引導式測驗 · 中等 等級

如 ESM 等蛋白質語言模型測驗

了解 ESM 序列表示、屏蔽令牌學習、嵌入、變體和結構預測,以及領域和評估限制。

相關引導路徑蛋白質語言模型 Esm
問題 1 的 8

What do protein language models learn from sequence databases?

預覽每個問題8 問題

當您參加測驗時,就會出現答案和解釋。

  1. What do protein language models learn from sequence databases?
    • Direct measurements of every protein property in an assay
    • A fixed alignment of each protein to one known structure
    • Statistical patterns in amino-acid sequences
    • A supervised label for each possible protein function
  2. What does a masked language model learn to predict?
    • A residue contact map from an experimental structure
    • Hidden amino-acid tokens inferred from sequence context
    • An experimentally measured activity value for the full protein
    • A protein family label supplied for every sequence
  3. How are protein embeddings commonly used after pretraining?
    • A substitute for an experimental activity measurement
    • A proof that two proteins have identical functions
    • A sequence-independent structural template
    • A learned sequence representation for downstream analysis
  4. Why can random sequence splits overstate downstream generalization?
    • Homologous proteins may appear in both training and test sets
    • Random splits automatically remove close homologs from both sets
    • They guarantee all test proteins come from unseen families
    • They convert predicted labels into experimental measurements
  5. What does an ESMFold output represent?
    • An experimentally resolved structure for the exact sample
    • A measurement of folding inside a living cell
    • A predicted structure from a protein sequence
    • A structure that includes every ligand and molecular partner by default
  6. Why is a variant likelihood score not a direct measurement of organism fitness?
    • It is calibrated to growth in the target organism without assay data
    • It reflects model sequence patterns, not the full experimental environment
    • It measures the effect of a variant under every possible condition
    • It establishes that the variant has no functional effect
  7. Which details should accompany downstream use of embeddings?
    • The protein identifier and sequence length alone
    • The embedding file format and storage size alone
    • The model training duration without preprocessing details
    • Checkpoint, sequence preprocessing, pooling and evaluation split
  8. How should a protein language-model prediction be used for a biological claim?
    • As a hypothesis to pair with domain review and validation
    • As an experimentally confirmed function label for any sequence
    • As evidence that family representation cannot affect performance
    • As a direct measurement of binding affinity