiency and local ribosome dynamics. RiboAI investigates how sequence, codon context, RNA structure and experimental context jointly shape these processes. The central idea is to combine harmonised transcriptomics data with machine learning models that are predictive enough to be useful, but still interpretable enough to generate biological hypotheses.

The project focuses on ribosome profiling and RNA-seq data, complemented by sequence-derived and structure-derived annotations. Rather than treating deep learning models as black boxes, RiboAI uses model interpretation to identify candidate regulatory features that can later be checked against known biology and tested experimentally.

People

Supervisors: Prof. Ivo Hofacker, Prof. Sebastian Tschiatschek and Prof. Stefan L. Ameres.

PhD students: Gabriele Martino focuses on machine learning models, while Denis Skibinski focuses on the bioinformatics workflow, dataset curation and ribosome profiling analysis.

Scientific Questions

RiboAI asks which mRNA properties control stability, translatability and ribosome occupancy along transcripts. Candidate signals include local sequence motifs, codon usage, untranslated-region context, transcript abundance, RNA secondary structure and local structures that may affect ribosome movement.

An important part of the project is to distinguish reproducible biological signal from dataset-specific technical effects. Ribosome profiling experiments differ in cell line, organism, stalling reagent, lysis conditions, nuclease treatment, adapter design, read length, UMI usage and other protocol details. RiboAI therefore models not only the ribosome profile itself, but also the experimental attributes and quality metrics that may explain systematic biases.

Data and Workflow

The computational workflow started with HEK293 and HeLa ribosome profiling cohorts and has been expanded into a broader collection of human cell lines, primary-like sample types, tissues and model organisms. Current and prepared datasets include HEK293, HeLa, LCL, HBE, Huh7, H1299, U2OS, A549, macrophages, monocytes, neurons, heart, liver and pancreas samples, as well as datasets from C. elegans, E. coli and S. cerevisiae.

All datasets are processed through harmonised Ribo-seq and RNA-seq workflows so that downstream comparisons are based on consistent transcript annotations, read processing, alignment, quantification and P-site assignment. Quality control is treated as a modelling input rather than only a filtering step. The project tracks metrics such as footprint length distribution, frame-0 periodicity, CDS enrichment, rRNA/tRNA contamination, transcript coverage and replicate correlation.

Feature Space

For each transcript, RiboAI combines the ribosome profile with a rich feature representation. This includes one-hot sequence encoding, codon-frame and region annotations, transcript abundance, local opening energies from RNAplfold, predicted local structures, exon-junction information, mappability and other transcript-level descriptors.

At the dataset level, the project also builds an attribute vector space that captures biological, chemical, technical and quality-related information. This makes it possible to ask whether a model is learning general translation biology, a cell-line-specific signal or a protocol-specific bias.

Modelling Strategy

Several existing ribosome-density models are used as baselines, including iXnos, Riboexp, RiboMIMO and Riboformer. These comparisons help determine how much of the ribosome profile can be explained by local codon context, how much requires longer-range sequence information and where dataset-specific correction becomes necessary.

The current modelling direction is a new architecture that separates biological signal from technical bias. Sequence-aware components such as recurrent and convolutional networks represent transcript-level signals, while dataset-aware components use experimental attributes and quality information to account for systematic differences between experiments. The loss functions are designed for sparse ribosome profiling data, where many codon positions have low or zero counts but some positions show strong peaks.

Outlook

The long-term goal is to turn predictive models into biological knowledge. Interpretable machine learning and explainability methods are used to identify sequence and structure features that may regulate mRNA stability, translation efficiency or ribosome stalling. Promising hypotheses can then be compared with known RNA-binding protein sites, microRNA regulation, RNA modifications, disease mutations and eventually tested with high-throughput reporter assays.

In the next phase, RiboAI will continue expanding the dataset collection, improve quality-aware modelling and refine the separation between reproducible translation signals and technical effects. The broader aim is to build a framework that helps explain how mRNA sequence and structure shape protein production across cell types, organisms and experimental protocols.