Stage M2 – Machine learning to uncover hidden mitochondrial microproteins in dark genomes

SFBI; Gif-sur-Yvette; On-site; EUR 12,000 - 17,000; Full time

Job summary

SFBI is offering a 6‑month internship (M2) in the Computational Biology of Noncoding Genomes team. You will characterize 35,000 sequences, explore physico-chemical descriptors and protein-language model embeddings, and build predictive models for mitochondrial targeting signals.

Collaboration with Sorbonne Univ researchers is expected. The project combines machine learning, bioinformatics and evolutionary biology.

Qualifications

• Proficient in Python programming.
• Basic knowledge of machine learning concepts.
• Interest in mitochondrial biology and noncoding genomes.

Responsibilities

• Characterize 35,000 candidate sequences using physico-chemical descriptors.
• Develop and benchmark ML models to predict mitochondrial targeting signals.
• Analyze sequencing data from the mito-Split GFP reporter experiment to identify targeting motifs.

Skills

Python

Machine Learning

Bioinformatics

Job description

Stage M2 – Machine learning to uncover hidden mitochondrial microproteins in dark genomes

Description

Context: Microproteins have recently emerged as an exciting frontier in biology, with the potential to reveal hidden mechanisms of cellular function, proteome evolution and disease. Encoded by non-canonical open reading frames, these tiny proteins of fewer than 100 amino acids have long escaped traditional genome annotation, and recent OMICS studies are now illuminating this dark proteome in eukaryotic genomes. A key challenge is to understand how they contribute to proteome diversification, through evolutionary innovation and in pathological contexts. Importantly, to perform its function, a protein must be present in the appropriate cellular compartment.

Mitochondria are particularly interesting in this context, as a growing number of microproteins have been found to localize there, some playing important roles in cellular metabolism. Yet, how newly emerged proteins acquire the ability to reach mitochondria remains poorly understood. Many nuclear-encoded mitochondrial proteins rely on short N-terminal amphipathic helices known as mitochondrial targeting sequences (MTS). These signals have no conserved sequence motif and are instead defined by their physico-chemical properties. Remarkably, even random sequences can promote mitochondrial import, suggesting that functional targeting signals may emerge by chance. We have now tested this experimentally on 35,000 sequences.

Preliminary results: We previously showed that many microproteins encoded by noncoding regions have folding potential, with some adopting amphipathic helices compatible with MTS. In collaboration with the team of M. Garcia (Sorbonne University), we generated a library of 35,000 sequences (20–50 amino acids) randomly sampled across the Saccharomyces cerevisiae genome and fused to a mito-Split GFP reporter that fluoresces upon mitochondrial localization. The library is being screened by cell sorting followed by high-throughput sequencing.

While current predictors assign mitochondrial localization to approximately 10% of these sequences, about 30% of cells in the unsorted library are fluorescent. These observations suggest that noncoding regions may harbor a substantial reservoir of cryptic mitochondrial targeting signals that existing tools fail to recognize.

Hypothesis: Canonical MTS may have evolved from sequences already possessing some intrinsic import capacity, subsequently refined by natural selection. As a consequence, predictors trained on canonical mitochondrial proteins may fail to recognize the weak or atypical signals carried by newly emerged microproteins that have undergone little or no selection at the protein level, even though such microproteins can play significant roles in adaptation or human disease.

Objectives: This internship aims to identify the properties that enable protein sequences, including those of newly emerged microproteins, to target mitochondria, and to develop a predictive model that generalizes beyond canonical, evolutionarily optimized proteins and can be applied to the thousands of microproteins now being uncovered by large-scale OMICS studies.

The student will first characterize the 35,000 candidate sequences using physico-chemical descriptors, including amino‑acid composition, charge distribution, hydrophobicity and amphipathic helix propensity. These properties will be compared with those of canonical mitochondrial proteins, stratified by evolutionary age using phylostratigraphy, an approach developed in the team. The student will also assess whether existing localization predictors perform differently across evolutionary age classes, and explore whether protein language model embeddings provide useful information beyond explicit sequence descriptors.

Building on our experimental screen, the student will then analyze the sequencing data to identify the targeting sequences hidden among the tested candidates. Machine-learning models will be trained on these experimentally characterized sequences and benchmarked against current tools, starting from interpretable approaches based on explicit physico-chemical descriptors and extending to representations learned by protein language models. The central question is whether a model trained on unevolved sequences can recognize targeting signals that conventional predictors overlook.

Positioning: This internship lies at the interface of machine learning, bioinformatics and molecular evolution. It addresses a fundamental question: how do functional targeting signals emerge from sequences that selection has never shaped, and can we predict them before they have been optimized by evolution?

Environnement: The internship will take place in the Computational Biology of Noncoding Genomes team at I2BC and will be in close interaction with our collaborator Mathilde Garcia (CQSB - Sorbonne Univ) who leads the experimental part of the project. Our team brings together complementary expertise in structural bioinformatics, genomics, OMICS analyses and molecular modelling, and includes many young researchers (PhD students, engineers), providing a dynamic environment for early‑career scientists. We are located in the new buildings of the I2BC, south of Paris, on the historic and green campus of the CNRS. The I2BC is a recent and highly dynamic institute with around 700 people, including about 150 PhD students and postdoctoral researchers. The student will benefit from our HPC/GPU resources.

Required Technical Skills: Programming in Python (mandatory) and basic knowledge of machine learning.
Duration: 6 months

Candidature

Contacts