Questions: AlphaFold and ML Prediction

Question 1

AlphaFold2 produces a confidence score (pLDDT) for each residue. A region with pLDDT < 50 should be interpreted as:

• A. A very accurate prediction that can be trusted for atomic-level analysis
• B. A region where the prediction is unreliable — likely intrinsically disordered, flexible, or lacking evolutionary information from the multiple sequence alignment. The 3D coordinates in this region should not be interpreted as a meaningful structure
• C. A region that must be confirmed by X-ray crystallography before any interpretation
• D. An artifact of the neural network that has no biological meaning

Answer: (B)

pLDDT (predicted Local Distance Difference Test) ranges from 0-100, with >90 indicating high confidence (comparable to experimental structures), 70-90 indicating good but less reliable prediction, and <50 indicating low confidence. Low-pLDDT regions typically correspond to intrinsically disordered regions (which genuinely lack a single stable structure), flexible loops, or regions where the MSA provides insufficient evolutionary information. Importantly, low pLDDT is often biologically meaningful — it often correctly identifies disordered regions. But the specific 3D coordinates in these regions should not be trusted, as the model is essentially guessing.

Question 2

AlphaFold has solved the protein structure prediction problem, making experimental structural biology unnecessary.

• T. True
• F. False

Answer: False

AlphaFold has dramatically advanced structure prediction, but experimental structural biology remains essential for several reasons: (1) AlphaFold predicts a single static structure, while many proteins function through conformational changes — experimental methods (cryo-EM classification, NMR dynamics) capture multiple states. (2) AlphaFold does not reliably predict ligand binding poses, protein-protein interaction geometries, or the effects of post-translational modifications. (3) AlphaFold cannot predict structures of proteins without homologs in evolutionary databases (orphan proteins). (4) Novel folds, engineered proteins, and de novo designed proteins lack the evolutionary information AlphaFold relies on. (5) Experimental validation is required for drug design, mechanistic studies, and any high-stakes structural interpretation. AlphaFold generates hypotheses; experiments validate them.

Question 3

Why does AlphaFold rely on multiple sequence alignments (MSAs), and what happens when the MSA is shallow (few homologs)?

Model answer: MSAs encode evolutionary information: which residues co-evolve (mutate in a correlated manner), reflecting spatial proximity in the 3D structure. These co-evolutionary signals are the primary source of information that AlphaFold uses to infer residue-residue contacts and ultimately 3D structure. When the MSA is shallow (the target protein has few homologs in sequence databases, as for orphan proteins or recently evolved sequences), co-evolutionary signals are weak or absent, and AlphaFold's accuracy drops significantly. Single-sequence methods (like ESMFold, which uses protein language model embeddings instead of MSAs) partially address this but are generally less accurate than MSA-based methods. The dependence on evolutionary information means AlphaFold is least reliable where it is most needed — for structurally novel proteins.

AlphaFold3 has expanded to predict protein-nucleic acid complexes, protein-ligand interactions, and post-translational modifications, addressing some limitations of AlphaFold2. However, the fundamental dependence on evolutionary information remains, and accuracy for novel targets and interactions continues to be lower than for well-represented protein families.