AlphaFold Changed Science. After 5 Years, It’s Still Evolving

Summary

AlphaFold — DeepMind’s AI that predicts protein structures — turned five and has reshaped biology and chemistry. Its database now holds on the order of 200 million predicted structures and is used by millions of researchers worldwide. The system evolved from AlphaFold2’s atomic-accuracy protein folding to AlphaFold3, which extends modelling to DNA, RNA and small molecules.

The article features an interview with Pushmeet Kohli, DeepMind’s VP of research, who outlines DeepMind’s problem-first approach, the move to diffusion models, and the continued use of a generation-plus-verification architecture to reduce hallucinations. He describes an “AI co-scientist” (multi-agent system built on Gemini 2.0) that accelerates hypothesis generation and debate, citing Imperial College work on viral phages as an example. Kohli highlights the next big ambition: simulating complete cellular systems — starting with the nucleus — to transform drug discovery, personalised medicine and broader applications like climate-relevant enzymes.

Key Points

AlphaFold’s database now covers roughly the known protein universe and is widely adopted by the global research community.
AlphaFold3 expands predictions beyond proteins to include DNA, RNA and small molecules, enabling more complex interaction modelling.
DeepMind pairs creative generative models (now diffusion-based) with rigorous verifiers and confidence scores to limit structural “hallucinations.”
“AI co-scientist” multi-agent systems (Gemini 2.0) act as virtual collaborators: generating, debating and refining hypotheses to speed discovery.
Real-world validation remains crucial — human-led experiments continue to confirm and interpret AI predictions.
Concrete use case: Imperial College used the co-scientist to explore how pirate phages hijack bacteria, advancing antimicrobial-resistance research.
Longer-term ambition: a reliable simulation of a complete human cell, beginning with understanding nuclear processes and gene readout dynamics.

Context and relevance

AlphaFold is a watershed for computational biology — it changed how researchers approach structural problems and accelerated many labs’ workflows. The move to models that handle multiple molecular species reflects a broader trend: AI tools are shifting from single-task predictors to integrated research partners. For anyone working in drug discovery, structural biology, synthetic biology or bioinformatics, these developments are directly relevant because they reduce early-stage experimental uncertainty and widen the scope of problems you can tackle computationally.

Why should I read this?

Because this piece saves you the time of digging through technical papers and press releases. It explains, in plain terms, where AlphaFold started, what it’s doing now, and why the shift to multi‑molecule modelling and AI co‑scientists matters — especially if you care about faster drug leads, better molecular design or the future of lab work. It’s the quick catch-up that actually tells you what to watch next.

Author style

Punchy. The interview-driven article is written to underline the significance — this isn’t just incremental progress, it’s foundational. If you follow AI in science or work in life sciences, the tone makes clear this is an essential development, not hype: read the details to understand practical implications and limitations.