A systematic review published in the Archives of Computational Methods in Engineering has mapped the landscape of generative artificial intelligence in bioinformatics, producing a finding that cuts against the dominant trend in commercial AI: specialized models trained on domain-specific data consistently outperform general-purpose architectures at biological tasks.
The review, led by researchers from multiple institutions and updated on July 23, follows the Preferred Reporting Items for Systematic Reviews and Meta-Analyses framework to evaluate generative AI strategies across genomics, proteomics, transcriptomics, structural biology, and drug discovery. It addresses six research questions covering model capabilities, the specialized-versus-general trade-off, benefits, methodological advances, limitations, and datasets.
The central finding concerns the performance gap between specialized and general-purpose models. Domain-specific architectures that undergo targeted pre-training on biological data routinely outperform their general-purpose counterparts, the review concludes. The advantage comes from context-aware design that reflects the underlying structure of biological problems, sequence patterns in proteins, regulatory relationships in gene expression, and the three-dimensional geometry of molecular interactions.
Areas of impact
The review identifies several domains where generative AI has meaningfully advanced the state of the art. In molecular analysis, models demonstrate improved accuracy in predicting protein structure and function, with reduced analytical error compared to traditional computational methods. In drug discovery, generative approaches are enabling the design of novel molecular candidates with desired properties, bypassing parts of the conventional screening pipeline. In integrative data analysis, models are learning to combine information across genomics, transcriptomics, and proteomics to produce more complete biological pictures.
The benefits are backed by established benchmarks. The review points to measurable improvements in structural modeling, functional prediction, and synthetic data generation, the latter being particularly valuable for training downstream models when real biological data is scarce or privacy-restricted.
Limitations and gaps
The review does not sugarcoat the field’s weaknesses. It flags poor scalability as a recurring problem: models that work well on small, curated datasets frequently fail when applied to genome-scale or population-level data. Data bias is another persistent issue, training distributions that over-represent well-studied organisms or disease pathways produce models that generalize poorly to understudied areas of biology. Restricted generalizability means a model trained on one type of cellular data may not transfer to another without significant retraining.
The authors recommend stronger evaluation practices and biologically grounded modeling approaches as remedies. They also highlight key datasets that underpin current progress, including UniProtKB and ProteinNet12 for molecular data, CELLxGENE and GTEx for cellular data, and PubMedQA and OMIM for textual biological knowledge.
For the broader AI industry, the review’s findings carry a practical implication: the one-model-fits-all approach that dominates commercial AI may be ill-suited to scientific domains where domain-specific knowledge is not a nice-to-have but a prerequisite for meaningful performance. The specialized architectures that win on bioinformatics benchmarks do so not despite their narrow focus, but because of it.
Sources: “Generative Artificial Intelligence in Bioinformatics: A Systematic Review” (arXiv:2511.03354, revised Jul 23, 2026); Archives of Computational Methods in Engineering