AlphaFold 3 扩展到分子相互作用

AlphaFold 2 让许多人第一次把单链蛋白质的三维形状，当成可以日常调用的假设。细胞与药物里的问题却常常更挤：蛋白质怎样和 DNA、RNA、小分子、离子或抗体待在同一张图里？单体骨架对了，结合姿态仍可能错；错的对象往往比形状差一点更能带偏下游设计。

2024 年 5 月 8 日前后，Google DeepMind 与 Isomorphic Labs 发布 AlphaFold 3。目标从“一个大分子的形状”，扩到多类生物分子的联合结构。模型用统一表示处理不同分子类型，以 Pairformer 等模块更新配对信息，再用扩散式结构生成直接采样三维原子坐标——相对 AF2 的结构模块路径，生成方式明显转向迭代去噪。论文在蛋白质—配体、蛋白质—核酸、抗体相关复合物等类别上报告优于多种专用基线的指标。这些数字绑在论文设定的基准与版本上，是研究评估，不是监管意义上的药证材料。置信度与错误模式仍在；同一输入多次采样也可能得到不同姿态，需要结合置信度与物理检查筛选。

发布形态与能力一样响亮。首发以 AlphaFold Server 等形式提供受限访问，完整训练代码与权重并未以 AF2 那种研究友好方式同步敞开。科学共同体对可重复性、基准复核与商业边界提出批评。后来的访问政策调整属于余波，细节随时间变化；“服务器优先”是发布时刻的事实，不能把日后的开放回填成首日故事。Isomorphic Labs 将相关能力接入药物设计管线，说明机构押注点在相互作用，而不只在单体收藏。服务器配额、可提交的分子类型与商业使用条款，共同构成发布时刻的真实获取条件。

扩散进入结构生物学主流工具，与图像生成里的扩散浪潮跨域呼应，误差代价却完全不同：错一个原子坐标，通常高于错一张风景图。口袋几何、立体化学与张力，也未必被基准分数直接捕捉。论文中的提升应读作“在报告的任务分割上优于专用基线”，而不是“可以自动出临床候选”。

AlphaFold 3 没有把湿实验变成可选项。它把“复合物长什么样”的先验加宽，也把开放科学与产品化发布之间的张力，写进同一篇 Nature 论文的阅读体验。能力与获取条件并读，才是这一事件的完整账本；删去其中一半，会得到半部历史。

当预测对象变成复合物，错误从“形状差一点”变成“可能绑错位置”。风险形态变化，发布策略却更封闭，二者叠加，使批评不仅来自竞争，也来自可重复性文化。Pairformer 与扩散模块改写结构生成路径，服务器配额改写谁能立刻复现。两条改写必须并排出现在记述里。

AlphaFold 2 made single-chain protein shapes feel like a daily hypothesis tool. Questions in cells and drug design are usually more crowded: how does a protein sit with DNA, RNA, small molecules, ions, or antibodies in one figure? A correct monomer backbone can still carry a wrong binding pose; the wrong partner often misleads downstream design more than a slightly imperfect shape.

Around 8 May 2024, Google DeepMind and Isomorphic Labs released AlphaFold 3. The target moved from one macromolecule’s shape to joint structures of many biomolecule types. A unified representation handles different species; modules such as Pairformer update pair information; a diffusion-based structure module samples three-dimensional atomic coordinates—clearly shifting generation toward iterative denoising relative to AF2’s structure path. On protein–ligand, protein–nucleic-acid, and antibody-related complexes, the paper reports gains over several specialist baselines. Those numbers are bound to the paper’s splits and versions: research evaluation, not regulatory clearance. Confidence scores and failure modes remain; multiple samples from the same input may yield different poses and need confidence and physical checks.

Access was as loud as capability. Launch offered limited access mainly through forms such as AlphaFold Server; full training code and weights did not open in the research-friendly way AF2 had. The scientific community pressed on reproducibility, independent benchmarks, and commercial boundaries. Later policy changes belong to the aftershock and shift over time; “server-first” is the fact of the launch day and must not be back-filled with later openness. Isomorphic Labs plugging related capability into drug-design pipelines shows the institutional bet: interactions, not a museum of monomers. Server quotas, allowed molecule types, and commercial terms jointly formed the real access conditions at launch.

Diffusion enters mainstream structural biology with a cross-domain echo of image generation and a wholly different error cost—one wrong atomic coordinate usually hurts more than one wrong landscape. Pocket geometry, stereochemistry, and strain are not always captured by a single score. Paper gains should be read as “better than specialist baselines on the reported splits,” not as “automatic clinical candidates.”

AlphaFold 3 does not make wet labs optional. It widens the prior on what complexes look like and writes the tension between open science and productized release into the same Nature reading experience. Capability and access conditions belong on one ledger; delete either half and the history is only half true.

When the prediction target becomes a complex, error shifts from “shape a bit off” to “maybe bound in the wrong place.” Risk changes shape while release strategy grows more closed; stacked, criticism comes not only from competition but from reproducibility culture. Pairformer and diffusion rewrite the structure path; server quotas rewrite who can reproduce at once. Both rewrites must sit side by side in the account.