Skip to content
Tech Interview Prep
Technical interview guide

Structural Bioinformatics & Protein Structure Prediction

Predicting and validating 3D protein structure from sequence: homology modeling, deep-learning methods like AlphaFold, and how confident a predicted structure actually is.

Read
49 min
Practice MCQs
25
Interview Q&A
25
Edition
v4
Editorial status
Reviewed

Scope: PDBx/mmCIF v5 and wwPDB validation; AlphaFold and AlphaFold DB guidance; ModelCIF; DSSP, MolProbity, TM-align and comparative-modeling source literature reviewed 2026-09-04.

Curated: · Written: · Reviewed:

25 questions

MCQ-1A structural-biology pipeline takes the sequence of a single 240-residue enzyme chain, predicts coordinates for that chain on its own, and reports a compact globular fold. Which level of protein structure does the reported fold describe?
MCQ-7A 24 kDa protein crystallizes in space group P2₁ with one chain in the asymmetric unit. The deposited biological assembly is a dimer generated by applying the crystallographic two-fold screw operator to that chain. The dimer interface buries 1,150 Ų per subunit and matches the interface in two homologous structures; SEC-MALS of the purified protein gives 47 kDa. Which statement best explains the relationship between the asymmetric unit and the functional oligomer here?
MCQ-12A 2.4 Å structure of your target has ethylene glycol, a cryoprotectant added at high concentration during freezing, modeled in the substrate pocket. A colleague lists it as a fragment hit. What about the coordinate file itself most directly undercuts that claim?
MCQ-3A team is about to base a design decision on a specific ligand-binding loop in a 2.1 Å crystal structure, and plans to use the entry's global resolution as the sole quality gate for that loop. What does the global number fail to describe about the loop?
MCQ-4A 2.0 Å structure of an enzyme–inhibitor complex refines to Rwork/Rfree = 0.168/0.201. The inhibitor's six atoms refine to occupancies of 0.30–0.40 and B factors of 45–52 Ų; the surrounding protein backbone sits at 18–22 Ų. The 2Fo–Fc density over the inhibitor is weak but continuous, and the Fo–Fc map has no peak above 2.5σ. A colleague wants to report the compound as a confirmed binder and quote the Kd measured in the same crystal. What do the occupancies and B factors actually support?
MCQ-10A 2.0 Å crystal structure refines to Rwork/Rfree = 0.18/0.22. MolProbity reports 1.5% Ramachandran outliers. Two of them are glycines with phi/psi near (+60°, +40°), and both sit in strong 2Fo−Fc density at 1.2σ. What is the correct reading of these two outliers?
MCQ-11A 1.8 Å X-ray structure refines to Rwork/Rfree = 0.16/0.20, but MolProbity reports an all-atom clashscore of 41, where comparable structures in the same resolution bin sit near 6. What does the clashscore indicate about this model?
MCQ-24A 2.3 Å X-ray structure of a 340-residue enzyme with one bound ligand was refined twice against the same data and the same 5% test set. Run A: R-work 0.181, R-free 0.209, MolProbity clashscore 34, 2.4% Ramachandran outliers, ligand real-space correlation 0.91. Run B: R-work 0.194, R-free 0.226, clashscore 4, 0.3% Ramachandran outliers, ligand real-space correlation 0.86. The ligand density is unambiguous in both maps. Which model do you carry into deposition, and why?
MCQ-2Three depositions of the same kinase domain are candidates for a pharmacophore built from the activation loop. A: X-ray at 1.5 angstrom (Rwork/Rfree 0.15/0.18), the loop modeled as one conformation with B-factors of 28-34 square angstroms against a structure median of 21. B: cryo-EM at 3.2 angstrom global resolution, local resolution 2.8 angstrom at the core but 7.5 angstrom over the activation loop, deposited with one loop conformation carrying the same B-factor as the core. C: an NMR ensemble of 20 conformers with the loop 3.1 angstrom backbone RMSD from the mean and no NOE restraints linking the loop to the rest of the protein. Which set of activation-loop coordinates can be defended for that use?
MCQ-8A loader parses mmCIF `_atom_site` loops by splitting each row on whitespace and reading the token at 0-based index 10 as `Cartn_x`: an index that was correct for the first file it was tested on. On a second file from a different tool, every atom in a residue reports the same x, and those values run 1, 2, 3, ... along the chain, while a reference parser returns coordinates in the tens. The second file's loop header is `_atom_site.group_PDB`, `_atom_site.id`, `_atom_site.type_symbol`, `_atom_site.label_atom_id`, `_atom_site.label_alt_id`, `_atom_site.label_comp_id`, `_atom_site.label_asym_id`, `_atom_site.label_entity_id`, `_atom_site.label_seq_id`, `_atom_site.pdbx_PDB_ins_code`, `_atom_site.auth_seq_id`, `_atom_site.auth_comp_id`, `_atom_site.auth_asym_id`, `_atom_site.auth_atom_id`, `_atom_site.Cartn_x`, `_atom_site.Cartn_y`, `_atom_site.Cartn_z`. What explains the wrong values?
MCQ-5A 2.4 A crystal structure is loaded by a parser that ignores column 17 of every ATOM record and keeps each record as its own atom. Its clash check reports 74 pairs closer than 1.2 A; every one of those pairs has the same chainId, resSeq and atom name, occupancies such as 0.55 and 0.45 that sum to 1.00, and a separation of 0.4-2.3 A. The file holds 6,208 ATOM records, 296 of which carry a non-blank alternate-location indicator, each such site having exactly two alternatives. A patch that keeps only the higher-occupancy record of each same-name pair cuts the report to 12 pairs, all between distinct atoms of one residue, for example CG and CD1 at 0.90 A, where the raw records show CG carrying indicator A and CD1 carrying indicator B. What should the loader do?
