protein-binder-design

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Protein Binder Design (workflow)

蛋白质结合剂设计（工作流）

Run a de novo binder design campaign by composing atomic NIM skills. This skill owns orchestration, handoff contracts, filtering, validation, and the run manifest. It does NOT duplicate per-NIM API details — defer those to each atomic skill's

SKILL.md

通过组合原子NIM技能运行从头结合剂设计流程。本技能负责编排、交接协议、过滤、验证以及运行清单的管理。本技能不会重复每个NIM的API细节——相关细节请参考各原子技能的

SKILL.md

文档。

Composed skills

组合技能

Step	Skill	Owns
Backbones	`rfdiffusion-nim`	binder backbone PDBs (contigs + hotspots)
Sequences	`proteinmpnn-nim`	sequences for each backbone
Co-fold / score	`boltz2-nim` or `openfold3-nim`	binder–target complex + confidence / ipTM
MSA (optional)	`msa-search-nim`	target A3M for higher-quality folding

The atomic NIM skills are recommended companions (one per NIM, from the BioNeMo NIM skill set). They are not required:

references/pipeline.md

carries the concrete request shape for every NIM call, so an agent with NIM access can follow this skill standalone. For endpoints/auth see Configuration below.

步骤	技能	负责内容
骨架生成	`rfdiffusion-nim`	结合剂骨架PDB文件（contigs + 热点）
序列生成	`proteinmpnn-nim`	每个骨架对应的序列
共折叠/评分	`boltz2-nim` 或 `openfold3-nim`	结合剂-靶点复合物 + 置信度/ipTM
MSA（可选）	`msa-search-nim`	用于更高质量折叠的靶点A3M文件

原子NIM技能是推荐的配套工具（每个NIM对应一个，来自BioNeMo NIM技能集）。它们并非必需：

references/pipeline.md

包含了每个NIM调用的具体请求格式，因此拥有NIM访问权限的Agent可以单独使用本技能。关于端点/认证信息，请参见下方的配置部分。

Pipeline

流水线流程

Target prep — get the target PDB + epitope; map epitope/hotspot author residue numbers to RFdiffusion
```
hotspot_res
```
strings; optionally build a target MSA with
```
msa-search-nim
```
.
Backbones (
```
rfdiffusion-nim
```
) — binder contig +
```
hotspot_res
```
; N backbones.
Sequences (
```
proteinmpnn-nim
```
) — k sequences per backbone; drop the native/WT row from
```
mfasta
```
.
Co-fold + score (
```
boltz2-nim
```
/
```
openfold3-nim
```
) — co-fold binder+target; collect interface confidence (ipTM) and binder pLDDT.
Self-consistency — CA-RMSD between the RFdiffusion backbone and the predicted binder (
```
scripts/metrics.py
```
).
Filter + rank — apply thresholds; rank survivors; write manifest + CSV.

Full handoff contracts, branching, and the cost funnel:

references/pipeline.md

靶点预处理 — 获取靶点PDB文件+表位；将表位/热点的原始残基编号映射为RFdiffusion的
```
hotspot_res
```
字符串；可选地使用
```
msa-search-nim
```
构建靶点MSA。
骨架生成（
```
rfdiffusion-nim
```
） — 结合剂contig +
```
hotspot_res
```
；生成N个骨架。
序列生成（
```
proteinmpnn-nim
```
） — 每个骨架生成k条序列；从
```
mfasta
```
中移除天然/野生型（WT）行。
共折叠+评分（
```
boltz2-nim
```
/
```
openfold3-nim
```
） — 对结合剂+靶点进行共折叠；收集界面置信度（ipTM）和结合剂pLDDT值。
自一致性验证 — 计算RFdiffusion生成的骨架与预测结合剂之间的CA-RMSD（参考
```
scripts/metrics.py
```
）。
过滤+排名 — 应用阈值筛选；对留存的设计方案进行排名；生成清单+CSV文件。

完整的交接协议、分支逻辑以及成本漏斗请参考：

references/pipeline.md

。

Handoff contracts (the fragile glue)

交接协议（关键衔接规则）

RFdiffusion
```
output_pdb
```
→ ProteinMPNN
```
input_pdb
```
(inline PDB text).
ProteinMPNN
```
mfasta
```
→ Boltz2 binder polymer
```
sequence
```
(exclude the native/WT row; pair scores only with designed rows).
Epitope author residue numbers → 1-based sequence indices: remap with
```
scripts/pdb_utils.py:remap_to_seq_index
```
. RFdiffusion
```
hotspot_res
```
uses chain+author strings like
```
"A50"
```
; Boltz2 pocket/contacts use 1-based indices.
Boltz2 complex
```
.cif
```
→ binder chain → self-consistency RMSD vs the backbone.

RFdiffusion的
```
output_pdb
```
→ ProteinMPNN的
```
input_pdb
```
（内嵌PDB文本）。
ProteinMPNN的
```
mfasta
```
→ Boltz2结合剂聚合物的
```
sequence
```
（排除天然/野生型行；仅将分数与设计行配对）。
表位原始残基编号 → 基于1的序列索引：使用
```
scripts/pdb_utils.py:remap_to_seq_index
```
进行重映射。RFdiffusion的
```
hotspot_res
```
使用链+原始编号字符串（如
```
"A50"
```
）；Boltz2的口袋/接触点使用基于1的索引。
Boltz2复合物的
```
.cif
```
文件 → 提取结合剂链 → 计算与骨架的自一致性RMSD。

Run manifest (reproducibility backbone)

