新智元报道
近日，来自中国的AI 制药公司觅深知源（Molence Bio）发布其首个生物分子基础模型 Mol-1，在基础结构预测、超大复合物组装和纳米抗体从头设计等方面取得了振奋人心的成果，其中：
- 在抗体–抗原结构预测评估中：Molence Bio 发布的 Mol-1，追平了前沿闭源模型，达到75%，与Deepmind孵化公司Isomorphic Labs的闭源模型IsoDDE的成功率相当，远超其他开源模型。在其他重要的药物模态，如多肽、小分子、分子胶的预测中，也全面领先于开源模型。
- 超大复合物组装任务中：基于 Mol-1 的长时运行智能体使用单张 NVIDIA H20 GPU，在评估的 10 个大型复合物中，有 9 个案例优于 Anthropic 此前报告的对应结果；后者曾使用配备 8 张 B300 GPU 的节点探索超大体系预测。
- 针对纳米抗体从头设计任务：基于 Mol-1 早期版本开发的内部抗体设计平台，在包括4个GPCR在内的 17 个靶点上均获得结合候选，单个靶点的结合命中率最高达到63%，针对BCMA的30个候选中，最高亲和力达到43 pM；单轮实验反馈后，4-1BB激动抗体设计的第二轮功能命中率达到85%；成功设计出GPCR拮抗抗体；并按需设计出可交叉结合序列一致性仅为54%的蛋白抗原的VHH。
Mol-1 目前已经在GitHub上公开了源代码和模型权重，并提供了在线网页服务：
开源代码：github.com/molence/Mol-1
在线试用：lab.molence.com
Mol-1 使用单个模型，即可预测抗体–抗原、蛋白质–多肽、蛋白质–小分子及分子胶等多种复合物。在本次评估的多个药物模态上，Mol-1 均显著领先于其他开源模型。
在 FoldBench-AB 抗体–抗原测试集上，Mol-1 的成功率达到 75.0%，与 Isomorphic Labs 开发的 IsoDDE 技术报告公布的 75.6% 相当；加入表位约束后，进一步提升至 88.3%（Mol-1 Constraint）。
为考察结果对随机种子选择的敏感性，Molence还分析了从 32 个种子中选取的全部 201,376 个五种子组合。其中，62.22% 的组合达到至少 75% 的成功率，17.85% 的组合超过 76%。
作为通用的生物分子基础模型，Mol-1 使用单个模型，即可预测抗体–抗原、蛋白质–多肽、蛋白质–小分子及分子胶等多种复合物。
Molence Bio 在小分子、多肽和分子胶等更多药物分子相关任务上对其进行了评测，Mol-1 在每一类评测中均为开源模型的最优水平。在分子胶预测任务上，Mol-1 的优势尤为明显。
分子胶能够诱导或稳定蛋白质间的相互作用；对于这类挑战性的三元复合物体系，Mol-1 在高同源性子集上的预测成功率达到 74.6%，高于第二名的 40.7%。
在蛋白质–多肽复合物预测任务上，Molence Bio构建了 MolBench-Peptide-26，覆盖 2026 年前三季度公布的 150 个 PDB 条目中的 172 个相互作用界面簇。
在 DockQ > 0.23、> 0.49 和 > 0.80 三个阈值下，Mol-1 的预测成功率分别达到 88.8%、77.3% 和 54.1%，均超过了其他开源模型。
在最严格的 DockQ > 0.80 阈值下，Mol-1 的成功率为 54.1%，表现最好的无约束对照模型为 52.3%；Mol-1-constraint 则将成功率进一步提升至 60.2%，展现出对多肽结合界面的高精度建模能力。
超大蛋白质复合物的结构预测，是当前结构预测模型最具挑战性的任务之一。
这类复合物通常由几十甚至上百条蛋白质链组成，包括蛋白酶体、p97等与肿瘤研究密切相关的重要分子机器。随着复合物规模增加，模型不仅需要准确预测局部相互作用，还需要保证数百条链在三维空间中的正确组装，对结构预测能力提出了极高要求。
