Abstract
Deep Learning (DL) has recently enabled unprecedented advances in one of the grand challenges in computational biology: the half-century-old problem of protein structure prediction. In this paper we discuss recent advances, limitations, and future perspectives of DL on five broad areas: protein structure prediction, protein function prediction, genome engineering, systems biology and data integration, and phylogenetic inference. We discuss each application area and cover the main bottlenecks of DL approaches, such as training data, problem scope, and the ability to leverage existing DL architectures in new contexts. To conclude, we provide a summary of the subject-specific and general challenges for DL across the biosciences.
Introduction: The Biological Deep Learning Divide
The rapid integration of deep learning (DL) into the biosciences has created a starkly uneven landscape. While some sub-fields have undergone paradigm-shifting transformations, others remain largely reliant on classical statistical methods. This review explores the factors that make a biological problem "prime" for DL, the technical limitations currently hindering progress, and the specific architectures best suited for diverse biological data types. By examining five key areas—protein structure, protein function, genome engineering, systems biology, and phylogenetics—we map the current frontiers of AI-driven biology.
Deep Learning Frameworks and Architectures
The choice of a DL architecture is fundamentally driven by the nature of the biological data. The authors categorize biological data into four primary "Key Ideas" that dictate model selection:
- Spatial Locality: Handled by Convolutional Neural Networks (CNNs), which are ideal for data with local features like protein contact maps or medical imaging.
- Sequential Nature: Addressed by Recurrent Neural Networks (RNNs) and Transformers, which excel at processing DNA or protein sequences where temporal or positional order is critical.
- Context Dependence: Managed through Attention Mechanisms, allowing models to capture long-range interactions within complex molecules.
- Data Distribution and Embedding: Utilized by Autoencoders (AEs) and Generative Adversarial Networks (GANs) for tasks like dimensionality reduction in single-cell transcriptomics or generating synthetic biological samples.
The Data Landscape and the Success Spectrum
A major bottleneck in biological DL is the disparity in high-quality, labeled data.
- Protein Structure: While the Protein Data Bank (PDB) contains only ~180,000 entries, researchers have successfully leveraged massive sets of unlabeled sequences from UniProtKB/TrEMBL to extract evolutionary information. This "unsupervised wealth" strategy was key to overcoming limited structural data.
- Systems Biology: Multi-omic repositories like GEO and TCGA provide rich data, but integration remains challenging due to noise and batch effects.
- Phylogenetics: This field faces a unique hurdle: the lack of "ground truth" for ancient evolutionary history, forcing a reliance on simulated data.
The impact of DL follows a clear spectrum. It has been paradigm-shifting in protein structure prediction (e.g., AlphaFold2) and has seen major success in protein function and genome engineering. However, its impact remains moderate in systems biology and minor in phylogenetic inference, where classical models still offer superior mathematical guarantees.
Key Breakthroughs and Results
The most significant leap in computational biology history was marked by AlphaFold2, which outperformed competitors in the CASP13 competition by nearly double. In genome engineering, models like DeepCas9 have drastically improved the prediction of CRISPR "off-target" effects, enhancing the safety of gene editing. Furthermore, autoencoders have become essential for "denoising" single-cell RNA-seq data, enabling the discovery of rare cell types that were previously obscured by technical noise.
Challenges and Future Directions
Despite these successes, several systemic barriers remain:
- The Explainability Crisis: High-accuracy "black box" models are often insufficient for clinical or experimental validation. The adoption of "Explainable AI" (XAI) tools is crucial for understanding the biological reasoning behind model predictions.
- Computational Costs: Training state-of-the-art models requires immense GPU resources, creating a financial barrier that favors large institutions.
- Data Bias: Genomic datasets are often skewed toward specific populations, risking the development of biased AI that performs poorly for underrepresented groups.
Looking forward, the authors suggest a shift toward Transfer Learning and the development of large "foundation models." By pre-training on general biological data, these models can be fine-tuned for specific, data-poor tasks, democratizing access to high-performance AI for smaller research teams.
Summary updated