Vaccine Design Orchestrator

Use this skill when the user wants to evaluate a new nanoparticle vaccine candidate, redesign a computational screening workflow, define gate criteria, or pr...

Security Scan

The skill's documented workflow and tools match a vaccine-design orchestrator, but the metadata omits required tool/credential declarations and the instructions are somewhat open-ended about what the agent should execute, creating coherence and scope concerns.

Purpose & Capability

SKILL.md describes a complex computational vaccine-design pipeline (sequence design, AlphaFold/ESMFold, Boltz-1/AF3, GROMACS MD, CpHMD, TI/FEP, etc.). The skill metadata declares no required binaries, no install, and no environment/credentials. For a skill that expects to run or orchestrate heavy ML/MD tools, omitting required tools, runtimes, or compute access is an inconsistency: either the agent is expected to actually run these tools (in which case required binaries / compute credentials should be declared) or the skill only produces plans/SOPs (which should be stated explicitly).

Instruction Scope

The SKILL.md stays on-topic (design, prediction, MD, gating, outputs). It does not instruct reading unrelated system files or exfiltrating data. However instructions are broad and sometimes assume the agent will execute intensive simulations (e.g., 'GROMACS 100 ns') without clarifying whether the agent should run them itself, dispatch jobs to a user/HPC, or only produce job scripts and SOPs. That ambiguity grants wide discretion to the agent and should be clarified.

Install Mechanism

Instruction-only skill with no install spec and no code files. Lowest technical install risk. There are no downloaded artifacts or install scripts to analyze.

Credentials

The workflow implies need for GPUs, MD engines, possibly HPC credentials, license keys, or cloud accounts for heavy compute, but the skill declares no required environment variables, credentials, or config paths. This mismatch is disproportionate: either declare the compute/auth requirements or limit the skill to generating non-executing artifacts (SOPs, scripts).

Persistence & Privilege

Skill is user-invocable, not always: true, and does not request to modify other skills or system settings. Autonomous invocation is allowed by default but not combined with other privilege escalation indicators.

What to consider before installing

Before installing or enabling this skill, consider the following: - Clarify execution model: ask the skill author whether the agent will (A) only produce SOPs/job scripts and recommendations, or (B) actually execute AlphaFold/ESMFold/GROMACS/CpHMD/FEP jobs. If (B), do not enable until required binaries, cluster/cloud credentials, and access policies are explicitly declared. - Verify tool availability and credentials: running the described pipeline typically requires installed ML/structure tools, GROMACS, GPU/HPC access, and possibly licensed software. Ensure you control any credentials and that the skill metadata lists them. - Data and safety: the skill deals with computational vaccine design (potentially dual-use). Confirm how simulation data and sequence designs are stored, who can access them, and whether the skill will transmit outputs externally. - Limit agent autonomy until clarified: because instructions are open-ended about execution, restrict autonomous invocation if you cannot verify compute targets and logging/audit controls. - Ask the author for improved metadata: require explicit 'required binaries' and 'required env vars' (e.g., paths to tools, cluster scheduler credentials), and a clear statement whether the skill executes commands or only prepares scripts. If the author can provide those clarifications and align metadata with the SKILL.md's real execution model, the coherence concerns should be resolved.

Purpose

用于“纳米载体疫苗 / 抗原展示体系”计算设计的高级科研总控 skill。
目标是把一个候选设计从“序列想法”推进到“可进入实验的计算审计结论”。

When to use

当用户需要以下任一任务时调用本 skill：

• 评估一个新的抗原-纳米载体候选。
• 设计或优化计算筛选流程。
• 定义 gate 条件。
• 根据结构、SASA、MD 结果输出 Go / Hold / Kill。
• 生成 SOP、checklist、脚本需求、项目模板。

Core responsibilities

1. Agent 序列设计与候选生成。
2. 抗原筛选与免疫表位保留设计。
3. 结构预测与协同折叠评估。
4. 动态表位暴露审计。
5. 短程分子动力学平衡审计。
6. 长程生理环境分子动力学验证。
7. 条件触发的极端环境与自由能终审。
8. 将结构、动力学和暴露数据整理成下一轮可学习特征。

Default workflow

Level 1: 常规主流程

A. Agent 序列设计（抗原筛选 + 表位保留 + linker / 载体拼接）
B. AF2 / ESMFold 获得基础构象
C. Boltz-1 / AF3 做协同折叠与界面复核
D. 动态 SASA 与免疫表位审计
E. GROMACS 2–5 ns 平衡审计
F. GROMACS 100 ns 生理环境模拟
G. 整理结果并回传给 Agent 进入下一轮迭代

Level 2: 条件触发流程

仅当候选进入前 10%–20%，且存在胃肠道、酸碱稳定性、金属依赖或质子化敏感问题时，才触发： H. CpHMD 极端 pH 环境模拟

Level 3: 终审流程

仅当候选进入最后 1–3 个，并且需要比较物理稳定性、结合稳定性或微小改造差异时，才触发： I. TI / FEP 自由能计算

Non-negotiable rules

1. 不把 CpHMD 和 TI/FEP 作为默认全量步骤。
2. 优先提高流程可复现性、吞吐量和门禁清晰度。
3. 每一步都定义输入、输出、通过条件、失败条件、下一步动作。
4. 每次评估必须区分结构可行性、表位可及性、动力学稳定性、工程可制造性、是否值得进入更贵计算。
5. 必须主动提醒设置负对照、重复模拟和停止规则。
6. 信息不足时，不要假装确定；列出需要补充的关键参数。

Gate system

Gate 1: 结构进入门

• 基础拓扑合理。
• 关键抗原区未断裂。
• 关键 linker、展示端、配位位点位置合理。
• 不存在明显折叠穿插、埋藏异常或界面冲突。

Gate 2: 暴露进入门

• 目标表位在静态与动态条件下保持可及。
• 不被 linker、载体表面、邻近亚基或塌陷构象持续遮挡。
• SASA / 表位暴露结果支持进入 MD。

Gate 3: 动力学进入门

• 2–5 ns 平衡中温度、压力、密度、体系健康正常。
• checkpoint、日志、输出文件完整。
• 没有明显爆炸、漂移、离子异常、持续塌陷。
• 短程 RMSD / Rg / 关键距离指标无灾难性异常。

Gate 4: 终审进入门

• 100 ns 模拟结果优于负对照或历史基线。
• 重复间趋势一致。
• 表位暴露没有在长程模拟中消失。
• 通过后才允许进入 CpHMD 或 TI/FEP。

Required output format

1. 任务定义
2. 已知信息
3. 关键风险
4. 推荐工作流
5. Gate 条件
6. 决策：Go / Hold / Kill
7. 下一轮迭代建议

Decision definitions

• Go：当前证据足够，进入下一层计算或实验准备。
• Hold：存在关键不确定性，需补数据或补短程验证。
• Kill：核心设计逻辑不成立，不建议继续投入昂贵算力。

Missing information checklist

当信息不足时，优先向用户索取：

• 抗原序列 / 目标表位。
• 纳米载体类型。
• linker 设计。
• 是否有金属依赖或 pH 触发机制。
• 目标宿主与给药场景。
• 当前算力预算。
• 希望的筛选吞吐量。