HERMS · INSIGHTS

AI in Life Sciences: The Latest Innovations Moving from Prediction to Experiment

Global innovation briefing · Updated 5 November 2026

The latest wave of life-sciences AI extends beyond generating text. It is helping researchers interpret DNA variation, design molecular candidates and connect computational work with laboratory measurement. The opportunity is substantial, but prediction, experimental confirmation and clinical evidence remain distinct stages.

Genomics: precomputed predictions at a new scale

On 8 September 2026, Google DeepMind introduced AlphaGenome Atlas, a research resource containing predicted molecular effects for nine billion possible single-nucleotide variants. The associated variant-impact score is intended to help prioritise variants and interpret the molecular processes that may be affected. These are model predictions, not nine billion experimentally validated clinical findings. [1]

Protein design becomes more programmable

On 21 October 2026, NVIDIA and the University of Washington's Institute for Protein Design announced RFD4-Proteina, an open model combining protein design with search. NVIDIA reported computational benchmarks across binding and design tasks. Candidate generation and benchmark performance still require laboratory testing before a useful biological function or therapeutic effect can be established. [2]

AI agents meet the wet lab

Google Research's 16 October report described agents connected with physical laboratory robotics to support iterative molecular-design experiments. The relevant shift is a measured feedback cycle: propose a candidate, run an experiment, evaluate the result and choose the next step. This remains a research setting, rather than a general authorisation for autonomous clinical or regulated manufacturing decisions. [3]

Platforms are linking models with scientific infrastructure

NVIDIA's 2026 BioNeMo announcements describe tools and collaborations spanning biological models, laboratory orchestration and automation. Such programmes indicate investment in the infrastructure between a model and an experiment. They should be evaluated as individual capabilities with documented maturity, rather than assumed to provide validated end-to-end discovery. [4]

What makes adoption defensible

HERMS interpretation: select a defined scientific or operational task and measure the decision it improves. For discovery, track experimentally confirmed hits and reproducibility. For document workflows, measure error rates, reviewer effort and traceability. Keep benchmark scores separate from evidence of patient benefit.

A controlled deployment should record source data, model and prompt versions, intended use, access permissions, uncertainty and reviewer decisions. Changes need evaluation against the original acceptance criteria. Human scientific and professional accountability remains essential when outputs influence patient safety, submissions or investment decisions.

For life-sciences leaders, the strongest near-term strategy is to connect AI with governed evidence and real experimental feedback. Faster hypotheses matter when the organisation can also explain which ones survived testing, why they were selected and what remains uncertain.

Sources and further reading

Explore the evidence, regulatory requirements and implementation questions behind your next innovation decision. Discuss your priorities with HERMS →