GSK Data Scientist Interview: Questions & Prep (2026)

GSK (GlaxoSmithKline) is a global biopharma company and one of the more active pharma employers hiring data scientists in India right now. The knok jobradar counted 64 open Data Scientist roles at GSK as of July 2026, against a national market of 937 Data Scientist positions live across all employers.

GSK uses data science across drug discovery, clinical trial analytics, commercial and marketing analytics, and patient outcomes research. Candidates typically go through several rounds: an initial HR or recruiter screen, a technical interview (or take-home case study), and one or two panel rounds with data science leads and business stakeholders. Candidates report that pharma domain context matters here, even if you come from a different industry.

What makes GSK interviews distinctive is the blend of rigorous statistical thinking (clinical trial design, hypothesis testing, survival analysis) with strong business communication skills. They want people who can talk to a medical director as clearly as they talk to an engineer.

These questions come up repeatedly, based on what candidates report from GSK data science interviews:

1. How would you design an experiment to measure whether a new drug marketing campaign changed prescription rates? Walk us through your hypothesis, metrics, and statistical approach.
2. Tell me about a machine learning project where you drove a measurable business outcome. What was the problem, what did you build, and what happened after deployment?
3. Clinical datasets often have significant missing data. What strategies do you use, and how do you decide which one to apply?
4. Explain Type I and Type II errors in plain language. If you were testing a drug's effectiveness, which error would you be more concerned about, and why?
5. How would you build a model to predict which patients are likely to stop taking their medication (non-adherence)? What features would you use, and how would you validate it?
6. How do you explain a complex predictive model to a non-technical audience, like a marketing director or medical lead?
7. GSK handles sensitive patient data. What does responsible data use mean to you in practice, not just in theory?
8. Describe a time when your model produced unexpected or wrong results. How did you diagnose and fix it?
9. How do you approach severe class imbalance, for example predicting a rare adverse drug event that occurs in very few patients?
10. Do you have experience with survival analysis or time-to-event models? Describe a scenario where you would use a Cox proportional hazards model.
11. After a model goes live, how do you track whether it is still working and delivering the expected business value?
12. How would you use NLP to extract insights from medical literature or clinical notes to support a research team?

Q: Tell me about a machine learning project where you drove a measurable business outcome.

*Situation:* At my previous company, the sales team was losing time chasing leads that almost never converted, and there was no data-driven way to prioritise their calls.

*Task:* I was asked to build a lead scoring model that could rank prospects by conversion likelihood, so the sales team could focus on the most promising accounts first.

*Action:* I worked with the sales ops team to pull several years of CRM data, engineered features around past purchase behaviour, engagement signals, and firmographic data, and trained a gradient boosting classifier. I ran cross-validation carefully, tracked precision and recall separately (not just accuracy), and pre-registered my primary success metric before showing results to the business, so there was no post-hoc cherry-picking.

*Result:* The model was deployed as a daily-updated dashboard. In the quarter after rollout, the sales team reported that top-ranked leads converted at a much higher rate than their previous manual approach, and management approved scaling it to two more product lines.

---

Q: Describe a time when your model produced unexpected or wrong results. How did you diagnose and fix it?

*Situation:* I had built a churn prediction model that performed well on the validation set, but when it went live, the alerts it generated were largely wrong and the business team quickly lost trust in it.

*Task:* I needed to diagnose what went wrong and either fix the model or explain why the business should retire it.

*Action:* I first checked for data leakage and found none. Then I compared the distribution of features in training data versus live data and discovered the product had undergone a major UI change after my training cutoff, which shifted user behaviour patterns significantly. I retrained the model on more recent data, added a data drift monitor, and set up an alert if key feature distributions shifted beyond a set threshold.

*Result:* The retrained model performed much more reliably. More importantly, the team now had a process for catching drift early rather than discovering it after the business had already lost confidence in the output.

---

Q: How do you explain a complex predictive model to a non-technical audience?

*Situation:* I had built a patient segmentation model for a healthcare client that grouped patients based on their treatment history and outcomes. The medical director needed to understand and trust the output before acting on it.

*Task:* I had to present findings to a room of clinicians and business leads who had no machine learning background.

*Action:* Instead of showing model architecture or feature importance plots, I led with the business question: 'Which patient groups are most at risk of readmission within a quarter?' I used plain-language cluster descriptions (for example, 'this group tends to be older patients with multiple chronic conditions who miss follow-up appointments') and backed each description with real patient examples from the data. I prepared one slide on how the model was validated, focused on what it got right and where it was uncertain.

