Deep Learning and Machine Learning Applications in Biomedicine

The rise of omics research, spanning genomics, transcriptomics, proteomics, and epigenomics, has revolutionized our understanding of biological systems. While the development of these technologies offers immense opportunities for exploring biological complexity, the sheer volume and complexity of multi-omics data present significant analytical challenges. In this context, Artificial Intelligence (AI), with its advantages in data processing and learning capabilities, has become a key tool for multi-omics data analysis. The applications of AI have expanded to include various fields such as disease diagnosis, precision medicine, drug discovery, and elucidating pathogenic mechanisms. This paper delves into the latest advancements of AI in the life sciences sector, with a particular emphasis on its application in crucial areas such as genomics, transcriptomics, and proteomics. By analyzing these successful cases, we aim to demonstrate the potential of AI in handling and applying multi-omics data and provide valuable insights and guidance to researchers.

Genomics, as a scientific field studying the genetic blueprint of organisms, is dedicated to decoding the genomes of living entities. The core objectives of this field include understanding genetic variations, gene functions within the genome, and their impact on an organism’s morphology, physiology, and disease occurrence. Particularly, deep learning (DL), a branch of Artificial Intelligence, has demonstrated significant potential in deciphering gene regulation, exploring genome structure, and analyzing variation effects. Tools like DeepVariant [1] and Clairvoyante [2] utilize deep learning to analyze DNA sequence data for variation detection, including identifying single nucleotide polymorphisms (SNPs) and structural variations. Compared to traditional methods, deep learning excels in capturing complex dependencies between sequencing reads, thereby enhancing accuracy and efficiency and enabling more precise genetic analysis. In the realm of cancer genomics, deep learning tools are instrumental in assessing the pathogenicity of variations, elucidating the specific effects of particular variations and informing treatment strategies and prognosis [3,4,5]. This application is particularly pivotal considering the unique genetic landscapes presented by various cancers. The journey towards understanding gene function has often been hindered by the inefficiency of experimental annotation. Researchers are increasingly employing Artificial Intelligence to develop computational tools in functional genomics. This includes the prediction of gene functions [6] and regulatory elements like enhancers and promoters [7,8]. Additionally, deep learning is utilized to mine and predict the functional impact of non-coding variations from large-scale genomic data, aiding in unraveling the complexities of gene regulatory networks and the potential causal mechanisms behind genetic variations [9,10,11]. Beyond these applications, deep learning also extends its influence into the realm of epigenomics. For instance, tools like DeepCpG [12] and DeepHistone [13] analyze DNA methylation and histone modification patterns, contributing to a more profound understanding of how these epigenetic factors influence gene expression and disease development.

With advancements in high-throughput sequencing technologies, the field of transcriptomics has broadened to include both single-cell and spatial resolution studies. The massive scale and complexity of raw transcriptomic data necessitate sophisticated computational algorithms and tools for preprocessing (quality control, dimensionality reduction, and clustering) and downstream analysis. Various renowned software packages, including Seurat [14] and Scanpy [15], offer comprehensive solutions for transcriptomic data analysis and are adept at tasks like data dimensionality reduction, cell clustering, and differential expression analysis. DL efficiently extracts rich, compact features from noisy, heterogeneous, and high-dimensional scRNA-seq data, thus enhancing downstream analysis. Unsupervised learning, employed for data mining and pattern identification in unlabeled data, is widely applied in scRNA-seq for dimensionality reduction and cell clustering [16,17,18,19]. In scRNA-seq, a low RNA capture rate frequently leads to dropout issues. Researchers utilize neural network algorithms for data imputation in scRNA-seq, effectively mitigating noise in gene expression profiles [20,21,22]. It is noteworthy that a significant advantage of DL in scRNA-seq data analysis is its capacity to handle nonlinear relationships between genes. In tasks like batch effect correction [23,24], cell type identification [25], and gene regulatory network [26,27] analysis, DL methods outperform traditional ones in terms of flexibility and efficiency. Deep learning also finds significant application in the field of spatial transcriptomics. Owing to sequencing technology limitations, emerging spatial transcriptomics has not yet achieved single-cell resolution in gene expression detection. Deep learning can be employed to synergistically analyze single-cell and spatial transcriptomic data, addressing this challenge [28,29]. Additionally, deep learning is utilized for spatial domain identification [30], cell–cell communication [31], 3D reconstruction [32], and detecting spatially variable genes [33]. Furthermore, the development of pre-training models such as Genefomer [34] is paving the way for more sophisticated analyses of specific downstream tasks.

Proteomics is one of the leading application domains for AI. Research includes predicting proteins’ three-dimensional structures and functions from primary sequences, studying protein interactions, and designing peptides [35]. Natural Language Processing (NLP) and Computer Vision (CV) methods, such as Transformers and Convolutional Neural Networks (CNNs), play a crucial role in the field of proteomics, particularly in protein residue modeling. A notable example is AlphaFold [36], which uses CNNs and RNNs to accurately predict protein spatial structures. DL also excels in identifying proteins’ biological functions based on amino acid sequences, aiding in both general and specific protein recognition [37,38,39,40,41,42,43,44]. In peptide research, it has revolutionized traditional methods, such as mass spectrometry, for peptide identification [45,46,47]. For protein sequence design, the ProtGPT2 model by Ferruz et al. [48] demonstrates DL ‘s capability in generating biologically consistent sequences. Analyzing post-translational modification (PTM) sites is another critical area where DL, particularly Transformer-based models, effectively classifies and predicts PTMs [49,50]. In the field of pharmacoinformatics, Artificial Intelligence has shown potential in predicting drug targets and drug–protein affinity [51,52]. Lastly, DL has significantly advanced single-cell proteomics analysis, improving proteome coverage and aiding in cell type/state identification from bulk tissue profiles [53].

Advancements in computing and algorithmic technology are broadening the scope of AI in life sciences. This evolution has drastically improved the processing and analysis of biological data, uncovering complex nonlinear correlations within biological systems and offering innovative approaches for disease research and drug development. AI’s role in precision medicine is increasingly critical, especially in biomarker discovery, sample classification, and interpreting disease processes. Despite these advances, DL in life sciences faces challenges related to dataset type and size, which affect its effectiveness and present uncertainties. Large-scale datasets demand greater computing power, while factors like model interpretability, data availability, and quality are pivotal, especially in vital areas like medical diagnosis and drug development, where understanding the influence of input features on predictions is crucial for trust in decision making. Future research might focus on improving algorithm efficiency and model interpretability to overcome these challenges. We are optimistic about the potential of DL in omics data analysis, anticipating that its ongoing development will yield new insights to propel bioinformatics and life science research forward. These advancements are anticipated to deepen our understanding of disease mechanisms and lay essential biological and computational groundwork for the development of future treatments and preventive measures.