Skip to content
Tech Interview Prep
Technical interview guide

Machine Learning in Computational Biology

Where ML is actually applied in genomics and drug discovery, and the specific data challenges (class imbalance, batch effects, limited labels) that make biological ML harder than typical tabular or image tasks.

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

Scope: scikit-learn 1.9 cross-validation, pipeline, evaluation and calibration guidance; DeepVariant, AlphaFold and Enformer source work; TRIPOD+AI, PROBAST+AI, Model Cards and Datasheets guidance reviewed 2026-09-04.

Curated: · Written: · Reviewed:

25 questions

MCQ-11A payments team scores a fraud model on 10,000 transactions, 1% of which are fraudulent. At the deployed threshold the model's ROC point is TPR 0.90 and FPR 0.01. Review capacity is justified by precision, so the team wants that number for this same threshold: what share of flagged transactions are actually fraud?
MCQ-10A payments team scores 50,000 accounts with a fraud model, and 1 percent of those accounts are fraudulent. At the model's current threshold it catches 90 percent of the fraudulent accounts and clears 95 percent of the clean ones. Every flagged account goes into a manual review queue. Of the accounts sent to review, roughly what share actually are fraudulent?
MCQ-13A hospital deploys a 30-day readmission model. On a 10,000-patient holdout set the mean predicted risk is 0.18, and 1,800 of those patients were readmitted within 30 days. The care-management team plans to auto-enroll every patient scoring 0.30 or higher. Among the 1,000 holdout patients scoring in [0.30, 0.40], 130 were readmitted. What should the team conclude before using 0.30 as an enrollment cutoff?
MCQ-14A sepsis risk model was developed and calibrated at Hospital A: across 2,000 admissions the mean predicted risk was 12% and 12% of patients developed sepsis. Hospital B deploys the same model unchanged. On 1,000 admissions at B the mean predicted risk is again 12%, but 216 patients develop sepsis. Discrimination is unchanged (AUC 0.79 at both sites). Which conclusion about the model at Hospital B is correct?
MCQ-12A hospital deploys a calibrated sepsis-risk model. On a 10,000-admission validation cohort, 200 admissions have sepsis. The ICU consult team can handle at most 500 alerts per 10,000 admissions. A missed sepsis case costs an estimated 50 units of harm; an unnecessary consult costs 1 unit. The table shows validation performance at five thresholds: | Threshold | TP | FP | FN | TN | | --- | --- | --- | --- | --- | | 0.05 | 190 | 1800 | 10 | 8000 | | 0.10 | 170 | 800 | 30 | 9000 | | 0.20 | 140 | 300 | 60 | 9500 | | 0.30 | 100 | 150 | 100 | 9750 | | 0.50 | 50 | 20 | 150 | 9780 | The team must pick one threshold before test. Which threshold should they select?
MCQ-24An insurer is about to ship a 30-day readmission model behind one referral rule: every patient scored at or above 0.10 is enrolled in a care-management program. Evaluation over one held-out year gives the table below (illustrative figures, not from a published study). | group | n | observed 30-day readmission | mean predicted risk | scored at or above 0.10 | AUC | | --- | --- | --- | --- | --- | --- | | primary payer Medicaid | 1,200 | 4.0% | 6.5% | 9.4% | 0.79 | | all other payers | 14,000 | 6.1% | 6.2% | 6.8% | 0.80 | | overall | 15,200 | 5.9% | 6.2% | 7.0% | 0.80 | Which change should be made before the 0.10 cutoff ships?
MCQ-16An apparel supplier places one order per week for a SKU and cannot reorder mid-week. A unit short at the end of the week loses 40 USD of margin; a leftover unit costs 10 USD in markdown and holding. The demand model publishes the week's predictive distribution: 10th percentile 700 units, 20th percentile 850, median 1,150, 80th percentile 1,600, 90th percentile 1,750, mean 1,250. Which order quantity should the buyer place?
MCQ-6A team has 50,000 labeled examples. They train on 35,000, tune hyperparameters on 10,000 validation examples, and keep 5,000 for a final test. After 40 validation runs they pick the best configuration, evaluate it on the test set, then adjust the configuration and re-evaluate on the same test set 30 more times. They report the best test accuracy, 94.2%, while the best validation accuracy was 92.8%. What does the 94.2% most likely represent?
MCQ-5A team has 10,000 labeled examples. They run nested 5-fold cross-validation: for each outer fold, an inner 4-fold split on the 8,000 training examples tunes 20 hyperparameter configurations, and the selected configuration is scored on the 2,000 outer validation examples. The five outer scores are 0.84, 0.86, 0.85, 0.83, and 0.87. A separate non-nested 5-fold run that tuned the same 20 configurations on all 10,000 examples reported a best mean validation accuracy of 0.91. Which number is the unbiased estimate of test accuracy?
MCQ-15A team has 6,000 labeled samples and a preprocessing chain of median imputation, log transform, and variance-based gene filtering. A colleague fits all three once on the full 6,000 rows rather than refitting them inside each training fold, then runs 5-fold CV and reports mean validation AUC 0.92. They argue that because none of the three steps uses the labels, the folds remain independent and the estimate is unbiased. Which statement about the 0.92 is correct?
MCQ-7A team predicts a 3%-positive outcome from 12,000 rows and 400 features. They standardize the whole matrix and run SelectKBest(k=20) on all 12,000 rows, then run 5-fold cross-validation on the reduced matrix with logistic regression: mean validation AUC 0.93. Moving both steps inside a Pipeline, so each is refit on the 9,600-row training fold, drops the mean to 0.79. Which explanation accounts for the gap?