*Result:* The medical director approved a pilot intervention programme targeting the highest-risk cluster. The clinicians felt they understood and trusted the segmentation well enough to act on it, which is usually the hardest part of applied data science in healthcare.

For experiment design questions (Q1 above), use this structure: state the null hypothesis, define your primary metric and why you chose it, explain how you would randomise and control for confounders, describe your sample size thinking (without invented numbers), and explain what a significant result would mean for the business decision.

For machine learning case questions (Q2, Q5, Q9), follow the problem-to-production arc: (a) define the business problem precisely, (b) describe the data you would need and where it comes from, (c) explain your modelling choices and why, (d) describe how you would evaluate the model, and (e) explain how you would monitor it after deployment. GSK interviewers pay particular attention to evaluation: a candidate who reaches for accuracy on an imbalanced dataset without discussing precision, recall, or AUC will lose credibility fast.

For ethics and compliance questions (Q7), show that you think beyond GDPR checkbox compliance. Talk about data minimisation, anonymisation, the risk of re-identification in small patient subgroups, and how you would raise a concern if you spotted a privacy risk in a project you were working on.

For stakeholder communication questions (Q6), the BLUF (Bottom Line Up Front) approach works well: lead with the decision the business needs to make, then support it with data. Avoid leading with methodology.

GSK data science interviewers typically look for four things, based on what candidates report:

Statistical rigour. Pharma sets a high bar here because the cost of a wrong conclusion is real. Expect to talk about p-values, confidence intervals, multiple testing corrections, and power calculations. If you have worked with clinical or real-world evidence data, say so explicitly.

Domain curiosity. You do not need a biology PhD. But candidates who have done basic homework on how clinical trials work, what RWE (real-world evidence) means, or what commercial analytics looks like in pharma tend to stand out. Read a GSK annual report and one or two public data science blog posts from pharma companies before your interview.

Communication that non-scientists can follow. GSK data scientists regularly work with medical directors, commercial leads, and regulatory teams. Interviewers will probe whether you can simplify without being misleading.

Ownership and honesty about failure. The sample answer about the model that went wrong is a real type of question that tests whether you debug systematically and learn from mistakes, rather than deflecting or blaming data quality.

Week 1: Technical foundations
Revise statistical hypothesis testing, A/B testing design, and experiment evaluation. Review gradient boosting methods (XGBoost, LightGBM) and how you would explain feature importance. Revise survival analysis basics if you have not used them recently.

Week 2: Pharma domain context
Spend a few hours learning how randomised controlled trials (RCTs) work and what real-world evidence (RWE) means in pharma. Read GSK's publicly available annual report for context on their therapeutic focus areas. Look up 'pharmacovigilance data science' to understand the adverse event detection problem.

Week 3: Practice and mock interviews
Practise answering the questions above out loud, not just in your head. Record yourself once and watch it back; most people find they use far more jargon than they realise. Do at least one mock interview with a peer who can give honest feedback.

All along: Build your story bank
For every STAR question, you need a concrete project from your own experience. Map your past work to the likely question themes: experiment design, model failure, stakeholder communication, class imbalance, and responsible data use. If you have a gap (for example, no clinical data experience), prepare an honest 'here is how I would approach it' answer instead of bluffing.

As of July 2026, knok shows 64 open roles at GSK in India across all levels. Knok checks 150+ job sites nightly, matches openings to your resume, and messages HR on your behalf so you do not miss a role while you are busy preparing.

Treating accuracy as the only metric. In pharma and healthcare, false negatives (missing a real signal) can be as costly as false positives. Always address precision, recall, and the business cost of each error type.

Skipping the 'so what'. Interviewers at a company like GSK want to know what decision your analysis enabled. Stopping at 'the model had good AUC' is not enough. Always follow through to the business outcome.

Overclaiming on domain experience you do not have. If you have not worked with clinical data, say so, then pivot to what transferable skills you bring and how you would ramp up quickly. Bluffing pharma domain knowledge is easy to detect.

Not asking clarifying questions in case studies. When given an ambiguous problem, candidates who jump straight to modelling without asking about the business goal, data availability, and constraints look like they lack product thinking.

Ignoring data privacy in your answers. Any answer involving patient data that does not mention privacy, anonymisation, or compliance will raise a red flag at a regulated company like GSK.

Underestimating the communication rounds. Some candidates prepare heavily for the coding or statistics rounds but walk into the stakeholder communication round unprepared. Practise explaining your best project to someone outside data science before the interview.