Pharmacology · Short notes

Artificial Intelligence in Pharmacology: Short Notes & Cheat Sheet

Pharmacology and Therapeutics II (Artificial Intelligence in Pharmacology and Therapeutics): machine learning basics, AI in drug discovery, repurposing, dosing and pharmacovigilance, model metrics and limits.

Short notes

Read these before you attempt the topic's MCQs.

Where AI enters the drug pipeline

1TargetStructure prediction2Hit findingVirtual screening3Lead optimisationADMET prediction4Clinical trialsPatient selection5PrescribingDecision support, dosing6Post-marketingSignal detection
AI tools now touch every stage, from finding a target to watching a medicine's safety after launch.

Core ideas

  • Artificial intelligence (AI) is software that performs tasks needing human-like judgement. Machine learning (ML) learns patterns from data instead of following hand-written rules.
  • Supervised learning uses labelled data for classification (toxic or not) or regression (predicting a value such as clearance). Unsupervised learning finds groups in unlabelled data (patient clustering). Reinforcement learning learns by reward.
  • Deep learning uses many-layered neural networks: convolutional networks for images (histology, retinal scans) and transformers for sequences, text and molecules.
  • Natural language processing (NLP) reads free text such as case reports and clinical notes; large language models generate text and can answer questions, but can also invent facts.

AI in drug discovery

  • Target identification and protein structure: AlphaFold predicts 3D protein structures from sequence; its developers shared the 2024 Nobel Prize in Chemistry.
  • Virtual screening and QSAR models rank millions of molecules before any are made; generative models propose new structures (de novo design).
  • ADMET prediction flags poor absorption, metabolism or toxicity (hERG blockade, hepatotoxicity) early, cutting late failures.
  • Drug repurposing: in 2020 an AI knowledge graph proposed baricitinib for COVID-19, and it was later shown to benefit hospitalised patients.

AI in therapeutics and practice

  • Clinical decision support screens prescriptions for interactions, allergies and renal dose adjustments.
  • Model-informed precision dosing uses Bayesian software to individualise doses from a few drug levels, for example vancomycin AUC-guided dosing.
  • Pharmacovigilance: ML and NLP detect safety signals in spontaneous reports, health records and social media faster than manual review.
  • Pharmacogenomics and precision medicine: models combine genotype and clinical data to predict response or toxicity.
  • Clinical trials: AI helps find eligible patients, predict drop-out and monitor data quality.

Judging a model, and its limits

  • Data are split into training, validation and test sets; cross-validation checks that performance holds on data the model has not seen.
  • Overfitting means the model memorises training data and fails on new data.
  • Metrics: sensitivity, specificity, precision and the area under the ROC curve (AUC, 0.5 = chance, 1.0 = perfect). Accuracy alone misleads when one outcome is rare.
  • Limits: biased or unrepresentative training data, 'black box' decisions that are hard to explain, data privacy, and hallucinated answers from language models.
  • AI supports but does not replace the pharmacist's judgement; outputs must be checked against references and the patient in front of you.

Cheat sheet

High-yield facts in one place. Cover the right-hand column and test yourself.

Learning type → pharmacy example

TypeExample
Supervised classificationPredicting hepatotoxicity (yes/no)
Supervised regressionPredicting clearance or solubility
Unsupervised clusteringGrouping patients by response
Deep learning (CNN)Reading histology or retinal images
NLPMining adverse event reports
Generative modelsDesigning new molecules

Metric → meaning

MetricMeaning
SensitivityTrue positives found among all positives
SpecificityTrue negatives found among all negatives
Precision (PPV)Positives predicted that are correct
AUC-ROC0.5 chance · 1.0 perfect
OverfittingGood on training data, poor on new data

Memory aids

Garbage in, garbage out
A model is only as good and as fair as the data it was trained on.

Exam traps

  • High accuracy can hide a useless model when the outcome is rare; check sensitivity and AUC.
  • AlphaFold predicts protein structure; it does not by itself prove that a drug will bind or work.
  • A correlation found by a model is not proof of cause.
  • Language models can produce confident but wrong doses; always verify against a reference.

Revised? Now test yourself.

Attempt the Artificial Intelligence in Pharmacology MCQs while the notes are fresh. Anything you get wrong is saved to your incorrect list for the next revision round.

More Pharmacology notes