SATURNA

RNA
is central to life and disease, its role highlighted by COVID-19
vaccines and recent Nobel Prize. Decoding its behavior promises
breakthroughs for complex disorders such as cancer and
neurodegeneration. Despite the growing availability of omics
datasets and curated biological knowledge, our understanding of
how RNA molecules define and regulate healthy and diseased cell
states remains incomplete, hindering the development of
transformative RNA-based therapeutics. This gap stems in part
from the fragmented and inherently limited nature of
experimental data, as well as the complex rules that govern RNA
function and regulation, emerging from context-dependent
multilayered interactions between RNA and diverse molecular
partners. This underscores the need for advanced AI methods that
can integrate existing data, infer missing information, identify
hidden patterns, and generalize beyond experimentally observed
conditions to uncover novel biological principles. 

The
overarching goal of this project is to advance the state of the
art of AI for RNA, through advances in RNA language models,
biological knowledge injection, interpretability and complex
interaction and trajectory representation, in order to generate
new biological insights into how RNA shapes physiological
processes and thereby contributes to disease cell states. To
achieve this, we have assembled an interdisciplinary team of
experts in machine learning, natural language processing,
knowledge engineering, bioinformatics, RNA biology, and clinical
neuroscience. This integrated setup uniquely positions our team
to achieve the project’s objectives: with a history of
successful collaborations, we have identified new timely and
clinically relevant biological questions and designed an
effective collaborative research setting that synergizes
cutting-edge AI technologies with publicly available omics data
and curated biological knowledge bases. Additionally, the wet
lab component ensures the generation of high-quality, multimodal
datasets, crucial for fine tuning robust AI models and
validating key discoveries. 

Our
project is accordingly structured around four intertwined
scientific workstreams (WS1-4; Figure 1). WS1 (responsible PI:
Lonneke van der Plas) interprets and improves RNA foundation
models to uncover the intricate statistical properties of
non-coding mRNA regions. WS2 (responsible PI: Janna Hastings)
develops biologically informed methods to augment RNA-LMs with
knowledge about higher-level domains, structures and functions.
WS3 (responsible PI: Pascal Frossard) develops self-supervised
and generative graph representation learning methods to (i)
model context-specific, multimodal RNA interactomes and (ii)
develop robust modeling of disease trajectories. WS4
(responsible PI: Raphaëlle Luisier) elucidates the function of
non-coding mRNA regions in driving the subcellular
compartmentalization of biomolecules underlying physiological
and disease cell-state transitions, in close collaboration with
project partner Rickie Patani, who investigates how deregulated
RNA metabolism contributes to ALS. The rich sequencing data
generated in WS4 by the project partner will enhance and further
validate the AI methods developed across all workstreams. 

By
co-developing next-generation RNA foundation models, interactome
frameworks, and multimodal cell state models with RNA biologists
and clinicians, we will foster a truly interdisciplinary
international community, driving breakthroughs from fundamental
RNA biology to the mechanistic understanding of cancer and ALS,
which remains a incurable neurodegenerative disease. By
harnessing the latest advances in foundation models and
extending them to biological data, this ambitious project
promises to deliver novel computational methods coupled with
transformative biological insights that can be directly applied
to clinical and translational research.In summary, our project
will train the next generation of interdisciplinary scientists.