High-throughput technologies are currently flooding life sciences with massive, highly complex multi-omics and biological imaging datasets. Classical machine learning and deep learning methodologies, while useful, often hit major roadblocks. These include high dependence on high-quality annotated data, weak generalization to unseen datasets, and significant computational bottlenecks when dealing with massive data scales.
To overcome these challenges, a collaborative research team from the College of Computer Science at Sichuan University, China has published a comprehensive review paper titled "Applications of large-scale artificial intelligence models in bioinformatics" in Quantitative Biology. The review systematically categorizes large-scale AI foundation models in bioinformatics into three modalities: large-scale language models (LLMs), large-scale vision models (LVMs), and large-scale multimodal models (LMMs), providing a roadmap for future bio-AI research.
LLMs for Biological Sequences:
Biological sequences (such as DNA, RNA, and protein amino acids) behave much like natural language. By treating biological parts as “words,” LLMs utilize tokenization, padding, masking, and positional embeddings to capture non-local dependencies within genomic and proteomic scales. Bidirectional pretraining models, such as Enformer and DNABERT-2, have significantly increased the accuracy of gene expression prediction and sequence function recognition. In proteomics, MSA-free (multiple sequence alignment-free) language models like ESM-2 and xTrimoABFold have reached competitive structural prediction accuracy while running up to 60 times faster than classical MSA-based methods.
LVMs for Biological Imaging:
Biological images (e.g., cell morphology, tissue pathology, and spatial localization) contain multi-scale dependencies. By combining the local feature extraction of Convolutional Neural Networks (CNNs) with the global context capabilities of Vision Transformers (ViTs), LVMs are redefining computational imaging. In spatial transcriptomics (ST), models like Hist2ST and THItoGene can predict spatial gene expression profiles directly from histological images. Other models, such as ViTScore, voxelize protein-ligand interaction pockets to accurately predict molecular docking poses.
LMMs for Multi-Omics Fusion:
Because biological systems are inherently multi-modal, LMMs fuse complementary information from text descriptions, sequences, 3D structures, and clinical images. Multi-omics models like DeepOmix integrate genome, transcriptome, and proteome data to predict cancer survival. Meanwhile, revolutionary tools like Evo, AlphaFold3, and Chai-1 unify sequence and structural prediction across DNA, RNA, proteins, and chemical ligands without the strict requirement of MSAs. In spatial transcriptomics, models like mclSTExp use multimodal contrastive learning to align tissue images with gene expression matrices, demonstrating massive prediction improvements.
Challenges and Future Outlook:
Despite their breakthrough achievements in bioinformatics, large-scale AI models still face critical challenges across multiple dimensions. In protein dynamics, current language models predominantly predict static structural snapshots, failing to capture highly dynamic conformational shifts in vivo; this necessitates exploring techniques like Parameter-Efficient Fine-Tuning (PEFT) and dynamic batching to balance prediction speed with conformational accuracy. Regarding data scarcity, the prohibitive cost of annotating spatial transcriptomics data drives the need for vision models that operate efficiently on limited annotated datasets or transfer to chromosomal banding images. For cross-modal alignment, biological multimodal datasets often suffer from missing modalities and conflicting signals, requiring advanced feature-fusion frameworks and CDS-Encoders to systematically reconcile multi-omics discrepancies. Finally, to address the high computational costs and massive parameter sizes of these models, model distillation, compression, and distributed computing are essential to democratize high-performance models for low-resource research environments.
DOI:10.1002/qub2.70026