Artificial intelligence has been reshaping biological research for years, but the pace of that transformation has accelerated to a point where even the scientists closest to it describe being caught off guard. Nature Methods, one of the field’s leading journals, has now dedicated a second special issue to the subject, just two years after its first. That interval alone says something significant: in most scientific disciplines, two years is barely enough time to replicate a single study. In AI-driven biology, it is enough time to require a comprehensive update.
AI Is Rewriting the Tools of Biological Inquiry
The breadth of fields now touched by AI in biology is striking. Researchers working in mass spectrometry-based proteomics, the study of proteins at scale, are exploring how AI can contribute to protein sequence analysis, protein interactions, spatial proteomics, and perturbation studies. The ambitions extend further still: multi-omics integration and AI-based virtual cell construction are now considered serious research directions, not speculative futures.
Microscopy is another area seeing rapid change. AI techniques drawn from computer vision are being applied to super-resolution microscopy to improve image reconstruction tasks including denoising, low-light enhancement, deblurring, and resolution enhancement. The goal is to extract meaningful biological information at the nanoscale, a scale where traditional imaging methods struggle. Stem cell research is also being reshaped, with AI architectures improving image analysis, generative models enabling data augmentation, and automated pipelines reducing the manual burden on researchers.
Perhaps the most conceptually ambitious application discussed in the issue is the idea of virtual embryos: fully digital reconstructions of the process of embryogenesis, designed to capture the complex, multiscale dynamics of how an organism develops from a single cell. This is not a metaphor. Researchers are working toward computational systems that model developmental biology in its full complexity.
Practical Tools, Built by Non-Engineers
Alongside these large-scale scientific ambitions, the issue highlights something more immediately practical: AI is beginning to change who can build scientific software and how quickly.
A contribution from Nelson Medina and Joergen Kornfeld describes how large language models can be used to generate custom software tools for scientific applications, without requiring dedicated software engineers. Their example is MOSS (Microscopy Oriented Segmentation with Supervision), a segmentation tool built by a single researcher with limited software development experience, completed in a matter of weeks. This is worth pausing on. Scientific software development has historically been a bottleneck, requiring either collaboration with engineers or years of self-taught programming. The suggestion here is that this barrier is lowering.
A separate contribution from Henry Pinkard and Nils Norlin raises a different practical opportunity. Scientific instruments routinely generate commands, data, and metadata that are typically discarded. These researchers argue that this overlooked data stream could be used to train AI systems to conduct experiments autonomously at a high level. The catch, and it is a significant one, is that realizing this potential requires rethinking what data gets stored and what quality control and human oversight mechanisms need to be in place. The data exists. The infrastructure to use it responsibly does not yet.
What This Means Beyond the Lab
Here is what most coverage of AI in biology misses: the scientific community is not simply adopting a new tool. It is grappling with a shift in how science itself is practiced, validated, and communicated.
Nature Methods is explicit about this. As AI models proliferate across biological fields, the absence of robust community standards for assessing their performance, reproducibility, and sustainability is becoming a genuine problem. Foundation models, large-scale machine learning systems trained on broad, heterogeneous datasets spanning multiple domains, are gaining traction across biology. But rigorous methods for evaluating their actual capabilities are still lacking. Researchers Julio Saez-Rodriguez, Gustavo Stolovitzky and colleagues have examined these gaps and proposed guidelines for more meaningful evaluations. A separate contribution addresses the importance of open and sustainable AI in life science research.
The journal’s editorial position is clear: transparency is not optional. New AI methods are only as useful as they are trustworthy. Community benchmarks are needed to assess whether new tools represent genuine advances or simply impressive-sounding outputs.
There is also a foundational point that deserves emphasis. AI systems in biology are entirely dependent on high-quality experimental data for training. No amount of algorithmic sophistication compensates for poor or insufficient data. The researchers, repositories, and infrastructure that generate and host biological data are not secondary to the AI story. They are its prerequisite. Ensuring that these resources remain supported is not a bureaucratic concern. It is a scientific one.
In Short
AI has moved from a promising addition to biological research into something closer to a structural feature of it. Virtual cells, AI-built microscopy tools, software generated without engineers, and instruments that could one day run their own experiments: these are not distant possibilities. They are active research directions documented in a leading peer-reviewed journal. The challenge now is not adoption. It is governance: building the standards, benchmarks, and oversight mechanisms that allow these tools to be trusted, reproduced, and fairly evaluated.
Based on reporting from Nature: Machine Learning.