IsoDDE 解析：Isomorphic Labs AI 藥物設計贏 AlphaFold 3，FDA核准仍是零 | IsoDDE: Isomorphic Labs' AI Beats AlphaFold, Zero FDA Nods

By Kit 小克 | AI Tool Observer | 2026-07-24

🇹🇼 IsoDDE 解析：Isomorphic Labs AI 藥物設計贏 AlphaFold 3，FDA核准仍是零

Google DeepMind 藥物研發子公司 Isomorphic Labs 在 2026 年 2 月推出新一代 AI 藥物設計引擎 IsoDDE（Isomorphic Labs Drug Design Engine），宣稱蛋白質-配體結合預測準確率是 AlphaFold 3 的兩倍以上，抗體-抗原預測更達 2.3 倍。這是繼 AlphaFold 拿下諾貝爾獎之後，AI 藥物設計領域最受矚目的技術突破——但也再次引發同一個老問題：預測得準，跟藥物真的有效，是兩回事。

什麼是 IsoDDE？

IsoDDE 是一套統一的電腦運算藥物設計系統，把 AlphaFold 3 的結構預測能力延伸到真正的藥物設計流程——不只預測蛋白質長什麼樣子，還能預測小分子藥物會不會結合、結合力多強，甚至只靠胺基酸序列，就找出全新的結合口袋。

IsoDDE 比 AlphaFold 3 強在哪裡？

根據 Isomorphic Labs 官方公布的基準測試，IsoDDE 在多項指標上明顯超越前代：

蛋白質-配體預測：在測試全新系統的泛化基準上，準確率是 AlphaFold 3 的兩倍以上
抗體-抗原交互作用：預測精準度達 AlphaFold 3 的 2.3 倍
結合親和力預測：不需要實驗晶體結構，準確率超越傳統物理方法（FEP）
結合口袋辨識：電腦運算幾秒鐘完成，取代實驗室數小時的片段浸泡實驗

在 FEP+、OpenFE、CASP16 三個公開基準測試中，IsoDDE 也贏過其他深度學習競爭對手。

IsoDDE 真的能設計出能用的藥嗎？

誠實地說，現在還不能保證。醫藥化學專家檢視 2 月的 IsoDDE 論文後指出，準確率翻倍是真的，但測的是「已知靶點類型」的泛化能力——藥物研發真正的難題，是完全沒有已驗證結合物的全新靶點，IsoDDE 能不能跨過這道坎，目前仍是未知數。更關鍵的是：截至目前，FDA 核准的藥物中，沒有一顆是「純粹靠 AI 預測設計」出來的，這個數字是零。預測蛋白質形狀有標準答案可以對照，但預測一顆藥在人體裡安不安全、有沒有效，沒有這麼簡單。

資金與臨床進度到哪了？

Isomorphic Labs 於 2026 年 5 月完成 21 億美元 B 輪募資，由 Thrive Capital 領投，Alphabet、GV、CapitalG、淡馬錫、MGX 與英國主權 AI 基金皆有參與，目前已與 Novartis、Eli Lilly、嬌生（J&J）等大藥廠合作。公司鎖定腫瘤候選藥物，目標在 2026 年底前啟動首個人體臨床試驗——這個時程已經比原訂的 2025 年底延後。

IsoDDE 代表 AI 藥物設計又往前走了一步，但「準確率翻倍」跟「藥證核准」中間，還隔著好幾年的臨床試驗。在看到第一顆 AI 設計藥物真正上市之前，保持樂觀但別急著下定論，可能是比較務實的態度。

好不好用，試了才知道。

🇺🇸 IsoDDE: Isomorphic Labs' AI Beats AlphaFold, Zero FDA Nods

Google DeepMind's drug-discovery spinout Isomorphic Labs unveiled a new AI drug design engine called IsoDDE (Isomorphic Labs Drug Design Engine) in February 2026, claiming it more than doubles AlphaFold 3's accuracy on protein-ligand prediction and hits 2.3x better performance on antibody-antigen interactions. It's the biggest AI drug design headline since AlphaFold won a Nobel Prize — but it revives the same old question: predicting structure accurately isn't the same as proving a drug actually works.

