整理日期:2026-10-31 承接文档:ETH 假设证伪对制药计算的影响 第三节末"对比:真正在降算力的东西"表 检索方式:全部通过
/anysearchskill(batch_search/extract)完成 标注约定:[事实] = 来源中可直接核实;[推断] = 我基于公开信息做的关联判断;[口径] = 厂商/媒体口径,非同行评议结论
原文档第三节给出的"真正在降算力"的 4 类手段是:
| # | 手段 | 原文给出的收益 |
|---|---|---|
| A | V-SYNTHES / Chemical Space Docking 类:组合空间剪枝,不枚举 10¹⁵ | ~5000× |
| B | ML 势(ANI、NequIP 等)替代经典力场做 MD | 10–1000× |
| C | AF3 扩散步数 / 一致性采样优化、GPU 化 | ~10× |
| D | MSA 复用、片段库预计算 | 10–100× |
本轮调研要回答:这 4 种手段中,哪些在最近的 AI 制药 / 癌症制药突破里被实际用上并被证明仍有继续研究价值?
| 手段 | 与近期突破的相关性 | 判断 | 代表性近期突破 |
|---|---|---|---|
| A 化学空间剪枝 / 超大规模筛选 | ★★★★★ | 价值最高、最直接。已被 2025–2026 的"单日 3000 亿分子"级筛选和生成式化学(Insilico / Isomorphic / Model Medicines)当作核心竞争力验证 | Model Medicines GALILEO 单日筛 3250 亿分子 → 首创类 BRD4 抑制剂 MDL-4102(肿瘤);Insilico rentosertib(Phase IIa→III)、CTX-439(CDK12/13,乳腺癌);Isomorphic IsoDDE(2026-09 进入人体试验) |
| B ML 势 / 加速 MD | ★★★★☆ | 价值确定、增量式。癌症药 zovegalisib(Phase III)整条路线建立在 MD 构象集合之上;ML 力场与 ML 增强 FEP 已进生产级工具;但仍是"更准更快的物理",不是范式突破 | Relay zovegalisib(PI3Kα 变构,乳腺癌,Phase III);Qubit Dual-LAO(RBFE 15–30×);MLFF 打分对标 FEP(chemrxiv);AIQM-PBSA(2026) |
| C 扩散采样 / 快速推理 | ★★★★★ | 增长最快、争议最少。结构预测与 binder 设计从"数月/数年"压缩到"数天",直接来自采样数、种子数、步数的削减 | Biohub ESMFold2(2026-05,5 个癌症/免疫靶点 binder,实验验证有效);Boltz-2(亲和力+FEP);BoltzGen(通用 binder 设计);RFdiffusion3(全原子从头设计);Protenix-v1 / HelixFold3 |
| D MSA 复用 / 片段库预计算 | ★★☆☆☆ | 研究价值最低、正在被绕过。AF3 复合物预测与 ESM 系(ESMC/ESMFold2)都不再依赖 MSA/模板;片段库在 de novo 设计时代边缘化。它从"研究课题"降级为"缓存与基建" | ESMFold2(免 MSA、SOTA 速度);Boltz 系走 ColabFold MSA server 缓存;Rowan 对 MSA 失效的讨论 |
一句话:A 与 C 是当前 AI 制药突破的实际承重墙,B 是可靠的第二引擎,D 正在退场。
[事实] Model Medicines(GALILEO 平台,Google Cloud)2025 年底完成单日 3250 亿分子规模的虚拟筛选,产出 MDL-4102——一个首创类(first-in-class)BRD4 抑制剂,用于肿瘤,且对同家族 BRD2/BRD3 无可测活性;ECFP4-Tanimoto 与所有既往临床 BET 抑制剂结构相异。
[事实] 其算力策略正是"剪枝"的极致形态:
与 V-SYNTHES 的关系:V-SYNTHES(Nature 2022)的核心是"用模型先剪掉绝大多数不可能结合的分子,再做物理对接"(11B 空间,需求降 5000×)。GALILEO 把同一范式推到 10¹¹×10² 量级。2026 年的新证据说明:2022 年这套剪枝思路不但没过时,反而成了肿瘤新药的直接产出口。
[事实] Isomorphic Labs 2026-09-29 官方文章《Building a new path to make medicines with AI》把方法论讲得非常直白:
[事实] 另有报道称 Isomorphic Labs 已就 AI 设计的癌症药物启动人体试验(chemdiv 报道,2026)。
[事实] Relay Therapeutics 的 zovegalisib(RLY-2608) 是首个突变体选择性变构 PI3Kα 抑制剂,用于 HR+/HER2 乳腺癌;其平台定位为 motion-based drug design——用蛋白动力学的计算模拟来识别传统结构基方法找不到的变构位点。临床数据:二线联合 fulvestrant 11 个月中位 PFS,ORR 39%,已启动 Phase III ReDiscover-2(对照 AZ 的 capivasertib)。
[事实] Relay 还在公开发表 MD 方法学工作:MDPath(J. Chem. Inf. Model.,"通过 MD 模拟解开药物靶点的变构通信路径"),并公开其 MD 模拟数据集(zenodo,含拓扑与轨迹文件)。
