AlphaFold 自 2020 年 11 月亮相至今已满 5 年，从解决蛋白质折叠这个「根节点」问题切入，带来可达原子级精度的三维结构预测。其资料库已累积超过 2 亿个预测结构，接近涵盖已知蛋白质宇宙，并被遍及 190 个国家的近 350 万名研究者使用。2021 年的 Nature 论文至今被引用约 4 万次，显示其在生物与化学研究中的扩散速度与影响力。

从 AlphaFold 2 到 AlphaFold 3 的演进，重点是从「单一蛋白」扩展到蛋白质、DNA、RNA 与小分子/药物的交互作用；这也提高了生成式模型（如扩散模型）产生「结构幻觉」的风险，特别是在内在无序区域。DeepMind 的策略是以「创造性生成 + 严格验证」的架构配合信心分数，让使用者能辨识可靠度较低的预测，同时依靠多年来反复的实验室检验来建立信任。

展望未来 5 年，Kohli 强调三个方向：让模型能更像研究伙伴般推理与协作、把工具普及到每位科学家手中、以及追求更宏大的目标如完整人类细胞的精准模拟。新推出的「AI co-scientist」以 Gemini 2.0 的多代理辩论机制来产生与批判假说，宣称可把假说生成阶段大幅「压缩」，例如协助研究「海盗噬菌体」与细菌基因转移；但关键判断与实验验证仍由人类负责。

Five years after its November 2020 debut, AlphaFold scaled protein-folding prediction to near atomic 3D accuracy and into routine scientific use. Its database now exceeds 200 million predicted structures—close to the entire known protein universe—and is used by nearly 3.5 million researchers across 190 countries. The 2021 Nature paper describing the system has accrued about 40,000 citations, a quantitative signal of rapid adoption and sustained downstream impact in biology and chemistry.

AlphaFold 3 extends the scope from proteins alone to joint interactions among proteins, DNA, RNA, and small molecules/drugs, increasing both ambition and risk. With diffusion models that are more generative, concerns include “structural hallucinations,” especially in intrinsically disordered regions. DeepMind’s mitigation keeps a “creative generator + rigorous verifier” harness and adds confidence scores to flag lower-reliability predictions, while pointing to repeated lab testing over the five-year period as the practical validation loop.

Looking five years ahead, Kohli frames three priorities: stronger models that reason and collaborate like research partners, broader access so tools reach every scientist, and bolder targets such as an accurate simulation of a complete human cell. The new “AI co-scientist,” built on Gemini 2.0 as a multi-agent debating system, is presented as compressing hypothesis generation by rapidly synthesizing decades of literature; examples include work with Imperial College on “pirate phages” and mechanisms tied to antimicrobial resistance, with humans still designing and interpreting experiments.