运行清单（可追溯性核心）

Every campaign writes

manifest.json

candidates.csv

) under a run dir via

scripts/manifest.py

. It records lineage, params, scores, artifacts, filter status, and controls — enabling ranking, resumability, validation, and the final report. Schema and usage:

references/manifest.md

每个设计流程都会通过

scripts/manifest.py

在运行目录下生成

manifest.json

（+

candidates.csv

）。它记录了溯源信息、参数、分数、产物、过滤状态以及对照信息——支持排名、可恢复性、验证和最终报告。Schema及使用方法请参考：

references/manifest.md

。

Filters (defaults)

过滤条件（默认值）

ipTM ≥ 0.8, binder pLDDT ≥ 80, self-consistency RMSD ≤ 2.0 Å.
Override per campaign and record overrides in the manifest
```
filters
```
.

ipTM ≥ 0.8，结合剂pLDDT ≥ 80，自一致性RMSD ≤ 2.0 Å。
可针对每个流程覆盖默认阈值，并将覆盖信息记录在清单的
```
filters
```
字段中。

Validation

验证机制

Always run controls and report a success rate, not just top scores. Negative controls via

scripts/controls.py

(scrambled sequences); positive controls = published binders re-scored through the same pipeline. Benchmark targets live in

assets/targets.json

(

scripts/registry.py

). Methodology and metric definitions:

references/validation.md

始终运行对照实验并报告成功率，而非仅展示最高分。阴性对照通过

scripts/controls.py

实现（打乱序列）；阳性对照=已发表的结合剂通过同一流水线重新评分。基准靶点存储在

assets/targets.json

中（参考

scripts/registry.py

）。方法学及指标定义请参考：

references/validation.md

。

Human-in-the-loop + cost

人机协作与成本控制

Confirm target, epitope/hotspots, binder length range, and hosted-vs-local with the user before generating backbones (AskUserQuestion).
Co-folding is the expensive stage: co-fold a capped shortlist, review, then expand. State hosted vs local once and reuse it across all NIM calls.

在生成骨架前，与用户确认靶点、表位/热点、结合剂长度范围以及托管式vs本地部署选项（AskUserQuestion）。
共折叠是成本最高的阶段：先对有限的候选列表进行共折叠，审核后再扩展范围。只需声明一次托管式或本地部署选项，并在所有NIM调用中复用该设置。

Responsible use

负责任使用规范

De novo binder design is dual-use. Decline requests aimed at enhancing pathogen fitness, toxin potency, or bioweapon function; keep designs to legitimate research and therapeutic intent.

从头结合剂设计具有双重用途。拒绝旨在增强病原体适应性、毒素效力或生物武器功能的请求；仅将设计用于合法的研究和治疗目的。

Configuration (NIM access)

配置（NIM访问）

Each composed NIM is reached over HTTP; choose hosted or local once and reuse it for every call:

Hosted (managed): base URL
```
https://health.api.nvidia.com/v1/...
```
per NIM at build.nvidia.com; set
```
NVIDIA_API_KEY
```
(sent as
```
Authorization: Bearer
```
). Read keys from the env — never hardcode them.
Local (self-hosted NGC containers): point each NIM at its local URL (e.g.
```
http://localhost:8000/...
```
); local NIMs need no auth header. To launch the NIMs yourself (docker run per NIM, persistent caches, health checks, and the GPU profile‑selection gotcha — some NIMs (e.g. Boltz2) need
```
NIM_MODEL_PROFILE
```
pinned on GPUs that have no bundled profile, while others (RFdiffusion/ProteinMPNN) auto‑select by compute capability): see references/local-nim-setup.md
.

Per-NIM paths, request/response schemas, and worked

curl

/Python examples live in

references/pipeline.md

每个组合的NIM都通过HTTP访问；只需选择一次托管式或本地部署，并在所有调用中复用该设置：

托管式（托管管理）：每个NIM的基础URL为
```
https://health.api.nvidia.com/v1/...
```
（来自build.nvidia.com）；设置
```
NVIDIA_API_KEY
```
（以
```
Authorization: Bearer
```
方式发送）。从环境变量中读取密钥——切勿硬编码。
本地（自托管NGC容器）：将每个NIM指向其本地URL（例如
```
http://localhost:8000/...
```
）；本地NIM无需认证头。如需自行启动NIM（每个NIM使用docker run，持久化缓存，健康检查，以及GPU配置文件选择注意事项——部分NIM（如Boltz2）在无捆绑配置文件的GPU上需要固定
```
NIM_MODEL_PROFILE
```
，而其他NIM（如RFdiffusion/ProteinMPNN）会根据计算能力自动选择）：请参考**
```
references/local-nim-setup.md
```
**。

每个NIM的路径、请求/响应Schema以及可用的

curl

/Python示例请参考

references/pipeline.md

。

Scripts

脚本说明

```
scripts/manifest.py
```
— campaign manifest (create / load / score / filter / rank / CSV).
```
scripts/pdb_utils.py
```
— PDB parse, chain extract, sequence, residue remap, CA coords.
```
scripts/metrics.py
```
— Kabsch CA-RMSD for self-consistency.
```
scripts/controls.py
```
— scrambled negative controls.
```
scripts/registry.py
```
+
```
assets/targets.json
```
— example benchmark target registry (illustrative epitopes — verify against the cited structure before a real campaign). Replace with your own targets.

```
scripts/manifest.py
```
— 流程清单（创建/加载/评分/过滤/排名/生成CSV）。
```
scripts/pdb_utils.py
```
— PDB解析、链提取、序列处理、残基重映射、CA坐标提取。
```
scripts/metrics.py
```
— 用于自一致性验证的Kabsch CA-RMSD计算。
```
scripts/controls.py
```
— 生成打乱序列的阴性对照。
```
scripts/registry.py
```
+
```
assets/targets.json
```
— 示例基准靶点注册表（示例性表位——在实际流程前请对照引用的结构进行验证）。可替换为您自己的靶点。