[推断] 这直接证明原文档的判断:"真正吞算力的是 MD 长程采样"。癌症药领域最成功的 AI 辅助药物之一,其差异化机制(变构位点)本质上只能用动力学采样得到,静态结构给不了。所以任何能降低 MD 成本 / 提高采样效率的技术,在这个方向上有明确的可付费研究价值。
| 路线 | 证据 |
|---|---|
| 采样算法改进 | Qubit Pharmaceuticals 2026-01-28:Dual-LAO(dual-topology + Lambda-ABF-OPES)把 RBFE 计算加速 15–30×(从几十 ns 降到几 ns),误差 0.5–0.6 kcal/mol 与经典 RBFE 持平,且对 scaffold 跳变、埋藏水排出、带电变化、开环/闭环等难变换更鲁棒;基于 Ansari / Aviat / Hénin / Piquemal / Lagardère 论文。[事实] |
| ML 打分替代 FEP | chemrxiv:Machine-learning force-field scoring rivals free-energy calculations(2026)[事实];AIQM-PBSA(chemrxiv 2026,把机器学习原子间势整合进端点自由能)[事实];JCIM 2025《Integrating Machine Learning into Free Energy Perturbation》[事实];KU:Machine Learning-Enhanced Calculation of Quantum-Classical Binding Free Energies [事实] |
| 神经势替代经典力场 | AceFF(arXiv 2601.00581,小分子专用 ML 势)[事实];Bifold 的"用碎片化学训练QM级力的蛋白 MD" [事实];NequIP/ANI 系列;Synopsys 博客声称 10,000× 加速 [口径,厂商] |
Boltz-2(MIT Jameel Clinic + Recursion + NVIDIA,2025-06)把结构 + 结合亲和力 + FEP 预测塞进一个开源模型,宣称在"速度、规模、精度"上建立新标准 [事实];其推理期 FEP 与 MD-FEP 的系统比较被广泛跟踪(如 Rowan 的 Boltz-2 benchmark 追踪页)[事实]。
[事实] Biohub 2026-05-27 发布 ESMC / ESMFold2 / ESM Atlas 三件套(开放):
这条直接落在手段 C 上:"数天 vs 数月/数年"的量级差异,本质是采样步数 × 种子数 × 结构预测单次成本三项的乘积被削减;同时 ESM Atlas(11 亿预计算结构)也正是"片段库预计算"思路的极端放大版——注意它属于手段 D 的精神延续,但技术基础已不是 MSA/片段。
| 证据 | 内容 |
|---|---|
| AlphaFold 3(技术报告) | 扩散架构取代 Evoformer,复合物预测路径不再依赖 MSA 与模板 |
| ESMC / ESMFold2(Biohub 2026-05) | 纯语言模型路线,不需要 MSA;ESM Atlas 用 28 亿序列 + 11 亿预计算结构替代"每次现算" |
| Boltz 系 | 通过 ColabFold 公共 MSA API 拉取 MSA 并缓存复用——MSA 从"算法问题"降级为"API 缓存问题" |
| Rowan 讨论 | 有专门文章讨论 MSA failures 与"ColabFold MSA 失效"对折叠管线的影响——问题从"怎么算"变成"缓存不一致/污染" |
| 片段库 | RoseTTAFold All-Atom / RFdiffusion 系列不再使用经典片段库;FRAGSION 等片段库工作停留在 de novo 结构预测的传统脉络 |
| ColabFold | 仍以"免本地装 ColabFoldDB + 复用公共 MSA 服务器"为主要加速手段 [事实] |
| 突破(时间) | 领域 | 核心省算力手段 | 强度 |
|---|---|---|---|
| MDL-4102(2025 末,Model Medicines GALILEO,单日 3250 亿分子) | 肿瘤(BRD4,首创类) | A 剪枝 + CPU 推理规模化 | ★★★★★ |
| ESMFold2 binder 设计(2026-05,Biohub,5 个癌症/免疫靶点,数天 vs 数月/数年) | 肿瘤+免疫 | C 快速采样;D 的"预计算结构库"变体 | ★★★★★ |
| rentoſertib / ISM001-055(2024 Phase IIa→2026 Phase III,Nature Medicine) | 纤维化 | A 生成+剪枝(PandaOmics/Chem42) | ★★★★☆ |
| CTX-439 CDK12/13 抑制剂(2026-02 预印本,乳腺癌 PDX) | 肿瘤 | A | ★★★★☆ |
| zovegalisib / RLY-2608(11 月中位 PFS,ORR 39%,Phase III) | 乳腺癌 | B MD/构象集合 | ★★★★☆ |
| Isomorphic IsoDDE + 癌症药人体试验(2026-09 起) | 肿瘤等 | A(+ B/C) | ★★★★☆ |
| REC-617 CDK7 抑制剂(Recursion,Phase I ELUCIDATE,实体瘤) | 肿瘤 | 不属这四类(细胞影像+深度学习) | — |
| Boltz-2 / BoltzGen / RFdiffusion3 / Protenix-v1 | 通用平台 | C(+ B) | ★★★★☆ |
