Kristin Brooks, Contract Pharma06.11.20
Given the wealth of information and historical data that pharmaceutical organizations have at their fingertips, it’s no surprise that initiatives involving advanced technologies are being put in place to further explore the benefits and possibilities of artificial intelligence (AI) in R&D. Many organizations understand the value of AI though not many know how to get there, or even how to start, as the scope of data is so immense.
Richard Lee, Director of Core Technology at ACD/Labs, discusses the possibilities of using AI in R&D, and what this outlook means for IT leaders at pharmaceutical organizations.
Contract Pharma: What is the significance of human evaluation of data gathered by AI?
Richard Lee: AI is only as reliable as the data put into it, so being that a human is the facilitator of this, AI relies on human input. Similar to how an experienced team member reviews a less experienced peer’s work, the same process between AI and human review should be in place. If the data is not well curated or validated, it runs the risk of being unusable. Only humans have knowledge from experience to ensure the best data is captured.
Given this, cheminformatics groups within larger pharmaceutical organizations will continue to onboard data scientists to input data into the framework and set up or develop the data model. Ultimately, the scientific data that gets generated by AI systems is always going to have to be reviewed by the chemist.
CP: Do you foresee a future where AI autonomously performs all or many steps in the scientific process?
RL: Computational modeling and AI have advanced leaps and bounds, but I believe there will always be a need for some form of human intervention. I don’t foresee a time when AI will develop a completely new drug for a new human condition, because right now, AI is based on the specific data that is put into it. Think about it – if a chemist has a specific drug target and wants to model along that site, all of the information is going to be geared toward that one site, along with any pre-existing interactions available. It’s unlikely that AI, at any point, will be able to look at information from an autonomous perspective, let alone find a new drug based on a new human condition.
Right now, AI is competent at processes such as controlling robotics, dispensing materials and chemicals, monitoring reactions, knowing when to stop a reaction, and more, but nothing can replace the human mind. A scientist’s experience, knowledge, and creativity gives them the ability to think outside of the box, which is often what leads to scientific breakthroughs.
On a similar note, unexpected circumstances often arise in the R&D process, especially in drug development as the bio-variability between animals and humans is so immense and unpredictable. Even if a drug is believed to have great kinetics from the animal testing phase, it’s in the human testing phase where most drugs fail. There are an infinite number of possibilities, complexities, and biochemical pathways that prohibit the success of a drug to be predicted. The human mind, from its experiences and other knowledge gathering, can take these instances on a case-by-case basis to make connections. Many scientific breakthroughs have come from creative leaps. Computers, and by extension AI, cannot make connections between random events to generate those sparks of genius. I believe humans will use a larger proportion of their brain capacity before AI becomes sentient enough to problem solve autonomously.
CP: How can scientists leverage current technology to develop drugs?
RL: Right now, commercial and open source platforms are available to predict physicochemical properties of structures. Scientists can leverage these platforms to help predict properties such as pKa, LogD, LogP, solubility, and more. They can also leverage these platforms to generate lists of hundreds or thousands of potential structures, depending on how narrow the scope, further aiding in the drug development process. Although not all structures will be valid for a particular purpose, a scientist can use chemical patterns and an organization’s historical data to refine the platforms and generate insights.
Looking ahead, I’d like to see these technologies evolve to use information from a target site to create a predicted list of chemical structures that will likely exhibit the desired behavior. As an extension of this, AI systems should also consider the synthetic viability of those compounds. Peer-reviewed publications are full of successful reactions but for many years I’ve been talking about leveraging failed syntheses. If AI could use failed reactions as inputs as well as successful reactions, the scope and breadth of chemistries could be more focused. Furthermore, while a compound may be synthetically accessible, pharmaceutical discovery groups must also consider other factors. For example, if a rare reagent is involved the route may not be considered viable based on cost, availability, and/or scalability for production.
Ultimately, the insights generated from AI and data science activities rely completely on the quality of the input data. Without well-curated data as a foundation, AI cannot be useful for pharmaceutical R&D.
