Questions: Protein Structure Prediction Basics

Question 1

What is the key input that enables AlphaFold2 to predict protein structures with high accuracy?

• A. The protein's mRNA sequence and ribosome binding sites
• B. A multiple sequence alignment of evolutionary relatives of the target protein
• C. The protein's experimentally measured circular dichroism spectrum
• D. The crystal structure of a closely related protein

Answer: (B)

AlphaFold2's architecture relies heavily on a multiple sequence alignment (MSA) of homologous sequences. The patterns of amino acid co-variation in the MSA encode information about which residues are in spatial proximity — if two positions co-vary across evolution, they likely interact in the folded structure. The Evoformer module processes this co-evolutionary information alongside pairwise residue features to build a representation that the structure module converts into 3D coordinates. When few homologs are available, prediction accuracy typically decreases.

Question 2

Homology modeling can only produce a reliable structure if the target protein shares at least 90% sequence identity with a template.

• T. True
• F. False

Answer: False

Homology modeling can produce useful structures at much lower sequence identity, though accuracy decreases with divergence. Above 50% identity, models are typically reliable for backbone placement and many functional inferences. Between 30-50%, models are useful for overall fold and some functional predictions but less reliable for side-chain positioning. Below 30% (the 'twilight zone'), homology detection itself becomes unreliable, and threading or ab initio methods may be more appropriate. The 90% threshold is far too conservative.

Question 3

Explain why a region of a protein with a low AlphaFold pLDDT score should be interpreted differently from a region with a high score.

Model answer: The pLDDT (predicted local distance difference test) score ranges from 0-100 and reflects AlphaFold's confidence in its prediction for each residue. High scores (>90) indicate that the predicted position is likely very close to the true structure and can be used for detailed structural analysis. Scores between 70-90 suggest the backbone is probably correct but side-chain positions may be approximate. Low scores (<50) often correspond to intrinsically disordered regions, flexible loops, or regions where the model lacks sufficient evolutionary information — these predicted coordinates should not be treated as reliable structural features.

Low pLDDT regions are not necessarily prediction failures — they may accurately reflect genuine structural disorder. The predicted aligned error (PAE) provides complementary information about the relative positions of domains, helping distinguish disordered regions from cases where domains are well-predicted individually but their relative orientation is uncertain.