| Eli Lilly × NVIDIA 10 亿美元(2026-01) | 行业 | A/B/C 全栈 | ★★★★☆ |
| 173 个 AI 药物项目在临床、肿瘤占 126 项(2026 初) | 全域 | A/B/C 合力 | 背景 |
| 原文档判断 | 本轮调研后的态度 |
|---|---|
| "ETH 证伪对制药算力的直接贡献 ≈ 0–2%" | 维持。本轮找到的所有近期突破,其省算力来源全部是数量维度(候选数、采样数、步数、轨迹长度、势函数速度),没有一项来自复杂度下界的松动 |
| "真正在降算力的是 V-SYNTHES(5000×)、ML 势(10–1000×)这类工程/算法改进" | 加强并细化:本轮发现 A 类在 2026 年被推到 10¹¹×10² 量级(3250 亿分子/天);C 类被证明能把 binder 设计的计算搜索从"数月/数年"降到"数天"(ESMFold2)。这两条是当前最有价值的赛道 |
| "P=NP 不会省算力,反而催生新负载" | 加强:近期突破的路径恰恰是"把原来算不起的枚举/采样变成常规操作"(3250 亿分子/天、11 亿预计算结构),这正是 P=NP 情景在实践层面已经在发生的事,只是驱动力是工程/模型而非复杂度理论 |
| 原文档未涉及 | 新增:D 类(MSA/片段库)研究价值在下降,原文档把它列为"10–100× 收益"的降算力手段,在 2026 年语境下应降级为 ops 项;此外原文档的"MD 吞算力"判断在癌症制药上得到具体验证(zovegalisib / MDPath) |
按"证据强度 × 与近期突破的耦合度"排序:
intuitionlabs.ai/articles/ai-ultra-large-virtual-screening-brd4modelmedicines.com/newsroom/record-scale-ai-screening-with-model-medicines-on-google-cloud-galileo-achieves-325-billion-molecule-throughput-for-oncology-drug-discoverysciencedirect.com/science/article/pii/S1359644626000218isomorphiclabs.com/articles/building-a-new-path-to-make-medicines-with-aiparticle.news/story/isomorphic-labs-raises-21-billion-to-scale-ai-drug-designmegaoneai.com/blog/ai-drug-discovery-trials-2026-first-results/nature.com/articles/s41591-025-03743-2;Insilico 官网 insilico.com/news/tnik-ipf-phase2adoi.org/10.64898/2026.02.20.706902scienmag.com/insilico-medicine-unveils-innovative-cdk12-13-dual-inhibitors-for-tumor-therapy-with-the-help-of-generative-ai/blog.qubit-pharmaceuticals.com/blog/going-15-30-faster-in-relative-binding-free-energy-calculationsrelaytx.com/pipeline/、relaytx.com/dynamo-platform/doi.org/10.1021/acs.jcim.5c01590;公开数据集 zenodo doi.org/10.5281/zenodo.15637794chemrxiv.org/doi/10.26434/chemrxiv.15008810chemrxiv.org/doi/10.26434/chemrxiv-2026-h0kn1doi.org/10.1021/acs.jcim.5c01449quantumcomputingreport.com(用于说明"计算降本"叙事的多元化)biohub.org/news/world-model-of-protein-biology/biorxiv.org/content/10.64898/2026.06.03.729735v1biorxiv.org/content/10.1101/2025.11.20.689494v2biorxiv.org/content/10.1101/2025.06.14.659707v1;jclinic.mit.edu/boltz-2-towards-accurate-and-efficient-binding-affinity-prediction/biorxiv.org/content/10.64898/2026.02.05.703733v1;github.com/bytedance/Protenixarxiv.org/abs/2408.16975v3nature.com/articles/s41586-023-06415-8axis-intelligence.com/ai-drug-discovery-2026-complete-analysis/ema.europa.eunature.com/articles/s41592-022-01488-1;github.com/sokrypton/colabfoldrowansci.com/blog/msa-failures-and-our-responsedoi.org/10.1093/bioinformatics/btw067bakerlab.org/wp-content/uploads/2026/05/s41586-026-10328-7.pdfcd .agents/skills/anysearch