Richard obtained his Ph.D. from McMaster University, Canada, where he focused on strategies for metabolite identification and metabolomics studies. From McMaster, he moved on as a Scientist at the Centre of Probe Development and Commercialization in Hamilton, Ontario, which developed radiopharmaceuticals as imaging agents and therapeutics for oncology. He has been with ACD/Labs for the past 8 years and during this time, has been responsible for inception and development of MetaSense—software to support metabolite identification. For the last 4 years he has been ushering new technology development, laying the foundations for the next generation of ACD/Labs software.
Richard Lee, Director of Core Technology at ACD/Labs, discusses the possibilities of using AI in R&D, and what this outlook means for IT leaders at pharmaceutical organizations.
Contract Pharma: What is the significance of human evaluation of data gathered by AI?
Richard Lee: AI is only as reliable as the data put into it, so being that a human is the facilitator of this, AI relies on human input. Similar to how an experienced team member reviews a less experienced peer’s work, the same process between AI and human review should be in place. If the data is not well curated or validated, it runs the risk of being unusable. Only humans have knowledge from experience to ensure the best data is captured.
Given this, cheminformatics groups within larger pharmaceutical organizations will continue to onboard data scientists to input data into the framework and set up or develop the data model. Ultimately, the scientific data that gets generated by AI systems is always going to have to be reviewed by the chemist.
CP: Do you foresee a future where AI autonomously performs all or many steps in the scientific process?
RL: Computational modeling and AI have advanced leaps and bounds, but I believe there will always be a need for some form of human intervention. I don’t foresee a time when AI will develop a completely new drug for a new human condition, because right now, AI is based on the specific data that is put into it. Think about it – if a chemist has a specific drug target and wants to model along that site, all of the information is going to be geared toward that one site, along with any pre-existing interactions available. It’s unlikely that AI, at any point, will be able to look at information from an autonomous perspective, let alone find a new drug based on a new human condition.
Right now, AI is competent at processes such as controlling robotics, dispensing materials and chemicals, monitoring reactions, knowing when to stop a reaction, and more, but nothing can replace the human mind. A scientist’s experience, knowledge, and creativity gives them the ability to think outside of the box, which is often what leads to scientific breakthroughs.
On a similar note, unexpected circumstances often arise in the R&D process, especially in drug development as the bio-variability between animals and humans is so immense and unpredictable. Even if a drug is believed to have great kinetics from the animal testing phase, it’s in the human testing phase where most drugs fail. There are an infinite number of possibilities, complexities, and biochemical pathways that prohibit the success of a drug to be predicted. The human mind, from its experiences and other knowledge gathering, can take these instances on a case-by-case basis to make connections. Many scientific breakthroughs have come from creative leaps. Computers, and by extension AI, cannot make connections between random events to generate those sparks of genius. I believe humans will use a larger proportion of their brain capacity before AI becomes sentient enough to problem solve autonomously.
CP: How can scientists leverage current technology to develop drugs?
RL: Right now, commercial and open source platforms are available to predict physicochemical properties of structures. Scientists can leverage these platforms to help predict properties such as pKa, LogD, LogP, solubility, and more. They can also leverage these platforms to generate lists of hundreds or thousands of potential structures, depending on how narrow the scope, further aiding in the drug development process. Although not all structures will be valid for a particular purpose, a scientist can use chemical patterns and an organization’s historical data to refine the platforms and generate insights.
Looking ahead, I’d like to see these technologies evolve to use information from a target site to create a predicted list of chemical structures that will likely exhibit the desired behavior. As an extension of this, AI systems should also consider the synthetic viability of those compounds. Peer-reviewed publications are full of successful reactions but for many years I’ve been talking about leveraging failed syntheses. If AI could use failed reactions as inputs as well as successful reactions, the scope and breadth of chemistries could be more focused. Furthermore, while a compound may be synthetically accessible, pharmaceutical discovery groups must also consider other factors. For example, if a rare reagent is involved the route may not be considered viable based on cost, availability, and/or scalability for production.
Ultimately, the insights generated from AI and data science activities rely completely on the quality of the input data. Without well-curated data as a foundation, AI cannot be useful for pharmaceutical R&D.
