AI in Drug Discovery Flashcards
Why is AI important in drug discovery?
AI accelerates molecular screening, improves prediction accuracy, reduces development time and costs, and complements rather than replaces traditional drug discovery.
What are five key benefits of using AI in drug discovery?
- Reduced animal use through in silico prediction,
- Early toxicity prediction,
- Faster synthesis via optimisation,
- Reduced screening burden,
- Cost reduction from minimising lab work.
What is the difference between AI and Machine Learning?
AI refers to systems that mimic human behaviour. Machine Learning (ML) is a subset of AI that learns from data using statistical models.
What are the three types of machine learning and their applications?
- Unsupervised ML - Clustering molecules;
- Supervised ML - Predicting binding affinity;
- Reinforcement ML - Molecule optimisation via feedback loops.
What are the characteristics and uses of Unsupervised ML in drug discovery?
Unsupervised ML learns patterns in unlabelled data, useful for data visualisation, clustering molecules, and detecting anomalies without predefined categories.
What did Aissa et al. 2021 show using unsupervised ML?
Clustering of single cells post-drug treatment into 12 subgroups, autonomously identifying control, tolerant, and sensitive cell populations.
How was clustering used with 10,000 molecular fragments in the lab?
Unsupervised ML identified initial hit compounds and helped find structurally similar analogues for drug discovery.
List limitations of Unsupervised ML.
- Requires human interpretation,
- Lacks explainability (black box),
- Difficult to assess accuracy.
What are the main concepts in Supervised ML?
Supervised ML requires labelled data and can be used for classification (e.g. active/inactive) and regression (e.g. predicting IC50 values).
How is binding affinity predicted using Supervised ML?
Drug features are input into a regression model trained on known IC50 values, which is iteratively adjusted to predict affinities for new compounds.
What tool is used for receptor activity prediction?
StarDrop (Optibrium), trained on data from known compounds to predict receptor binding of new candidates.
List limitations of Supervised ML.
- Requires extensive labelled data,
- High computational cost,
- Difficult to interpret,
- Expensive data acquisition.
What is the principle behind Reinforcement Learning?
RL uses trial and error with positive/negative feedback to optimise actions and improve molecule design over time.
How was RL used in designing DRD2 ligands?
RL model generated new molecules and learned from feedback loops to produce high-affinity DRD2 ligands.
What are the pros and cons of Reinforcement Learning?
Pros: Works with fewer data points, optimises through feedback.
Cons: Slower due to repeated cycles, needs manual or simulated feedback.
Match ML types to applications in drug discovery.
Unsupervised - Clustering & biomarker discovery;
Supervised - Binding prediction & docking scores;
Reinforcement - Molecule optimisation via feedback.
What is DDR1 and why is it important?
DDR1 is a collagen-binding kinase linked to kidney disease, liver cancer, and fibrosis. Selectivity over DDR2 is critical to avoid off-target effects.
Summarise the AI-driven workflow to discover DDR1 inhibitors.
- Literature & data curation,
- Autoencoder generates 30,000 new molecules,
- Supervised IC50 prediction,
- Docking filters down to 40 molecules,
- Patent filtering narrowed to 6 candidates,
- In vitro testing 4/6 active.
What was the result of the DDR1 AI discovery process?
From 6 candidates, 4 were active in vitro, 1 entered Phase IIa trials, with the whole process taking only 46 days (23 for computation).
List four key takeaways from the AI in drug discovery lecture.
- AI accelerates drug discovery,
- ML types vary in needs and speed,
- AI has produced clinical candidates,
- AI complements traditional approaches.
