Exercise: Match the Question to the Method
No computer needed. Pen, paper, and argument.
How this works
Six scenarios, all drawn from real Southern African genomics research. For each one, decide:
- Is this supervised, unsupervised, generative, or not an AI problem at all?
- What method would you actually use?
- What could go wrong?
Question 3 is the one that matters. Work in pairs, 20 minutes, then we discuss.
“Not an AI problem” is correct more often than people expect. Do not talk yourself out of it.
Scenario 1 — Genotype assignment
You have 6,000 HBV sequences. For 500 of them a reference lab has assigned a genotype (A, B, C, D, E). You want genotypes for the other 5,500.
- Supervised, unsupervised, generative, or not AI?
- What would you use?
- What could go wrong?
Supervised — you have labels and want to extend them.
But before training anything: this is a solved problem with a known mechanism. Genotype is defined by sequence divergence, so alignment against reference genotypes (or a tool like jpHMM) is the standard approach and it is interpretable — you can point at the evidence.
What could go wrong: the 500 labelled sequences may not be a random sample. If the reference lab preferentially genotyped unusual cases, your labels are biased and the model inherits that. Recombinant sequences break everything — if a genome is part A and part E, a whole-sequence classifier must pick one and will be confidently wrong. That is precisely why jpHMM scans along the genome instead of classifying it as a unit.
Scenario 2 — Structure in haplogroups
You have 338 mtDNA haplogroup calls. Nobody has grouped them. You want to know whether the cohort contains distinct ancestry clusters.
Unsupervised — no labels, looking for structure.
But note that haplogroups are already a hierarchical classification built by phylogenetics. Clustering them is close to re-deriving a tree that exists. The honest first move is count() and a bar chart.
What could go wrong: clustering always returns clusters. With 62 fine-grained labels over 338 people, most labels have a handful of members — you will find “structure” that is sampling noise. This dataset is used in the case study; it collapses to 4 macro-groups for exactly this reason.
Scenario 3 — Drug resistance from sequence
You have 200 sequenced isolates with known resistance phenotypes. You want to predict resistance for new isolates from sequence alone.
Supervised — and the tempting answer is “train a classifier”.
With 200 examples and thousands of positions, you have far more features than samples. A flexible model will memorise. Start with known resistance mutations from the literature — mechanism you already have — and use regression on those positions.
What could go wrong: severe class imbalance. If 15 of 200 are resistant, a model predicting “susceptible” always is 92% accurate and clinically dangerous. And a model trained on one subtype may not transfer to another.
Scenario 4 — Does haplogroup associate with HIV status?
338 participants, mtDNA haplogroup and HIV status for each. Does ancestry associate with infection status?
Not an AI problem. Four groups, one binary outcome, 338 people. This is a contingency table and Fisher’s exact test.
What could go wrong: everything — and not for statistical reasons. In the real version of this dataset, every male is HIV-negative and every HIV-positive participant is female. Sex is perfectly confounded with outcome, so no method — Fisher, regression, or a neural network — can separate a haplogroup effect from a sex effect.
The result is not “no association”. It is “this design cannot answer the question”. You would find this with one table() call and never find it by tuning a model. This is the case study.
Scenario 5 — Writing your methods section
You want an LLM to draft the methods section describing your pipeline.
Generative — and legitimate, with conditions.
Useful for: restructuring text you wrote, tightening prose, drafting boilerplate you will verify.
What could go wrong: LLMs produce fluent, plausible, wrong text. They will invent a citation that looks perfect. They will state a version number you never used. Everything factual — versions, parameters, references — must be checked against your actual pipeline.
Also: journals require disclosure of AI assistance, and you remain responsible for every sentence. “The model wrote it” is not a defence.
Scenario 6 — Predicting who progresses
You have 50 patients with longitudinal viral load and want to predict who progresses to liver disease.
Supervised, and the honest answer is you probably cannot do this yet.
Fifty patients, with perhaps 12 events. The rule of thumb is ~10 events per predictor, so you can justify roughly one variable. Not a model — a variable.
What could go wrong: with n=50 you can build something that looks impressive and means nothing. Informative censoring makes it worse: patients lost to follow-up may be the sickest.
The valuable contribution here is saying so — and specifying what cohort size would be needed. Telling a team their study is underpowered before they spend three years on it is worth more than any model.
Wrapping up
Across six scenarios drawn from real work:
- Two were not AI problems (4, and arguably 1).
- One was underpowered regardless of method (6).
- One had a fatal design flaw no method could repair (4).
- The failure modes were about data and design, never about algorithms.
The hard part of applying AI in biology is almost never the model. It is knowing whether the question can be answered by the data you have. That judgement is the thing this workshop is actually trying to teach.
Next: AI and Application.