PREM KUMARSCIENCE · LEADERSHIP · AI
← All stories
AI & Discovery

Artificial Intelligence: When Machines Began Learning Biology

AI can find patterns across biological data at remarkable scale. The harder question is how to turn predictions into reliable scientific understanding.

A new colleague in the laboratory

Imagine a researcher facing thousands of measurements, each representing a small part of a biological system. The problem is not simply a lack of data. It is knowing which patterns matter, which are accidental and which might point towards a useful experiment. Machine learning offers new ways to explore this complexity. Algorithms can identify statistical structure in sequences, images, molecular measurements and clinical datasets. The prospect is exciting: tools that help scientists notice relationships that would otherwise be difficult to detect. Yet the promise comes with a condition. A pattern recognised by a model is not automatically a biological explanation.

Biology was always a data problem

Modern biology generates extraordinary quantities of information. Sequencing instruments produce genomic data; microscopes capture detailed images; proteomics and metabolomics describe molecules across complex samples. Traditional statistical methods remain essential, but some questions benefit from models capable of learning intricate relationships. AI is not one technology. It includes classification systems, generative models, language models, computer vision and methods for predicting molecular structure or function. Each approach has different assumptions and failure modes. The first question should therefore be what scientific task needs solving, not whether AI can be added to it.

From sequences to structures

Protein structure prediction has become one of the most visible examples of AI supporting biological research. Predictive systems can provide structural hypotheses that help scientists plan experiments and interpret molecular function. These outputs are valuable, but confidence varies across regions, interactions and biological contexts. A predicted structure is not a direct observation of how a protein behaves inside a living cell. Experimental confirmation may still be necessary, particularly when conclusions depend on dynamics, binding or unusual molecular environments. The strongest workflow treats computation and laboratory evidence as partners, with each exposing limitations in the other.

The danger of a convincing answer

Generative AI can produce fluent explanations, code and literature summaries. Fluency is not evidence. Models may invent citations, misread study designs, overlook contradictory results or express uncertainty poorly. In scientific work, those failures can propagate into manuscripts, protocols and decisions if they are not caught. A responsible research assistant should distinguish retrieved evidence from generated interpretation, identify what it has not checked and make it easy for a scientist to inspect the source. The relevant benchmark is not whether an answer sounds knowledgeable. It is whether its claims survive independent verification.

Data quality sets the ceiling

Biological datasets often contain batch effects, missing information, measurement artefacts and sampling biases. A model trained on one laboratory’s data may perform poorly elsewhere because instruments, populations or procedures differ. Randomly splitting highly related observations into training and test sets can create misleadingly optimistic results. External validation, held-out testing and careful documentation of data provenance are therefore essential. Model performance should be reported with meaningful uncertainty and evaluated against appropriate baselines. Without those safeguards, sophisticated algorithms may simply learn the peculiarities of a dataset rather than generalisable biology.

From prediction to experiment

AI is most useful when it improves the cycle of scientific reasoning. A model may prioritise candidate targets, suggest hypotheses or help design a more informative experiment. The next step is to test those ideas against reality. Negative results matter because they reveal where the model’s assumptions were wrong. Reproducible experiments, transparent methods and independent replication remain central. Automation can increase the speed of analysis, but speed becomes valuable only when the resulting knowledge is dependable. Scientists should be rewarded for testing predictions critically, not for producing impressive-looking outputs.

The governance question

AI systems in health and biology can affect privacy, safety, research integrity and public trust. Sensitive datasets require careful stewardship, and clinical applications need appropriate regulatory evaluation. WHO has published guidance on the ethics and governance of large multimodal models in healthcare, emphasising oversight and risk management. In research settings, practical safeguards include access controls, documented model versions, independent review, clear authorship disclosures and restrictions on unsupported claims. Governance should enable responsible innovation rather than serve as a substitute for understanding the technology.

Keeping the scientist in the loop

The most important change may be cultural. AI can make certain technical tasks easier, but it cannot take responsibility for the truth of a scientific conclusion. Researchers still need to ask whether a question matters, whether evidence is adequate and what alternative explanations remain. The future laboratory may contain more algorithms, automation and interconnected data systems. Its success will still depend on curiosity, critical thinking and the courage to challenge a result that appears too convenient. Machines may help us see more of biology. Understanding what we see remains a human scientific responsibility.

Sources and further reading

References are provided for the principal factual claims and for readers wishing to explore the evidence. This article is an editorial interpretation, not a systematic review or clinical guidance.

  1. WHO: Ethics and governance of large multimodal models
  2. AlphaFold Protein Structure Database
  3. Nature: AlphaFold protein structure prediction

Publication status: Public draft · Version 0.1 · Not peer-reviewed · Subject to revision.

← Explore all articles