Bio-BLIP: A Multimodal Architecture for Transferable Reasoning in Genomic Variant Interpretation.

Area of Science:

• Computational Biology
• Genomics
• Artificial Intelligence

Background:

• Scientific hypothesis generation in biology necessitates integrating diverse data types like DNA, genes, proteins, and literature.
• Current multimodal AI systems often rely on textification or fine-tuning language models, limiting their generalizability.
• Task-specific optimization restricts the adaptability of existing AI models for complex biological reasoning.

Purpose of the Study:

• To introduce Bio-BLIP, a multimodal Q-former based architecture designed for generalizable biological reasoning without task-specific fine-tuning.
• To leverage biological embeddings and a large language model (LLM) for integrating heterogeneous biological data.
• To develop a model capable of handling multiple data modalities for complex reasoning tasks.

Main Methods:

• Developed a novel neural network architecture, Bio-BLIP, integrating DNA, genes, proteins, and text data through a master Q-former model.
• Utilized biological embeddings and a LLM backbone, feeding integrated modality information as a fixed-length prefix.
• Pretrained Bio-BLIP on human genetic variant annotation to establish baseline performance.

Main Results:

• Achieved a 29.8% increase in accurate variant feature generation compared to frontier LLMs.
• Demonstrated superior performance in zero-shot evaluation for variant prioritization and target gene prediction.
• Outperformed alignment-free genomic language models in regulatory variant prioritization for Mendelian diseases.

Conclusions:

• Bio-BLIP offers a natively multimodal and generalizable reasoning approach for biological domains with multi-scale data and varied tasks.
• The model provides transparent reasoning traces, enhancing interpretability.
• Bio-BLIP represents a significant advancement in AI for complex biological hypothesis generation and genomic analysis.