Worked solutionOne step at a time
Align the letters
A global alignment (Biopython PairwiseAligner, BLOSUM62, gap open −11, extend −1) pairs 142 residues of myoglobin (153 long) with leghemoglobin (143 long). Only 29 pairs are the same amino acid.
Ask what random gives
Shuffle leghemoglobin's letters (same composition, no order) and align again, 50 times. The aligner still finds about as many matches, because with gaps allowed it can always line up some letters by chance. In 32 of the 50 shuffles the junk scored at least as high as the real protein.
Compare the shapes, not the letters
Superpose the Cα backbones with the Kabsch algorithm. On the 142 letter-aligned pairs the RMSD is 5.1 Å, because the letter alignment slips in the loops. A structure-based alignment (refit, keep pairs closer than 3.5 Å, repeat) finds 114 residues that sit on top of each other.
Answer: B, the pattern of oily and water-loving positions
A globin is eight helices packed around a greasy core. What the fold needs is that the positions facing the core stay hydrophobic and the positions facing water stay polar. Leucine, isoleucine, valine, phenylalanine or methionine can each fill a buried slot, so the letters drift apart over evolution while the pattern, and with it the fold, stays.
What the computation showedThe pattern survives where the letters do not
Everything in the reel is computed from the two crystal structures (1MBN, 1BIN) in Python with Biopython and numpy.
- Letters: 29 of 143 identical (20.3%). Against 50 shuffles of leghemoglobin, 20.6% ± 2.7%. By letters alone the pair is indistinguishable from noise.
- Shape: 114 residues superpose within 3.5 Å, with a core RMSD of 1.7 Å. Helices A, B, C, E, F, G and H pair with the same-named helix in the other protein; only the short D helix does not.
- The hidden pattern: call a residue buried if its Cα has more Cα neighbours within 10 Å than the protein's median, and hydrophobic if it is A, V, I, L, M, F, W or C. At the 47 structurally aligned positions that are buried in both proteins, the identical letter appears only 12.8% of the time, but the hydrophobic-or-not label agrees 66.0% of the time. Shuffling leghemoglobin's letters over the same positions gives 48.4% ± 6.9% agreement, so the real pattern sits about 2.6 standard deviations above chance while identity sits at chance.
Wrong answersWhy each one is tempting
| Pick | What went wrong |
|---|---|
| A · exact letters | The textbook "sequence determines structure" read too literally. Sequence does determine structure, but many sequences map to one fold. |
| C · chain length | Similar length is a symptom of a shared fold, not its cause; unrelated folds come in every length. |
| D · the heme | The heme is the shared job, and the pocket is conserved, but the helices form without it. It does not explain how different letters build the same pocket. |
Follow-ups an interviewer asks nextTry these before opening them
BLAST misses this pair. How would you find it?
Use a profile, not a single sequence: PSI-BLAST or an HMM (HMMER, HHblits) built from many globins captures which positions must stay hydrophobic. Better still, compare structures directly (Foldseek, DALI, TM-align) or predicted structures from AlphaFold.
Why is 20–35% identity called the twilight zone?
Above about 35% over a long alignment, shared ancestry is almost certain. Below about 20% you are in the noise, as the shuffles here show. In between, true relatives and chance matches overlap, so identity alone cannot tell you which you have.
Why does the RMSD drop from 5.1 Å to 1.7 Å?
The 5.1 Å uses the letter alignment, which pairs residues that are a turn or two out of register in the loops and helix ends. The structure-based alignment pairs residues by where they sit in space and leaves out the parts that genuinely differ. Always say which pairs an RMSD is computed on.
How would you design a new sequence for this fold?
Keep the binary pattern: put hydrophobic residues where the helices face the core and polar ones outside, with helix-friendly residues along each helix. Hecht's binary-pattern experiments built folded four-helix bundles this way from libraries of otherwise random sequences. Today you would run ProteinMPNN on the backbone and check the result with AlphaFold.
