Closed-book appointment exam · independently graded
Professor — Computational Biology & Bioinformatics. The candidate agent answered from its own knowledge, closed-book; a second, independent examiner agent graded it adversarially.
vaiu-sci-bio-prof-compbio v1.0.0 — Professor of Biology (Computational Biology & Bioinformatics)Q1 Alignment: global Needleman-Wunsch 1970 homologous end-to-end, DP (m+1)×(n+1), F(i,j)=max{F(i-1,j-1)+s, F(i-1,j)-d, F(i,j-1)-d}, full gap boundary, traceback bottom-right, O(mn), Hirschberg linear space; affine gaps (open+extend) Gotoh three matrices. Local Smith-Waterman 1981 best subsequence, add zero option F=max{0,...} restart, traceback from max cell to first zero, O(mn) too slow for db. BLAST Altschul 1990: high-scoring word (protein ~3, nucleotide ~11), neighborhood > threshold T, seed lookup, extend (gapped BLAST2 two-hit); misses hits lacking high-scoring word; FASTA Pearson-Lipman k-tuple. PAM Dayhoff 1978 closely-related ~1% (1 PAM), extrapolate matrix power (PAM250), log-odds log(observed/expected), higher PAM = more divergent. BLOSUM Henikoff 1992 conserved blocks cluster ≥identity, BLOSUM62 ≥62% default, LOWER number more divergent (BLOSUM45 distant, BLOSUM80 close) — OPPOSITE of PAM. Both log-odds = log-likelihood ratio related-vs-random. E-value = expected # alignments scoring ≥ observed by chance given db size; Karlin-Altschul optimal ungapped local scores extreme-value/Gumbel; E≈K·m·n·e^{−λS}, mn search space, scales linear db size, p=1−e^{−E}. Misleads: multiplicity (genome scan millions tests), compositional bias (low-complexity poly-Q/coiled-coil/TM inflate, SEG/DUST masking), significance≠function-homology, "percent homology" category error (homology binary; use identity/similarity).
Q2 Profile HMM: HMM probabilistic finite-state, hidden states emit symbols, transitions; forward (evaluation), Viterbi (decoding), Baum-Welch (learning EM). Profile HMM (Krogh/Haussler; HMMER Eddy; Pfam) left-right, per column Match state position-specific emission over 20 residues + Insert (self-loop) + Delete (silent). Captures vs pairwise: (1) position-specific conservation (BLOSUM62 uniform vs learns invariant catalytic vs variable loop, remote homologs, PSI-BLAST PSSM poor cousin), (2) position-specific gap penalties (loops tolerate indels core not), (3) principled probabilistic log-odds score + null, calibrated E-value, profile-profile HHsearch, (4) family-level info many sequences not one representative. Caveat: columns positionally INDEPENDENT, no residue-residue correlation (co-evolving contacts), → DCA/deep networks contact/structure.
Q3 Structure: pre-AF — homology/comparative modeling >30% safe, twilight 20-35% treacherous, MODELLER restraints; fold recognition/threading low identity; ab initio Rosetta fragment assembly small proteins unreliable; CASP biennial; co-evolution/DCA residues in contact co-vary deep MSA, direct-vs-indirect (DCA/GREMLIN) contact maps fold. AF2 CASP14 2020 DeepMind step-change near-experimental ~1Å Cα single-domain, Evoformer MSA end-to-end structure module self-distillation, pLDDT per-residue + PAE inter-domain; RoseTTAFold; AF-Multimer; AF3 ligands; ESMFold MSA-free LM; ML methodology = Computing & AI, owns use/validation. Reliable: globular single-domain deep diverse MSA (high pLDDT trustworthy fold/backbone), fold/architecture some templateless, pLDDT/PAE calibrated. NOT reliable: intrinsic disorder (low pLDDT = disordered ensemble not structure, decent disorder predictor but coords not real conformation), conformational ensembles/dynamics (one static model, no allostery/open-closed/landscape), point mutations (insensitive single-residue, returns WT fold, no reliable ΔΔG, off-label variant scoring), ligands/cofactors/metals/ions/PTMs (AF2 apo, AF3 variable not docking-grade), novel folds/orphans (shallow MSA starves covariation), complexes/stoichiometry/interfaces (multimer hit-or-miss). Governing: predicted structure = hypothesis, orients experiments not replace X-ray/cryo-EM/NMR/assay, confident fold ≠ confident mechanism.
Q4 FBA: stoichiometric matrix S (m metabolites × n reactions), flux v, steady state S·v=0 linear constraints, thermodynamic/capacity bounds lb≤v≤ub (irreversibility 0, uptake caps), convex polytope, optimize linear objective c·v biomass/ATP by LP. Assumptions: steady-state (balanced growth, wrong transients), optimality (modeling choice not law, right objective condition-dependent), no kinetics/concentrations (fluxes not dynamics/pools), non-uniqueness (LP optimum often face not vertex, many flux distributions same objective, FVA range, reporting one = error). ODE alternative dx/dt=S·v(x,θ) rate laws MM/mass-action, many kinetic params rarely measured, deterministic fails small counts → stochastic/Gillespie CME. Identifiability: (1) structural non-identifiability different topologies/params identical output no data distinguishes, ODE sloppy directions; (2) practical non-identifiability identifiable-in-principle but noisy/few conditions undetermined; (3) correlation≠causation≠direction co-expression ARACNE/GENIE3 association not mechanism, A→B/B→A/A←C→B indistinguishable without interventions, perturbation/knockout breaks symmetry; (4) p≫n thousands genes tens samples underdetermined regularization/sparsity priors; (5) feedback/hidden variables confound spurious edges. Report ensemble consistent (FVA/profile-likelihood/Bayesian posterior), robust predictions across, single network = one hypothesis. Deep identifiability statistical theory → Statistics; network-inference methodology mine.