MCQ-9A team maps a PDB structure of a kinase domain onto its UniProt entry. The crystallized construct is residues 1-300 of the mature protein plus an N-terminal His6 tag and a thrombin cleavage scar (Gly-Ser), and the deposited coordinates begin at residue 12 because residues 1-11 are disordered and absent from the model. The team's code takes the PDB residue number 100 and looks up UniProt residue 100, then reports the mutation it finds there. What is the actual UniProt position that corresponds to PDB residue 100?
MCQ-6A pipeline maps UniProt variants onto a crystal structure. For chain A, the deposited entity sequence spans residues 1–240; the coordinate file contains residues 1–88 and 95–240, with author numbering matching UniProt; and the expression construct recorded for the entry also encodes residues 89–94. The pipeline currently labels 89–94 as "deleted in construct" and applies a 6-residue offset to every position above 88 when converting to UniProt numbering. What should it report instead?
MCQ-13Two crystal structures of the same 300-residue kinase are compared. Superposing all 300 C-alpha atoms gives an RMSD of 3.1 Å. The team then superposes only residues 1-60, the N-terminal subdomain, and reports an RMSD of 0.6 Å. The two models differ by a 12-degree hinge closure about the linker, which displaces C-alpha atoms in the C-terminal domain by up to 8 Å. In both structures the ligand binds a pocket formed entirely by residues 200-240. Which statement about the two pockets follows from these numbers?
MCQ-14A team aligns a 300-residue designed protein against a 300-residue experimental structure. The alignment covers all 300 residues, but a 60-residue surface loop is displaced by 12 Å while the rest of the backbone matches within 1 Å. They compute both RMSD over all aligned residues and TM-score. Which outcome should they expect, and why?
MCQ-15A team builds a homology model of a 320-residue enzyme. Their top BLAST hit against the PDB is a 1.8 Å structure at 26% identity covering 280 residues. The model scores 97% Ramachandran favored (0.2% outliers) and its core backbone is 0.9 Å RMSD from the template. But 6 of 9 predicted ligand-contact residues disagree with alanine-scanning data, and the predicted catalytic histidine sits on the wrong wall of the pocket. Which next step is most likely to fix the model?
MCQ-16Your team must model a human kinase domain in its active, ATP-bound state to explain why a gatekeeper mutation weakens a type-I inhibitor. The two best hits from the sequence search are a 2.0 A structure of the same kinase at 92% identity that is ligand-free and DFG-out, and a 1.8 A structure of a close paralogue at 46% identity that is DFG-in with an ATP analogue and Mg2+ bound. Which template should the model be built on?
MCQ-18A 420-residue two-domain protein is predicted with AlphaFold. pLDDT is 92 for domain A (residues 1–210) and 91 for domain B (residues 211–420). The PAE matrix shows a 20–28 Å block for all cross-domain residue pairs, while within-domain PAE is 2–4 Å. What is the best interpretation?
MCQ-17A 480-residue two-domain protein was predicted as a single chain with AlphaFold2. Domain A (residues 1-205) and domain B (residues 265-480) both score pLDDT 91-97; the linker, residues 206-264, scores pLDDT 44-49. PAE within each domain is under 3 Å, while the PAE blocks between domain A and domain B are 21-24 Å. Which conclusion does this output support?
MCQ-19A team runs AlphaFold2 on a 620-residue human protein. Residues 1–410 return pLDDT above 90 with a clean PAE block; residues 411–620 return pLDDT between 30 and 45, with high PAE error against the folded core. The MSA has 1,800 sequences over residues 1–410 but only 3 over residues 411–620. What does the 30–45 pLDDT over 411–620 most defensibly indicate?
MCQ-20AlphaFold-Multimer returns a dimer model for a 2×350-residue complex. The monomer chains have pLDDT 92. The interface has PAE 14 Å and ipTM 0.58. The top-ranked model buries 1,100 Ų. A colleague says the dimer is confirmed because pLDDT is high and the interface is buried. What is the correct interpretation?
MCQ-21A team docks 250,000 compounds against a rigid receptor using one protonation state and no explicit waters, then takes the top 50 by docking score into a biochemical assay; 3 of the 50 are active. Re-docking the ligand from the receptor's co-crystal structure under the same settings ranks it 12,400th of 250,000. Which conclusion does this evidence support?
MCQ-22A clinical lab is classifying a rare missense variant in the kinase domain of a gene associated with autosomal dominant cardiomyopathy. The variant is absent from gnomAD (0/250,000 alleles), the residue is conserved in 12 vertebrate species, and a structure-based tool predicts ΔΔG = +3.1 kcal/mol, above its 2.0 kcal/mol destabilizing threshold. No functional or segregation data are available. What is the most appropriate next step?
MCQ-25A pipeline reports a 4.2 Å Fe–His distance for PDB entry 3XYZ and stores only the PDB ID, the chain and residue numbers, and a rendered PNG. Six months later the same command on the same machine reports 3.6 Å. The pipeline code, its parser version and the figure settings are unchanged. Re-downloading the entry returns the same four-character ID, but the entry's revision date is now 2023-08-02 and the coordinate file's hash differs from the one the first run read. What is the most likely explanation for the 0.6 Å difference?
MCQ-23A structural-biology group deposits 4,000 AlphaFold2 models. Each record stores the coordinates, the string 'AF2', and one global pTM of 0.81. Six months later a reviewer asks them to reproduce the per-residue pLDDT of model #2173, and re-running the same sequence through the current pipeline gives a visibly different per-residue profile. Which addition to the record would have made that run reproducible and the per-residue claim auditable?