BPS3012 Chap.3 Artificial Intelligence in Drug Discovery
Artificial Intelligence in Drug Discovery
Artificial Intelligence in Drug Discovery separates AI, machine learning, deep learning and natural language processing, then evaluates applications in candidate prioritisation and clinical trials. The chapter follows data, task, output, decision and validation, with attention to leakage, representation, bias, accountability and monitoring.
Use the terms at the right level.
Artificial intelligence is the broad capability, machine learning learns patterns from data, deep learning uses multilayer representations, and natural language processing works with language data. Naming a method does not establish value. A pharmaceutical application must specify its input, task, output and downstream decision.
A model might rank candidate molecules, predict a property, extract evidence from text or identify trial participants. Its usefulness depends on whether the output improves that defined decision under realistic conditions.
Data provenance is part of the mechanism of performance.
Training labels may be noisy, duplicated compounds can leak across splits, and a benchmark may overrepresent familiar chemical or patient groups. High internal accuracy can therefore reflect memorisation or an easy data partition rather than transferable learning.
Ask how examples were collected, whether train and test sets are genuinely independent, which population or chemical space is represented, and whether the metric matches the cost of errors. These checks turn a headline result into an evaluable scientific claim.
Validation must resemble intended use. Retrospective testing is useful, but temporal, external or prospective evaluation can probe generalisation more directly.
A good comparison includes the existing workflow: does the model find better candidates, find them faster, reduce resource use or merely reproduce expert ranking? Human oversight does not automatically remove bias or responsibility.
The workflow needs clear escalation, documentation and monitoring so that changing data or practice does not silently degrade performance after implementation.
In an exam or essay, organise evaluation as data → task → model → output → decision → validation. At each arrow, state one possible failure and the evidence that would detect it.
Avoid claims that AI will replace discovery or guarantee success; discovery failures can arise from biology, translation, safety, manufacture and clinical design even when a prediction is technically strong. A bounded conclusion recommends adoption only for the defined setting, names the incumbent comparison and identifies the prospective outcome that could reverse the recommendation.
What this chapter covers
- 01
AI, machine learning, deep learning and NLP
- 02
Data provenance, labels and split design
- 03
Candidate ranking and prospective assays
- 04
Clinical trial design and recruitment
- 05
Representation, privacy and consent
- 06
Human review and accountability
- 07
Monitoring, drift and subgroup performance
Worked example · free
Evaluate a candidate-ranking system
- +1Define the endpoint and inspect data provenance, labels and class balance.
- +1Check the split for duplicate or closely related chemistry leakage.
- +1Compare against an appropriate baseline using the same information.
- +1Prospectively select compounds before outcomes are known.
- +1Run an assay aligned with the claimed mechanism.
- +1Judge whether the ranking improves experimental yield or decision quality.
Key terms
- Artificial intelligence
- The broad field of computational systems performing tasks associated with human intelligence.
- Machine learning
- A subset of AI in which patterns are learned from data to make predictions or decisions.
- Deep learning
- A subset of machine learning using multilayer neural networks.
- Data leakage
- Information from the evaluation set entering training or model selection and inflating apparent performance.
- Model drift
- Loss of performance as data, practice or populations change after deployment.
Artificial Intelligence in Drug Discovery FAQ
Is every artificial intelligence tool deep learning?
No. Deep learning is one subset of machine learning, which is itself a subset of artificial intelligence. Natural language processing names a language-related task area and can use several model types. In an evaluation, the category label is less important than the defined input, task, output and decision. Two systems both described as AI may require completely different data and validation.
Avoid using the broad label as evidence of accuracy, novelty or pharmaceutical usefulness.
Why is prospective validation important?
It fixes the model, decision rule and experiment before outcomes are known. This tests whether the system improves a real decision rather than merely recovering patterns in historical data. The comparison should include the incumbent workflow and outcomes that matter to intended users, not only an internal accuracy metric.
Prospective testing can reveal distribution shift, operational friction and error consequences that a retrospective benchmark misses. Monitoring is still needed after deployment because data and practice can change.
How can AI affect clinical trial recruitment?
Algorithms may support site selection, eligibility screening or outreach, but they must be evaluated for representation, privacy, false exclusions, retention and subgroup performance as well as speed. Compare the full workflow with current practice and identify who is affected by each error. A faster screening step has limited value if it systematically excludes eligible people or shifts burden to manual review.
Prospective outcomes and continuing subgroup monitoring are therefore essential parts of the claim.
Assessment move
Translate every AI claim into data, task, output and decision. Then add the baseline, prospective test, affected groups and monitoring plan. Practise identifying whether a reported metric supports model accuracy, workflow value or a clinical outcome, because those evidence levels cannot be substituted for one another.
