AI in drug development

what are the problems in drug development that could be solved with the use of AI

problems

• Expensive- 2-3 billion dollars to make a single drug

• Time consuming 10-15 years- researchers have to search through large amounts of biological and chemical data

• High risk- 90% of candidate drugs fail before approval

  • Failure usually due to lack of efficacy, toxicity, poor PK- which could have been predicted earlier

  • Failure at phase II waste billions

how AI can help

• AI can address this by processing biological, chemical and clinical data to allow for earlier and more accurate predictions at each stage of the pipeline

• This helps reduce attrition, costs, shorten development (3-8 years), improves target selections, also allows for personalised medicine, improved safety profiles

describe the traditional computational methods used in drug development

• In silico screening- HTS in computer, fits molecules into protein target and scores how well they fit. Top ranked molecules progress to in vitro assay

• De novo design- screen library of fragments that fit into the target site and modify them to see how they affect the binding at target site

• QSAR- predict biological activity of a molecule, detailed computer modelling of ligand properties and this is correlated with activity data at target site e.g. CoMFA predicts biological activity of molecules based on 3D shapes (3D QSAR)

describe machine learning and how it works

• we don’t have to have preset algorithms and ideas been tested

• the machine sources out the information. It uses large datasets to train the algorithms to improve through experience

• 2 types- supervised and unsupervised

describe the 2 types of machine learning including their use in drug development and pros/cons

supervised

• Human input is used to train the AI what to look for

  • classification- algorithm is used to assign data set into specific categories

  • Regression- algorithm is used to understand relationship between dependent and independent variable

• Uses- predict drug-target interactions, toxicity and efficacy

• Pro- high accuracy and easier validation 

• Con- requires high quality labelled datasets

unsupervised

• ML discovers hidden patterns in data without need for human intervention, though used at the end to filter what is needed

  • clustering- grouping unlabelled data based on similarities and differences

  • Association- uses different rules to find relationships between variables

  • Dimensional reduction- used to remove noise from the data to reduce the dataset to manageable size

• Uses- drug clustering, patient stratification, biomarker discovery

• Pros- discover new relationships and new disease subtype

• Con- identifies meaningless correlations

describe deep learning

• allows the system to learn from data on its own and interrogate the data more

• Uses neural networks made from layers of nodes that mirror the neural framework of our brain with the way it inputs, processes and outputs data

• Structure-

  • Input layer- this is where all the data goes in

  • Hidden layer- do the heavy thinking, analyse patterns and learn from the data

  • Output layer- gives the final prediction

• Pro- more powerful, allows integration of highly complex data beyond traditional statistics

give an example of a deep learning AI system used for drug discovery

AlphaFold

• Used for target identification and structure prediction

• Problem- many proteins are potential drug targets but drug design often requires knowledge of the 3D structure

• Traditional methods such as x ray crystallography, cryo-EM,  NMR are expensive and technically difficult

• Alpha fold uses deep learning to predict protein folding and protein structure directly from amino acid sequence

• knowing protein structure is important as it determines ligand binding, receptor activation, enzyme activity and knowing structure allows rational drug design and identification of binding pockets

• used in drug development to generate biological knowledge that was previously inaccessible

give an example of an AI-driven drug

Insilico Medicine (INS018_055)

• Disease- idiopathic pulmonary fibrosis

• Drug discovery for fibrosis is difficult because multiple pathways involved and poorly understood mechanisms and high failure rate

• Machine learning analysed biological datasets and identified novel pathways involved in fibrosis

• Generative AI designed candidate drugs rather than screening millions of existing compounds

• Candidate drug progressed from target discovery to preclinical candidate in approx 18 months compared to 4-6 years

name the 4 phases in drug development pipeline where AI can be implemented

target identification

Hit identification/LI/LO

preclinical testing

clinical trial optimisation

describe how AI can be implemented in target identification phase

• Used to analyse large scale biological datasets to identify novel therapeutic targets involved in disease pathology

• AI can identify patterns, correlations and disease-associated molecular signatures that may not be apparent to humans

• Can analyse differences between healthy and diseased tissues to identify genes, proteins, signalling pathways associated with disease

• Accelerates target discovery but reduces likelihood that good targets will be overlooked due to the complexity of the biological data

describe how AI can be implemented in HIT identification (LI/LO) phase

• Can screen vast numbers of compounds and predict which molecules are most likely to interact effectively with a therapeutic target

• AI driven screening allows researchers to computationally evaluate billions of compounds

• Can predict which molecules are most likely to bind to a target protein based on structure-activity relationships

• Can estimate important relationships such as binding affinity, selectivity, solubility- poor candidates eliminated earlier and resources focused on promising candidates

• Reduces time and cost required to identify a viable lead compound

describe how AI can be implemented in Preclinical testing phase

• Used to predict how candidate drugs will behave in biological systems before extensive animal and laboratory testing

• AI can predict ADME properties - used to estimate whether a compound will reach its target tissue and achieve therapeutic concentrations

• Can predict potential toxicities and DDI and identify safety concerns earlier so unsuitable compounds are removed before costly animal studies are done

• Can simulate how drugs interact with cells, tissues and biological pathways which decreases likelihood of late-stage failure

describe how AI can be implemented in Clinical trial optimisation phase

• Being used to improve clinical trial design, patient recruitment, monitoring and data analysis

• Increases trial success rates and reduces development costs

• Challenges of clinical research- identifying patients who will most likely benefit from treatment

• AI can analyse genetic data, biomarkers, medical records to identify patient populations more appropriate for a particular trial

• Supports precision medicine by enriching clinical trials with patients who possess molecular characteristics relevant to the drugs MOA

describe an example of AI in drug repurposing

Baricitinib for COVID 19 (Benevolent AI)

• Original indication- RA

• MOA- JAK1/2 inhibitor that suppresses the inflammatory cytokine cascade driving the destruction of joints in RA

• The role of AI

  • used to identify current drug that inhibits viral entry or replication and/or suppress the inflammatory response

  • used AI-driven knowledge graph which contained biomedical data on drugs, genes, proteins alongside mechanisms, processes and pathways of already approved drugs with anti-inflammatory activity

  • KG identified AAK1 (AP2-associated protein kinase 1) as a key regulator of clathrin-mediated endocytosis- used by the SARS-CoV-2 virus to enter cells

  • Among the already FDA-approved drugs that inhibit AAK1, baricitinib ranked the highest in efficacy; became key part of the COVID-19 treatment protocols

    • dual MOA- JAK1/2 +AAK1 inhibitor = anti-inflammatory and blocked viral entry

• Timeline

  • AI prediction to clinical trial entry ~ 3 months

  • From prediction to FDA approval ~ 28 months compared to usual 12-15 yrs

  • However baricitinb success was also helped by the high demand during that global pandemic; there was regulatory flexibility (emergency use authorisation)

describe the limitations of using AI in drug discovery

• Data quality and availability

  • AI is only as good as the data used and biological datasets are often incomplete, bias towards well-studied diseases (white EU populations) or unavailable in public domain

• Black box problem

  • Deep learning models produce accurate predictions without mechanistic explanations of the drugs MOA that is needed by regulatory agencies

  • The internal decision-making process is difficult for humans to interpret

  • Drug development requires scientifically justified reasons so without a mechanistic interpretation the data cannot be trusted

• Novelty ceiling

  • Generative AI can produce molecules that are synthetically inaccessible or non-producible and that puts researchers back at square 1