python3 scripts/anysearch_cli.py batch_search --queries '[{"query":"V-SYNTHES virtual screening 5000x compute reduction Nature 2022 chemical space pruning drug discovery","max_results":6},
{"query":"machine learning interatomic potential NequIP ANI molecular dynamics speedup drug discovery 2026","max_results":6},
{"query":"AlphaFold3 diffusion sampling steps reduction faster inference 2026 AI drug discovery","max_results":6},
{"query":"MSA reuse fragment library precomputation protein structure prediction speedup","max_results":6}]'
python3 scripts/anysearch_cli.py batch_search --queries '[{"query":"AI-designed CDK12/13 inhibitors resistant cancers Insilico clinical trial","max_results":6,"freshness":"year"},
{"query":"Isomorphic Labs 2026 drug discovery milestone clinical trial AI designed","max_results":6,"freshness":"year"},
{"query":"Recursion Exscientia 2026 pipeline results AI drug discovery physics platform","max_results":6,"freshness":"year"},
{"query":"Qubit Pharmaceuticals 2026 machine learning potential free energy perturbation speedup","max_results":6,"freshness":"year"}]'
python3 scripts/anysearch_cli.py batch_search --queries '[{"query":"HelixFold3 Protenix open source fast protein structure prediction diffusion speedup minutes","max_results":6,"freshness":"year"},
{"query":"LiteMol diffusion molecular generation faster sampling 20x Microsoft","max_results":6},
{"query":"machine learning force field free energy perturbation pharmaceutical 2026 accuracy speedup production use","max_results":8,"freshness":"year"},
{"query":"AI-discovered drug phase 3 2026 oncology clinical success","max_results":8,"freshness":"year"}]'
python3 scripts/anysearch_cli.py batch_search --queries '[{"query":"Relay Therapeutics zovegalisib RLY-2608 motion-based drug design molecular dynamics platform allosteric","max_results":6,"freshness":"year"},
{"query":"Boltz-2 free energy prediction FEP BoltzGen binder design 2026 milestone","max_results":6,"freshness":"year"},
{"query":"Schrodinger machine learning force field 2026 production use pharma free energy calculations speedup","max_results":6,"freshness":"year"},
{"query":"MSA-free protein language model ESM3 evolutionary design oncology binder 2026","max_results":6,"freshness":"year"}]'
