3. Where They Meet

AI has genuinely changed parts of biology. It has changed other parts not at all. Knowing which is which is a research skill.

Learning objectives

  1. Name problems where AI has produced real biological advances.
  2. Name problems where classical methods remain better.
  3. Apply a short test to decide whether your problem suits an AI approach.
  4. Explain why interpretability matters more in clinical settings.

Where it genuinely worked

Protein structure prediction. AlphaFold changed this field. The problem had the exact properties deep learning needs: a large corpus of solved structures, a well-defined input (sequence) and output (coordinates), and a clear measure of success. Structures that took years now take hours.

Basecalling. Nanopore basecalling is a neural network turning electrical signal into bases. Accuracy improvements here came from better models, not better chemistry.

Variant effect prediction. Predicting whether a missense variant is damaging is now dominated by learned models.

Image-heavy tasks. Cell segmentation, histopathology screening — anywhere the input is pixels.

Where it mostly has not

Small-sample association studies. With 338 participants and 4 groups, a contingency table and Fisher’s exact test are the right tools. A neural network would overfit instantly and tell you less.

Anything where you must explain the answer. If a clinician must act on the output, “the model said so” is not sufficient. A logistic regression with interpretable coefficients may be worth more than an opaque model that scores two points higher.

Problems with a known mechanism. We understand how sequences align. Use the algorithm.

Confounded designs. No model repairs a broken study. This is the lesson of the case study — and it is worth stating plainly: AI cannot fix data that never contained the answer.

A short test

Before reaching for machine learning, ask:

  1. How many examples do I have? Hundreds → classical statistics. Tens of thousands → ML becomes reasonable.
  2. Do I have trustworthy labels? ML inherits every bias in your labels.
  3. Do I need to explain the answer? If yes, weight interpretability heavily.
  4. Is there a known mechanism? If yes, use it.
  5. Would a simple model do? Try it first. It is your baseline, and it is often the winner.
TipAlways fit the simple model first

A logistic regression takes two minutes and gives you a number to beat. If your deep model cannot beat it, that is the finding — and you learned it cheaply.

The honest summary

AI is a set of tools with a specific profile: excellent when you have many examples, a clear objective, and tolerance for opacity. Southern African genomics research frequently has none of those things — cohorts are small, questions are causal, and results inform clinical decisions.

That is not a reason to avoid these methods. It is a reason to use them where they fit, and to be the person in the room who can say why they don’t.

ImportantWhat we want you to leave with

Not “AI is overhyped” and not “AI will solve this”. Both are lazy. The useful position is: this problem has these properties, therefore this method — and being able to defend that sentence.

Check yourself

  1. You have 200 sequenced isolates and want to predict drug resistance from sequence. Neural network or regression? What would you need to change your answer?
  2. Why is interpretability more valuable in a clinic than in a screening pipeline?
  3. Your model beats logistic regression by 1%. Is it better?

Take these into the exercise.