MCQ-2A team trains a single-cell classifier on 48,000 cells from 12 donors (4,000 cells each). A random 80/20 split by cell gives AUROC 0.97. When the split is redone so every cell from a donor stays entirely in train or test, AUROC drops to 0.63. The model will be deployed on cells from donors not seen in training. Which action and reported estimate follow?
MCQ-4A variant-effect CNN is trained on 1,000 bp windows tiled along each chromosome at a 200 bp stride, about 2.4 million windows genome-wide. An 80/20 split by window gives a test AUC of 0.97, while holding out whole chromosomes gives 0.68. Under the random split, how much of a test window's input is also in the training set?
MCQ-3An enzyme-class predictor is trained on 96,000 protein sequences and evaluated on a random 10% split: 0.91 top-1 accuracy on 10,667 sequences. The same checkpoint scores 0.62 on 3,400 sequences with no training-set hit above 25% identity (MMseqs2). A check finds that 68% of the random-split test sequences have a training sequence at 50% or more identity. The two test sets have near-identical median length (342 vs 351 residues), and training accuracy is 0.99. Which conclusion is correct?
MCQ-18A team trains a protein function classifier on 12,000 sequences drawn from Pfam. Before release they evaluate on a held-out set built by randomly assigning individual sequences to train or test. Test accuracy is 94%, but on 200 sequences from a newly sequenced bacterial phylum with no close relatives in Pfam, accuracy drops to 61%. The training and test sets share many near-identical homologs. Which change to the evaluation split would most directly expose the 61% figure before release?
MCQ-19A team trains a CNN on 5-kb windows from 1,000 individuals to predict a regulatory activity assay. A random window split gives AUROC 0.94; a chromosome-held-out split gives 0.71. They must choose the estimate to report for new loci. Which conclusion is best supported?
MCQ-20A team trains a 30-day readmission model on 12,000 admissions from hospitals A, B and C (2019–2021). They report "external validation AUROC 0.83" on 2,400 admissions from hospital D (2022). The methods note that the final feature set was chosen using hospital D's 2022 data; the 0.18 threshold was set on the A/B/C holdout. What does the 0.83 estimate?
MCQ-9In a hypothetical audit, a team trains a classifier to separate tumor from normal using RNA-seq from two sequencing batches. Batch A is 80% tumor and 20% normal; batch B is 20% tumor and 80% normal. Five-fold random cross-validation gives AUC 0.94. When they hold out all of batch B and train on batch A, AUC falls to 0.61. The top features are ribosomal RNA depletion, GC content, read length and flowcell lane. What should they conclude?
MCQ-1A team integrates 12 scRNA-seq batches with Harmony, passing donor as the batch key, and the treated and control cells then overlap almost completely, erasing the disease signature. In this dataset every treated donor was processed in one set of batches and every control donor in another. Which mechanism best explains the lost signal?
MCQ-25A team trains a transcription-factor binding classifier with ChIP-seq peaks as positives and uniformly sampled intergenic windows as negatives. Cross-validated AUC is 0.97, but the top-weighted features are GC content and CpG density, and on a held-out cell line the model barely beats a GC-only baseline. Which negative set would most directly test whether the model learned motif-level binding specificity?
MCQ-8An alert for Sepsis-3 onset is trained with the ICD-10 sepsis code as its label and reports AUC 0.88 against held-out coded labels. Before buying it, the hospital audits the label itself. Across the 10,000 admissions of the validation year, 900 carry the sepsis code; chart review of a random 500 coded admissions confirms Sepsis-3 in 300 of them; review of a random 500 uncoded admissions finds 50 meeting Sepsis-3. If a model reproduced the code label perfectly, about what sensitivity against chart-confirmed sepsis could it reach?
MCQ-17A team trains a gradient-boosted model on 4,000 RNA-seq samples to predict drug response. On the held-out set, shuffling gene G1 drops AUC by 0.09, the largest of any feature. G1 and G2 are co-regulated (Pearson r = 0.97). Retraining without G1 costs 0.004 AUC; retraining without G2 costs 0.11 AUC. Every sample sequenced at site B has G1 in the top decile and also responded to the drug. Which conclusion does this evidence support?
MCQ-22A team integrates 120,000 PBMCs from four donors profiled on 10x Chromium 3' and Smart-seq2, treats donor and platform together as the batch label (eight batches), and holds per-platform cell-type labels from the unintegrated data. The integrated UMAP interleaves donors, but the plasmacytoid dendritic cells (about 1% of the cells, present in all four donors) now sit inside the monocyte cloud. The check must confirm the eight batches mixed and, at the same time, flag the individual cells that lost their identity, without first choosing a clustering resolution. Which one does that?
MCQ-23A hospital's sepsis-risk model was trained on ICU data from 2019-2021 and deployed unchanged in 2023. The service is healthy: p99 latency 90 ms, zero 5xx, feature pipeline schema unchanged, and the model file hash matches the training artifact. Over the last quarter, the alert rate on the 18-40 age band fell from 11% to 4%, while the observed sepsis rate in that band is unchanged at 9%. Calibration on the full population still looks fine (Brier 0.08). Which signal best explains the missed alerts in that band?
MCQ-21A network predicts two outputs from a protein–ligand complex: for each ligand atom i, a 3D displacement vector Δr_i = r_i^bound − r_i^input (the shift from the input pose to the predicted bound pose), and a single scalar ΔG. The team requires exact consistency under a global rigid transform g = (R, t) applied to the input: the predicted bound pose must be the input pose transformed by g, and ΔG must be unchanged. Which pair of transformation laws must the two heads satisfy?