AI in Drug Discovery

Question 1

Where can AI be used in drug discovery?

Answer 1

• Molecular library screening
• Target identification
• Preclinical studies
• Drug repurposing
• De novo drug design

Question 2

What are the benefits of AI-driven approaches?

Answer 2

• Reduced animal use
• Toxicity predictions earlier in the clinical pipeline
• Faster and easier synthesis
• Significantly reduced screening requirements
• Significantly less expenditure

Question 3

AI vs. Machine learning vs. Deep learning

Answer 3

• AI is any technique where machines attempt to mimic human behaviour
• Machine learning is a subset of AI whereby statistics is used to enable machines/algorithms to improve with experience
• Deep learning is a subset of ML which makes the computation of multi-layer neural networks feasible

Question 4

Subtypes of machine learning

Answer 4

• Unsupervised learning: Unlabelled data is given to an algorithm and it tries to find patterns
• Supervised learning: Labelled data is given to an algorithm and it tries to fit or understand how the labels relate to the data
• Reinforcement learning: An algorithm tries to interpret data with constant positive and/or negative feedback

Question 5

Properties of unsupervised machine learning

Answer 5

• Data is analysed in an unbiased manner
• Requires very little manual intervention - less arduous
• Can be used to discover anomalies in data
• Identifies sets of items that often occur together
• Is heavily used for data visualisation and interpretation (the algorithm doesn’t know what the categories are but it sorts them out for you)

Question 6

Disadvantages of unsupervised machine learning

Answer 6

• Human interpretation is needed to see if the predicted clustering visualisation makes sense
• You cannot easily get precise reasons for why the clusters were assigned in a particular way
• Accuracy of the clustering is hard to measure

Question 7

Supervised machine learning subtypes

Answer 7

• Classification: The algorithm tries to learn how to predict a label for a sample given its features
• Regression: The algorithm tries to learn how to predict a value for a sample given its features

Question 8

Properties of supervised machine learning

Answer 8

• The ultimate goal is to be able to take a set of features from unseen data and predict their labels or values
• This is done by learning from a previously generated set of data and generating a model that is able to predict labels or values based on the features of the new data
• Fitting the data at different iterations and optimising the line of best fit

Question 9

Disadvantages of supervised machine learning

Answer 9

• Requires labelling of data or assignment to groups (can be costly)
• Data requirements can be high (minimum hundreds of data points)
• Interpretation of algorithms can be hard
• Can be computationally costly

Question 10

Reinforcement learning

Answer 10

• The algorithm learns through trial and error by making predictions and receiving positive or negative feedback and adjusting itself to improve
• Much slower than other ML types as it involves feedback loops whereby new data is collected or labelled
• Reinforcement learning has the potential to learn accurate models with significantly fewer data points than supervised learning
• Requires lots of manual interaction

Question 11

Case study: DDR1 inhibitor

Answer 11

• A novel DDR1 inhibitor with high patentability was discovered in only 46 days using ML
• Existing compounds and their IC50 values against DDR1 were used in a supervised learning model to predict IC50 for 30000 potential compounds
• These compounds were reduced to 40 with the best IC50
• These 40 were reduced to 6 with the best patentability
• Of these 6, 4 were effective and one compound has now completed Phase 2a clinical trials