技能 esm
🧬

esm

安全 🌐 網路存取📁 檔案系統存取

Generate and design proteins with ESM

也可從以下取得: K-Dense-AI

Protein engineering requires specialized AI models to analyze sequence-structure-function relationships. ESM provides state-of-the-art protein language models for generating novel sequences, predicting structures, and designing functional proteins through multimodal AI.

支援: Claude Code(CC)
📊 71 充足
1

下載技能 ZIP

2

在 Claude 中上傳

前往 設定 → 功能 → 技能 → 上傳技能

3

開啟並開始使用

測試它

正在使用「esm」。 Design a novel enzyme with catalytic activity using ESM3

預期結果:

  • Generated sequence with 238 amino acids
  • Predicted structure shows typical enzyme fold
  • Chromophore region designed with catalytic residues
  • Saved to novel_enzyme.pdb
  • Confidence score: 0.84

正在使用「esm」。 Find similar proteins to this sequence using embeddings

預期結果:

  • Generated 512-dim embedding for query sequence
  • Compared against 100 protein database
  • Found 3 high-similarity matches (>0.85 cosine similarity)
  • Top match: membrane_protein_variant_A

正在使用「esm」。 Inverse fold this PDB structure

預期結果:

  • Loaded structure with 342 residues
  • Generated novel sequence matching the fold
  • Sequence identity: 42%
  • Output saved to designed_sequence.fasta

安全審計

安全
v5 • 1/17/2026

This is a legitimate scientific documentation skill for ESM protein language models. All findings are false positives: backticks are markdown formatting, eval() is PyTorch's model.eval() method, and the network+credentials+code pattern is expected for an API client. No malicious code, no credential harvesting, no suspicious system operations.

6
已掃描檔案
2,921
分析行數
2
發現項
5
審計總數
審計者: claude 查看審計歷史 →

品質評分

45
架構
100
可維護性
87
內容
20
社群
100
安全
91
規範符合性

你能建構什麼

Design novel proteins

Generate new protein sequences with desired properties using ESM3 multimodal generation across sequence, structure, and function tracks.

Screen protein variants

Create and analyze libraries of protein variants using embeddings for clustering and similarity analysis.

Predict structures

Predict 3D structures from sequences or design sequences for known structures using inverse folding.

試試這些提示

Generate sequence
Generate a 200-amino acid protein sequence for a fluorescent protein using ESM3. Start with a masked sequence and use function conditioning for green fluorescent protein properties.
Predict structure
Predict the 3D structure for this protein sequence using ESM3 structure track. Output the coordinates and save as PDB format.
Inverse fold
Load this PDB structure and use inverse folding to generate a sequence that folds to the target structure. Use temperature 0.7 for diversity.
Embed sequences
Generate ESM C embeddings for these protein sequences and compute pairwise similarity using cosine similarity. Cluster similar sequences together.

最佳實務

  • Start with smaller models (esm3-sm-open-v1) for prototyping before production
  • Use temperature scheduling (high to low) for better generation quality
  • Validate generated sequences with structure prediction before experimental work
  • Cache embeddings for repeated analyses to reduce compute costs

避免

  • Using generated proteins directly in clinical applications without validation
  • Ignoring biosafety considerations when designing novel proteins
  • Processing extremely long sequences without chunking strategies
  • Skipping error handling for GPU memory errors and API rate limits

常見問題

What models are available locally vs via API?
esm3-sm-open-v1 (1.4B) and ESM C models run locally. Larger ESM3 models (7B, 98B) require Forge API access.
What are typical sequence length limits?
ESM C supports up to 1024 residues efficiently. ESM3 handles similar lengths with chunking for longer sequences.
Can this integrate with my existing ML pipeline?
Yes, ESM C embeddings work as features for classifiers, clustering, and custom neural networks.
Is my protein data sent to external servers?
Only when using Forge API. Local models process data on your infrastructure.
Why is my GPU running out of memory?
Use half precision (.half()), process in batches, clear CUDA cache, or use smaller model sizes.
How does this compare to AlphaFold?
ESM3 combines sequence generation with structure prediction in one model. AlphaFold is primarily structure prediction from sequence.