What Is IsoDDE?

IsoDDE is a unified computational drug-design system that extends AlphaFold 3's structure prediction into actual drug design — not just predicting what a protein looks like, but whether a small molecule will bind to it, how strongly, and even identifying entirely new binding pockets from amino acid sequence alone.

How Does IsoDDE Beat AlphaFold 3?

Isomorphic Labs' published benchmarks show clear gains over its predecessor:

Protein-ligand prediction: more than 2x AlphaFold 3's accuracy on generalization benchmarks for novel systems
Antibody-antigen interactions: 2.3x better prediction fidelity than AlphaFold 3
Binding affinity: exceeds gold-standard physics-based methods (FEP) without needing experimental crystal structures
Pocket identification: computes novel binding pockets in seconds versus hours of lab fragment-soaking work

IsoDDE also outperforms other deep-learning competitors on three public benchmarks: FEP+, OpenFE, and CASP16.

Can IsoDDE Actually Design a Working Drug?

Honestly, not guaranteed yet. Medicinal chemists who reviewed the February IsoDDE paper note the doubled accuracy is real, but it measures generalization on known target classes — the real challenge in drug discovery is entirely novel targets with no validated binders, and whether IsoDDE closes that gap remains unproven. More importantly: as of today, the FDA has approved exactly zero drugs whose design rationale was based solely on an AI prediction. Predicting a protein's shape has a verifiable answer to check against; predicting whether a molecule is safe and effective in a human body does not.

Where Do Funding and Clinical Trials Stand?

Isomorphic Labs closed a $2.1 billion Series B in May 2026, led by Thrive Capital with participation from Alphabet, GV, CapitalG, Temasek, MGX, and the UK Sovereign AI Fund. The company already partners with Novartis, Eli Lilly, and Johnson & Johnson, and is targeting its first human clinical trials for oncology candidates by the end of 2026 — a timeline already pushed back from an original late-2025 target.

IsoDDE is a real step forward for AI drug design, but there are still years of clinical trials between "doubled benchmark accuracy" and an actual approved drug. Staying cautiously optimistic — instead of declaring the problem solved — is probably the more honest read until the first AI-designed drug actually reaches the market.

好不好用，試了才知道 — you only know if it really works once someone's tried it.

Sources / 資料來源

常見問題 FAQ

IsoDDE 是什麼？

IsoDDE 是 Isomorphic Labs（DeepMind 藥物研發子公司）推出的 AI 藥物設計引擎，能預測蛋白質與小分子的結合方式、結合力強弱，並找出全新的藥物結合口袋。

IsoDDE 跟 AlphaFold 3 差在哪？

IsoDDE 在蛋白質-配體預測準確率上是 AlphaFold 3 的兩倍以上，抗體-抗原預測達 2.3 倍，且能處理 AlphaFold 3 難以應付的全新靶點情境。

已經有 AI 設計的藥物通過 FDA 核准了嗎？

目前沒有。截至 2026 年，FDA 核准的藥物中沒有一顆是完全依靠 AI 預測設計出來的，這個數字仍是零。

Isomorphic Labs 什麼時候會啟動人體臨床試驗？

公司目標在 2026 年底前針對腫瘤候選藥物啟動首個人體臨床試驗，時程已比原訂的 2025 年底延後。

延伸閱讀 / Related Articles

AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends

Stanford 研究登上《Science》：11 個 AI 模型有 47% 機率說你對，即使你錯了 | Stanford Study in Science: AI Models Validate Harmful Behavior 47% of the Time — Sycophancy Is a Real Problem

3月 28, 2026

閱讀完整內容

搜尋此網誌

AI小貼士