Two-Stage Probability-Enhanced Regression on Property Matrices and LLM Embeddings Enables State-of-the-Art Prediction...

Area of Science:

• Biotechnology and Bioinformatics
• Computational Chemistry
• Genomic Medicine

Background:

• Six small interference RNAs (siRNAs) approved since 2018 highlight their therapeutic potential via selective gene knockdown.
• siRNA design is complex, with chemical modifications critical for stability and therapeutic half-life.
• Machine learning (ML) offers advanced analysis for predicting siRNA efficacy and off-target effects.

Purpose of the Study:

• To develop a novel pipeline for predicting the gene knockdown activity of chemically modified siRNAs.
• To leverage composition-aware property matrices and large language model (LLM) embeddings for enhanced siRNA design.
• To benchmark various LLMs for target gene encoding in predicting siRNA activity.

Main Methods:

• A novel pipeline integrating siRNA chemical composition-aware property matrices and LLM embeddings for target gene encoding.
• Benchmarking of general-purpose and domain-specific LLMs, including Mistral 7B, for predicting siRNA activity.
• A two-stage probability-enhanced model to address data imbalance and improve prediction accuracy.

Main Results:

• The Mistral 7B LLM demonstrated superior performance compared to models pre-trained on genomic data.
• The proposed model achieved state-of-the-art quality with R² = 0.84 and RMSE = 12.27% on unseen data.
• Leave-one-gene-out experiments confirmed the model's ability to generalize to unseen genes, indicating robust feature and embedding representation.

Conclusions:

• The developed pipeline effectively predicts gene knockdown activity for chemically modified siRNAs.
• The model enhances siRNA design by integrating chemical properties and gene embeddings, improving therapeutic efficacy.
• This work advances the field of composition-aware siRNA design for next-generation nucleic acid therapies.