AI-driven protein research grew about 71% between 2024 and 2025. Protein–drug interactions led, rising from 49.9% to 54.4% of papers, while protein-structure prediction's share fell from 28.7% to 23.9%. The headline trend, though, is that smaller, specialized models are now beating much larger ones.

Smaller beats bigger

• MSAPairformer, a 111-million-parameter protein language model, surpassed previous state-of-the-art methods on the ProteinGym benchmark — at a fraction of the training and parameter budget. 2024's 98-billion-parameter ESM3 gave way to efficiency-first successors.
• GPN-Star, a 200-million-parameter genomics model, outperformed Evo 2 (40 billion parameters) on multiple variant-effect prediction tasks — a model nearly 200× larger.
• After AlphaFold 3, cofolding models converged on a similar parameter scale rather than continuing to grow. AlphaFold 3's FoldBench performance has yet to be significantly surpassed, even by larger models released since.

Virtual cells, and a data bottleneck

• Virtual-cell models emerged as a 2025 frontier: Evo 2 (Arc Institute), STATE, and DeepMind's AlphaGenome aim to predict cellular responses to drugs and genetic perturbations without wet-lab experiments — though current systems still require experimental validation.
• Development is now bottlenecked on data, not architecture. Training sets expanded from hundreds of thousands to tens of millions of entries via distilled AI-predicted structures and combined experimental sources; Meta's OMol25 alone holds 100M+ quantum-mechanical calculations.
• Biomni, a general-purpose biomedical AI agent from Stanford, mapped 25 subfields with 150 specialized tools, 105 software packages, and 59 databases — pairing digital reasoning with physical lab validation.

AI Overviews already top health searches

Google's AI-generated 'AI Overviews' now appear atop most health-related search results — on average 84%–92% of health queries triggered one across five query types. Symptom and common-health questions were the most likely to trigger an overview (92%), followed by treatment-related queries (90%) and condition-based ones (84%–88%).

Assistance, not autonomy

• Patients endorse AI in assistive roles rather than autonomous decision-making, especially in high-stakes clinical contexts.
• Preserving the human relationship is a consistent theme — patients name the potential loss of empathic care as a primary concern.
• Provider endorsement is a key determinant of patient acceptance, and transparency and disclosure of AI use are prioritized across populations.

Of all medical-AI publications in 2025, 43.4% discussed ethics topics — up from 37.1% in 2024 (medical-AI-and-ethics papers rose from 1,114 to 2,378). The volume is climbing, but the focus is uneven.

Governance-heavy, biosecurity-light

Ethical discussion skews toward governance. In 2025, only 14 publications discussed biosecurity, with even fewer directly addressing the ethical implications of misuse or dual use — a notable blind spot as biological models grow more capable.

Global health is the exception

Global health departs from the governance-dominated pattern: among 2025 global-health publications, 51.8% (100 of 193) also mentioned ethics, and societal concerns — equity, justice, and accessibility — ranked highest, surpassing both governance and algorithmic concerns.