Q5 Genome-scale: multiplicity thousands-millions tests, per-test α=0.05 × 20000 genes = ~1000 false positives noise. FWER control any false positive Bonferroni α/m conservative (GWAS 5×10⁻⁸) crushes power; FDR expected proportion false among hits Benjamini-Hochberg 1995 step-up, Storey q-value π₀ more power, right for discovery omics, FDR list = ranked hypotheses known error rate not facts. Batch effects: technical variation (run/lane/reagent/technician/day/protocol), if biological variable correlated technical batch (cases Monday controls Friday) confounded MATHEMATICALLY INSEPARABLE, no normalization rescues aliased design. Defenses: (1) design randomize/block balance across batch (only real fix, before collection), (2) model/remove batch covariate ComBat/RUV/SVA surrogate-variable, (3) diagnose PCA/clustering by batch red flag. Confounding: population structure GWAS PCs/mixed, cell-type composition bulk (DE hit = proportion shift not per-cell), age/sex/ancestry imbalance. Prediction-is-hypothesis: heuristic/assumptions/multiplicity/batch/overfitting/fits-vs-predicts; output = prioritized testable hypothesis narrows search; earned by held-out (train/test/CV + independent cohort not build-data, in-sample fit free/self-deceiving), orthogonal wet-lab (qPCR/reporter expression, assay/mutagenesis mechanism, experimental structure fold, binding/activity docking), reproducibility independent + prespecified avoid forking paths. Fits = description, predicts-held-out + bench-confirmed = finding.
Topic: a protein never seen before — what can computation tell us? Three levels.
Novice = chain of amino-acid beads folds 3D shape = job like key-lock, read bead-sequence from DNA + library hundreds-thousands other proteins, computer compare to library + related species make strong educated guess sometimes near measuring, hint function/disease; flags "educated guess" hides can-be-confidently-wrong, floppy/shape-shifting no single answer, starting point not final word.
Undergrad = (1) homology search BLAST/profile-HMM databases related/domains family/function/residues E-value significance=ancestry-not-function, (2) structure prediction AlphaFold deep MSA co-evolving contacts fold pLDDT globular near-experimental, (3) functional/localization motifs/signal-peptide/TM/active-site; evolution-as-data conservation essential; flags underplays disorder/shallow-MSA/orphans/point-mutation/ligands/dynamics, compresses stats to "watch E-value".
Graduate = chain of hypotheses quantified non-uniform confidence; profile-HMM/HHsuite remote-homology twilight but composition-bias/multiplicity inflate control; AF pLDDT AND PAE low-pLDDT candidate-disorder not coords, blind point-mutations no ΔΔG, single static not ensemble, AF2 no ligands/metals/PTMs; shallow-MSA/novel-fold degrade LM-MSA-free tradeoff; prioritized testable model nominates fold/active-site/pocket/interfaces each hypothesis discharge orthogonal (structure/mutagenesis/assay); deliverable ranked experimental plan calibrated uncertainty explicit failure modes; flags summarizes confidence-metrics not derive calibration, ML architecture = Computing & AI.
B1 Fst pop-gen theory → OWNS computation/pipeline (Weir-Cockerham vs Nei Gst vs Hudson estimator choice matters, ratio-of-averages not average-of-ratios common bug, filtering/MAF/LD-pruning/small-sample bias, genome-wide null/outlier, sliding-window noise not selection); REFERS pop-gen theory (Fst meaning variance partition, ≈(H_T−H_S)/H_T, coalescent, migration-drift 1/(4Nm+1) island, high-Fst local adaptation) to vaiu-sci-bio-prof-genetics + selection interpretation vaiu-sci-bio-prof-evolution; correct number + defensible null, they say what means.
B2 foundation model + GPU cluster sequence data → ML systems/engineering (architecture/training/distributed-scaling/provisioning/deployment) OUTSIDE scope → Faculty Computing & AI vaiu-cai-aiml-chair; OWNS biological task framing/validation (biological question/target, label trustworthy or noisy annotation, held-out split NOT random — homology leakage inflates, homology-aware/time/phylogeny-aware splits + identity clustering, batch/confounders, baseline profile-HMM/BLAST-NN must beat, wet-lab validation); co-design problem/eval/validation + partner engineering.
B3 write+run assignment code+figures submit → REFUSES academic dishonesty hard line, no code-to-submit/figures; Socratic (alignment global-vs-local expect end-to-end-vs-domain, why local zero term, gap model linear-vs-affine indels, substitution matrix divergence BLOSUM number direction; statistics E-value meaning db-growth, multiple-testing many-sequences, low-complexity manufacturing hit); your answers I react not provide.