在 Anthropic 此前发布的工作中，曾使用 8张 B300 GPU 来预测大型分子复合物，最大准确预测规模为大约一万个氨基酸。但当体系扩大至约 3.1 万至 7 万残基时，在这 7 个体系中，均未能准确预测整体结构。
针对这一挑战，Molence Bio 构建了一套长时运行的智能体，将生物学知识带入建模过程，组织 Mol-1 结构预测与自动化组装，并根据中间结果持续调整策略。在 10 个大型复合物的评估中，有9 个预测结果超越 Anthropic 此前报告的结果，其中 8 个复合物的 TM-score 达到 0.9 以上。
最大的成功预测体系达到6万氨基酸残基、120 条蛋白质链，TM-score 高达 0.997（最高为1），规模为 Anthropic 此前最大准确预测对象的 6 倍。更重要的是，整套工作流仅使用单张 NVIDIA H20 GPU，使得算力需求大幅降低，研究者不需要最先进的GPU即可完成相关工作。
Molence Bio 使用 Mol-1 的早期模型版本，开发了内部的抗体从头设计平台，并在内部实验室进行了设计分子的实验验证，在测试的 17 个靶点上均获得结合候选，单个靶点的结合命中率最高达到 63%。
针对来自 13 个靶点的 30 个 VHH-Fc 候选，系统评估了疏水性、热稳定性、自相互作用及非特异性结合等关键生物物理性质。其中 23 个候选满足了至少三项筛选标准，13 个靶点中有 12 个获得至少一个满足要求的候选分子，显示出优良的后续开发潜力。
BCMA 是多发性骨髓瘤治疗中的重要靶点，已广泛应用于TCE和CAR-T的开发。
Molence Bio 从头设计了 30 个 VHH 候选，其中 12 个可与 BCMA 结合，最优候选在 SPR 实验中测得的 KD 为 43 pM，团队还进一步测试了设计的 VHH 对细胞表面 BCMA 的识别能力， EC50 最低达到 0.16 nM，显著低于对照 belantamab 的 4.1 nM。
利用 Mol-1 ，Molence Bio 还设计了针对 4-1BB 的激动型VHH，希望通过抗体结合激活这一重要的免疫共刺激受体。
在首轮设计中，30 个候选中有 2 个表现出激动活性，其细胞激活 EC50 分别为 0.05 nM 和 0.22 nM，均低于同一实验中 Utomilumab 对照的 0.43 nM，显示出较强的受体激活活性。
更值得关注的是，团队仅经过一轮实验反馈，第二轮 60 个设计中有 51 个表现出功能活性，功能命中率从首轮的约 7% 大幅提升至 85%。
这一结果表明，Molence Bio的抗体设计平台，可以通过实验反馈持续优化，将结构预测、抗体设计与实验验证结合成高效的迭代闭环。
在 GPCR 靶点 CXCR6 上，设计的 VHH 不仅能够结合细胞表面的 CXCR6，EC50值为39 nM，还能够抑制其天然配体 CXCL16 诱导的下游信号，展示了利用该平台设计具有特定拮抗功能的抗体的可行性。
Molence Bio 进一步尝试了设计具有预设交叉反应性的抗体，以结合两种序列一致性仅为 54% 的细胞因子为目标，希望设计出能够同时识别两种抗原的 VHH。
在首轮 30 个设计中，有 5 个候选成功结合两种抗原，证明模型能够根据预先指定的多靶点识别要求进行从头设计。这一结果展示了从头设计能够同时识别两种非直系同源抗原的 VHH 的可行性。
从精度领先的全模态分子结构预测，到上百条蛋白质链的超大复合物组装，再到VHH按需从头设计的全面实验验证，Mol-1 展示了结构预测基础模型在药物研发场景下的广泛应用潜力。
目前，Molence Bio 已在 GitHub 上公开 Mol-1 的源代码和基础模型权重，同时提供 Mol-1 和 Mol-1-constraint 的在线预测服务。科研人员不仅可以直接通过网页使用模型，而且可以在自己的计算环境中部署 Mol-1，构建更复杂的预测、设计和智能体工作流。
编辑：大卫
不喜欢
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
New Ji-won.
