Abstract:
RNA is far more than a passive messenger between DNA and protein. Its ability to fold into diverse and dynamic structures creates an additional layer of information that regulates gene expression and enables plants to respond to their environment. Understanding this “RNA language” requires approaches that capture structural diversity and uncover functional signals hidden within vast datasets.
In this talk, I will describe our journey from probing RNA structures in living plants to using artificial intelligence to decode their regulatory information. Single-molecule RNA structure profiling reveals alternative RNA conformations and their dynamic responses to environmental cues, exemplified by the cold-responsive long non-coding RNA COOLAIR. I will then introduce PlantRNA-FM, a foundation model pretrained on 1,124 diverse plant transcriptomes that learns general representations of plant RNA and uncovers sequence and structural features associated with biological functions, including translation efficiency.
Moving beyond prediction toward discovery and design, we use AI to identify translation-associated RNA structural motifs, predict their positional effects, and quantitatively tune gene expression. Finally, I will introduce PlantScience.ai, a virtual plant biology scientist. Together, these advances demonstrate how integrating RNA biology, large-scale data and foundation models can move us from reading RNA language toward understanding and ultimately writing its regulatory code.
Bio:
Yiliang is Group Leader at the John Innes Centre, Norwich. She received her BSc in Plant Sciences from Shanghai Jiao Tong University and her PhD in Biology from the University of East Anglia and the John Innes Centre. She undertook postdoctoral training at the University of Dundee and at Penn State University. Her research is on the interface of RNA biology, computation and plant science, developing and applying machine learning approaches to decode RNA structure and function in plants.
Summary:
Central Dogma: DNA->RNA->Protein
RNA can have complex 3D structures: RNA triples, Pseudoknot, stem-loop, G-quadruplex
In-vivo RNA structure profiling: identify nucleotide pairing status
DNA->pre-mRNA
RNA translation
RNA degradation
miRNA silencing
Studies in plants show that RNA structure changes over its lifetime
When studied in-silico a single RNA sequence can have many different conformations
PacBio sequencer can scan the individual conformations of different RNA molecules in a cell
Differences in RNA structures affect plant flowering
How to find functional RNA motifs from Big Data
Which aspects of the structure are important for gene function/protein translation/RNA degradation?
PlantRNA-FM foundation model
1124 diverse plant transcriptomes, 25M RNA sequences, 54B RNA bases
Pretrain RNA LLM
Objectives:
Masked Nucleotides Modelling
Secondary Structure Annotation (from in-silico structure prediction)
RNA Region Annotation: 5’ UTR, CDS, 3’UTR
Specialized Tokenizer
Fine-tuning
Infer annotations for different RNA sequences
Infer function: high-translation vs low-translation gene
Compared to a PlantRNA-FM variant where high-/low-translation was randomized
Identify Translation Related RNA structure motifs
Strong stem with 4 GC pairs that is associated with low translation
Weak stem with AGCU sequences associated with high translation
Changing between these stems in RNA in-vivo is very effective at changing translation effectiveness
G-rich RNA folds to complex RNA g-quadruplex
GGN1-7 or GGGN1-9 repeated: acts as a suppressor to translation
Moving this segment through the RNA sequence makes its suppression effect stronger or weaker
Goal: design RNA to change translation
Can put this motif into any gene to control its translation
Same foundation model relevant to diverse downstream tasks
Fine-Tuning
RNA Decoding
RNA Design
Challenges in using generic LLMs for plant science: hallucination, term ambiguity, no traceability
Design
AutoSKG: assembly of plant data/knowledge corpus
Plant Science Knowledge Graph
Online question answering (questions can trigger acquisition of new papers and data)
Global community is using tool: 4338 registered users across 60 countries