python3 scripts/anysearch_cli.py batch_search --queries '[{"query":"BoltzGen binder design sampling steps diffusion number of samples inference cost compute","max_results":6,"freshness":"year"},
{"query":"machine-learning force-field scoring rivals free-energy calculation FEP accuracy 2026","max_results":6,"freshness":"year"},
{"query":"AACR 2026 AI-driven drug discovery oncology virtual screening generative chemistry breakthroughs","max_results":8,"freshness":"year"},
{"query":"ColabFold MSA server cache reuse protein folding 2026 speedup AlphaFold3 no MSA","max_results":6,"freshness":"year"}]'
python3 scripts/anysearch_cli.py batch_search --queries '[{"query":"AlphaFold 3 does not use MSA templates single sequence input diffusion","max_results":6},
{"query":"MSA server ColabFold reuse cached MSA cost structure prediction pipeline 2026","max_results":6,"freshness":"year"},
{"query":"fragment library de novo protein design 2026 RFdiffusion RoseTTAFold all-atom compute efficiency","max_results":6,"freshness":"year"},
{"query":"ESMFold2 binder design cancer immunology targets preprint 2026 compute days","max_results":6,"freshness":"year"}]'
python3 scripts/anysearch_cli.py batch_search --queries '[{"query":"machine learning potential NequIP protein ligand free energy 2026 order of magnitude faster simulation pharmaceutical","max_results":6,"freshness":"year"},
{"query":"Relay Therapeutics compute platform 2026 simulations daily motion based drug design throughput","max_results":6,"freshness":"year"},
{"query":"AI ML potential molecular dynamics cancer drug discovery 2026 allosteric site conformational ensemble compute","max_results":6,"freshness":"year"}]'
# 全文抽取
python3 scripts/anysearch_cli.py extract "https://blog.qubit-pharmaceuticals.com/blog/going-15-30-faster-in-relative-binding-free-energy-calculations"
python3 scripts/anysearch_cli.py extract "https://www.isomorphiclabs.com/articles/building-a-new-path-to-make-medicines-with-ai"
python3 scripts/anysearch_cli.py extract "https://megaoneai.com/blog/ai-drug-discovery-trials-2026-first-results/"
python3 scripts/anysearch_cli.py extract "https://intuitionlabs.ai/articles/ai-ultra-large-virtual-screening-brd4"
python3 scripts/anysearch_cli.py extract "https://www.geneonline.com/insilico-medicine-showcases-four-ai-driven-oncology-drug-discovery-abstracts-at-aacr-2026/"
python3 scripts/anysearch_cli.py extract "https://biohub.org/news/world-model-of-protein-biology/"
注:
pmc.ncbi.nlm.nih.gov对extract返回反爬页面(浏览器校验),故 PMC 链接仅作指针列出,正文内容改由 bioRxiv / 官方新闻页 / 机构页交叉验证。