In recent days, the AI pharmaceutical company from China, seeking the first Biomolecular Base Model (Molence Bio) publish, has achieved encouraging results in infrastructure forecasting, super-large composite assembly and nano-antibodies design, among which:
- In the antibody-antigen structure prediction assessment: Mol-1 of Molence Bio Z94Z, the front-line closed-source model was smoothed to 75 per cent, comparable to the closed-source model of Isomorphic Labs of Deepmind Incubation, which is success rate of IsodDE, far above other open-source models. Other important drug models, such as polyplyptunium, small molecules and molecular glues, are also fully ahead of open source models.
- Super-composite assembly missions: single NVIDIA H20 GPU for long run AI agent based on Mol-1. Of the 10 large compounds assessed, 9 cases were better than the corresponding results reported by Anthropic; the latter used 8 B300 GPUs to explore hypersystem predictions.
- Top-down design missions for nanoantibodies: an internal antibody design platform based on the early version of Mol-1 was selected on 17 target points, including 4 GPCRs, with a maximum combined impact rate of 63 per cent for individual targets and 43 pM for BCMAs out of 30 candidates; a second-round functional target rate of 85 per cent for 4-1BB agitated antibodies was achieved following feedback from single-wheel experiments; a GPCR resistant body was successfully designed; and a VHH designed for cross-linking sequence consistency was only 54 per cent of the protein antigens.
Mol-1 has now published the source code and model weights on GitHub and provides online web services:
Open source: Github.com/molence/Mol-1
Online trial: lab.molence.com
Mol-1 uses a variety of complexes such as predictable antibodies - antigens, proteins - polyazine, proteins - small molecules and molecular glue. Mol-1 is significantly ahead of other open-source models in a number of drug models in this assessment.
success rate of Mol-1 reached 75.0 per cent on the FoldBech-AB Antigen Test Set, similar to the 75.6 per cent published in the Isodode technical report developed by Isomorphic Labs; it was further raised to 88.3 per cent (Mol-1 Constrant) after the tablet binding.
To examine the sensitivity of the results to random seed selection, Molence also analysed all 201,376 combinations of seeds selected from 32 seeds. Of these, 62.22% had a combination of at least 75% success rate, and 17.85% had a combination of over 76%.
As a generic biomolecule base model, Mol-1 uses a variety of complexes such as predictable antibodies - antigens, proteins - polyazine, proteins - small molecules and molecular glue.
Molence Bio has assessed the additional drug molecule-related tasks of small molecules, peptide and molecular glue, and Mol-1 is the best level of open source models in each category. The advantages of Mol-1 are particularly evident in molecular glue forecasting missions.
Moll-1 predicts that success rate will reach 74.6 per cent on the high homogenic subset, up from 40.7 per cent in the second place.
In the protein-polymer compound prediction mission, Molence Bio built MolBech-Peptide-26, covering 172 interfaces of 150 PDF entries published in the third quarter of 2026.
Under the three thresholds of DockQ > 0.23, > 0.49 and > 0.80, Mol-1 projects that 88.8 per cent, 77.3 per cent and 54.1 per cent, respectively, exceed other open source models.
At the strictest DockQ > 0.80, the success rate for Mol-1 was 54.1 per cent and the best unbridled contrast model was 52.3 per cent, while Mol-1-constraint further raised success rate to 60.2 per cent, demonstrating a high-precision modelling capability for the polypermium interface.
Structural prediction of hyperprotein compounds is one of the most challenging tasks of the current structural prediction model.
Such compounds typically consist of dozens or even hundreds of protein chains, including protein enzymes, and important molecular machines such as p97, which are closely related to oncology research. As compounds grow in size, the model requires not only accurate prediction of local interactions, but also assurance of the correct assembly of hundreds of chains in a three-dimensional space, placing very high demands on structural prediction capabilities.
