L17: Artificial Intelligence in Drug Development Flashcards
what is machine learning?
- Machine learning is just a “thing-labeler”
- Input: description of something
- Output: what label the thing should get
- example: if we were a machine algorithm and we saw a picture of a cat, we would go through all the schemas in our head to decipher what the stimuli is
- So in machine learning, instead of providing precise instructions to the machine about “what is a cat”, you provide examples: a couple photos of cats, a couple photos of not cats
- Once the machine has “learned” on the training set of data, you ask the machine to label (categorize) new items for you. You don’t provide the labels. ex. provide images of cats and other animals and hope that the machine can categorize correctly, if not – tweak the code
what category of items is machine learning good for?
- it is well suited for ineffable things
- defn: adjective, too great or extreme to be expressed or described in words.
- this is why we use photos in the training for machine learning
why is machine learning helpful for drug development?
- since in machine learning we use examples and not instructions, the machine was able to learn and distinguish between what made a cat - a cat?
- same theory applied for drugs: what makes a drug - a drug? we can just feed the machine instructions
- we don’t have to know the answer but the machine can find the patterns in the data and turn it into instructions that we couldn’t write yourself
- thus machine learning and AI “automates the ineffable”
how is AI used in drug discovery? (5)
1- drug design
- modelling the shape of drug targets
- predicted all the human proteins shapes
- starting to predict drug protein interactions
- program called alpha-fold
2- polypharmacology
- designing biospecific drug molecules
- designing multitarget drug molecules
3- chemical synthesis
- predict reaction yield
- predicting retrosynthesis pathways
- developing insights into reaction mechanism
- designing synthetic route
4- drug repurposing
- identifying therapeutic targets
- prediction of therapeutic use based on new structure compared to old
5- drug screening
- prediction of toxicity
- prediction of bioactivity
- prediction of physicochemical property
- identification and classification of target cells
how is AI used for in silico drug discovery – finding a good inhibitor for a specific protein
- Model a “good” inhibitor (antagonist) for a protein compared to non-inhibitors {2 categories} – feed in the machine
- We start with what we know (known inhibitors) to “train” the computers on what a good inhibitor is like
- Alternatively, predict a structure, predict inhibitors and then test the chemicals in a lab {are they really inhibiting the target?}
- Re-educate the machine about how good its predictions were
Case study for drug targets: using AI to study protein-protein interactions
ex. for new treatments of covid-19
which mechanism forms covid in our body and what would be the AI approach to help design a drug?
- we know that covid comes from novel protein protein interactions to occur
- these proteins are human + viral proteins
- we ask ourselves and the computer to disrupt a particular protein-protein interaction to cure covid infections
- we could repurpose drugs but we need to see which drugs and which targets to repurpose
how big is the protein-protein interaction problem?
- Humans have 20,000 different proteins
- thats 200 million interactions between them
- chemicals or disruptors of the interactions are not yet known
what is the artificial intelligence company at UofT
- the Vector Institute: Independent, not-for-profit corporation
- dedicated to research in the field of artificial intelligence (AI)
- excelling in machine and deep learning. * launched in March 2017
- Highly collaborative
what did the group in sanfran find when they made a protein interaction map for SARS-CoV-2?
- they made this map to reveal targets for drug repurposing
- to do so they used AI to analyze millions of proteins
- they found 26 relevant SARS-CoV-2 proteins
- they found 332 human protein-protein interactions
- of the 332 human proteins they found:
- 66 druggable human proteins
- 69 compounds target these 66 proteins
- 29 approved drugs
- 12 drugs in clinical trials
- 28 investigational drugs (preclinical stage)
how can AI be used to model pharmacokinetics of a drug (clinical trials in the computer instead of in bodies)
- Mathematical predictions of a drug’s pharmacokinetics
computer looks at the inputs the data gave:
* physiology variables (blood flow, body fat composition, etc)
* Drug’s chemical properties
* Drug-specific preclinical information
Determines the outputs of the drug, ie how/when/where it will be
* Absorption
* Distribution
* Metabolism
* Elimination
Has AI been used for testing how a drug should be taken? (5)
The AI pharmacokinetic model is used frequently for exploring the effects of:
* Other medications taken at the same time
* Food or empty stomach
* Disease status (e.g. limited kidney or liver function)
* Age
- Also to test how these factors effects the drugs exposure (how much of the drug reaches the patient in a given time)
- FDA receives this type of data frequently in regulatory submissions
- Used to guide the recommended dosing of other drugs given at the same time
- Not routinely used for regulatory decisions (approvals)
how is AI used for digital health in clinical trials?
AI can be used to via:
* Devices to monitor blood pressure, etc
* Vast amounts of data
* Smartphone technology / health information / privacy
this is used in cases where remote health monitoring is needed (for ex in a pandemic) and also to test compliance (did the person in the clinical trial actually take the pill every day for 6 months?)
- for example chat boxes with digital nurses
what is HuMod
- computer used for in silico clinical trials
- HumMod – “the most complete mathematical model of human physiology ever created”
- considers our > 10,000 variables
how are digital clinical trials made to be more inclusive?
- Use an AI tool called a trial pathfinder where it uses patient data to optimize the selection of people within the eligibility criteria while maintaining patient safety
- this is done with the goal to expand inclusion criteria for cancer clinical trials
- Uses electronic medical records and is open source
- Several partnerships with large pharmaceutical companies
- finders are James Zou and Ruishan Liu at stanford university
how are in silico clinical trials going global? who recommended this?
- European union has recommended for the use of this technology by the European medicines agency, before the drugs are given to humans
this is recommended by:
* >500 experts
* 35 countries
* 22 of 28 members of the European Union countries
* 18 month process