In Anthropic's previous work with publish, eight B300 GPUs were used to predict large molecular complexes, with a maximum accuracy projection of approximately 10,000 amino acids. But when the system was expanded to about 31,000 to 70,000 remains, none of the seven systems had been able to predict the overall structure accurately.
In response to this challenge, Molence Bio constructed a long-lived set of AI agentes, bringing biological knowledge into the modelling process, organizing the Mol-1 structural prediction and automated assembly, and continuously adapting strategies to intermediate results. Of the 10 large compounds assessed, 9 projected results exceeded the results previously reported by Anthropic, of which 8 were TM-score with more than 0.9.
The largest success projection system reached 60,000 amino acid residuals, 120 protein chains, TM-score up to 0.997 (up to 1) and was six times the size of anthropic ' s largest accurate projection target. More importantly, the entire package of workflow uses a single sheet of NVIDIA H20 GPU, which reduces the need for computing power significantly and does not require state-of-the-art GPU to complete the work.
Molence Bio, using an earlier model version of Mol-1, developed an in-house antibody top-down design platform and conducted an experimental test test of design molecules in an in-house laboratory, which obtained a combined candidate for each of the 17 targets tested, with a combined impact rate of up to 63 per cent for each target.
For the 30 VHH-Fc candidates from 13 target points, the critical biophysical properties of hydrophobicity, thermal stability, self-interacting and non-specific combinations were systematically assessed. Of these, 23 candidates met at least three screening criteria and 12 of the 13 target points received at least one candidate who met the requirements, indicating a good development potential.
BCMA is an important target point in the treatment of multiple osteoporosis and has been widely used in the development of TCE and CAR-T.
Mollence Bio designed 30 VHH candidates from the top, 12 of which could be combined with BCMA, the best candidate measured in the SPR experiment was KD 43 pM, and the team further tested the design VHH's ability to identify BCMA on the cell surface, with EC50 having a minimum of 0.16 nM, significantly lower than 4.1 nM compared to belantamab.
Using Mol-1, Molence Bio also designed an exciting VHH for 4-1BB to activate this important immuno-stimulant receptor through antibody combinations.
In the first round of design, 2 out of 30 candidates showed an agitated activity, with EC50 cell-activation of 0.05 nM and 0.22 nM, respectively, which was lower than the 0.43 nM comparison of Utomilumab in the same experiment, showing a strong receptor-activation activity.
Even more noteworthy is the fact that the team has only undergone a round of experimental feedback, with 51 out of 60 second-round designs showing functional activity and a significant increase in the functional impact rate from about 7 per cent in the first round to 85 per cent.
This results suggest that the antibody design platform of Molence Bio can be continuously optimized through experimental feedback, combining structural prediction, antibody design and experimental validation into an efficient iterative closed loop.
On the GPCR target CXCR6, the VHH designed not only combines CXCR6 with EC50 values of 39 nM on the surface of the cell, but also inhibits downstream signals induced by its natural formulation CXCL16, demonstrating the feasibility of using the platform to design antibodies with specific resistance functions.
Molence Bio has further tried to design antibodies with predefined cross-reactive properties, aiming at combining the two series of cytogens, which are only 54%, with the hope of designing VHHs that identify both antigens.
In the first round of 30 designs, 5 candidates successfully combined the two antigens, demonstrating that the model can be designed from scratch according to pre-defined multi-target identification requirements. This results demonstrate the feasibility of a top-up design that identifies both non-directive congeners of antigens.
From an accuracy-leading full-modular molecular structure predictions, to the formation of hundreds of hypercomplex protein chains, and to a comprehensive experimental validation of a needs-based top-down design by VHH, Mol-1 demonstrates the potential for extensive application of structural prediction infrastructure models in drug development settings.
The source code and base model weights for Mol-1 are currently available on the GitHub, along with the online forecasting services for Mol-1 and Mol-1-construint. Researchers can not only use models directly through web pages, but also deploy Mol-1 in their own computing environment to build more complex predictions, designs and AI agentworkflow.
Editor: David
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.