Why Are Medicinal Chemists Reluctant to Use AI/ML Tools?

Posted by
Phil McHale, Dóra Barna
on 2023-11-02

Why Are Medicinal Chemists Reluctant to Use AI/ML Tools?

There is a lot of current chatter about how AI/ML tools like ChatGPT will soon take over several areas of human endeavour: so should medicinal chemists be worried that they too will soon be replaced by algorithms?  

The use of AI/ML in drug discovery and design is a burgeoning area, with > 200 startups active in the space and > $10B investments raised, and a claimed first AI-developed drug recently entered Phase II clinical trials - so why are some medicinal chemists slow to jump on the band wagon? 

Part of this may be due to their experience of living through multiple previous Gartner Hype Cycles relating to other drug design phases (e.g. computer-aided drug design, rational drug design, combinatorial chemistry, high throughput screening, etc.), and their natural scepticism about anything new and over-hyped like AI/ML.

But putting the scepticism aside, do medicinal chemists have bona fide issues and concerns about the use of AI/ML in drug design and development? This paper focuses on ML-based models rather than other AI tools (e.g. robotics, etc.) and discusses several areas where chemists’ concerns may be justified. A follow-on article will describe ways to mitigate and overcome these issues. 

Higher Level Concerns

There are some genuine higher-level concerns about the use of AI/ML tools in drug design that are beyond most medicinal chemists’ pay grade to solve, and they are mentioned here for completeness:

  • High cost and infrastructure requirements. AI/ML techniques in drug design and development may require significant computational resources and infrastructure; and training complex models and handling large datasets can be computationally intensive and expensive. To address these computational challenges, organizations increasingly turn to cloud-based services to try to mitigate the expense: but this in turn raises other issues. See https://chemaxon.com/blog/chemistry-in-the-cloud for a discussion of these.
  • Regulatory and ethical considerations. Regulatory agencies may require transparent and explainable models, which can be challenging with complex ML algorithms; and there may be ethical considerations when using AI to make decisions that can impact human health, such as predicting toxicity or efficacy. 

Medicinal Chemists’ Concerns

As long ago as 2018, a technical director at a medchem consultancy suggested that "AI won’t replace medicinal chemists, but medicinal chemists who use AI will replace those who don’t" – so greater interest in and use of the tools might be expected if chemists’ concerns can be overcome. These fall into three main areas, each of which is an impediment to successful and productive uptake of AI/ML tools.

  • Lack of quality and availability of suitable data. AI and ML algorithms rely on large amounts of high-quality data for successful model training and validation. There is a scarcity of data itself, and the data that is available is often in disparate formats and of insufficient quality to derive high quality predictions and models. These issues, coupled with a lack of suitable pre-processing tools for data and chemical structure cleaning and standardization can lead to the generation of inaccurate or misleading models and predictions, which in turn diminish the value of the tools for medicinal chemists.
  • Lack of interpretability and transparency of the underlying decision-making.  Many AI and ML models such as deep learning neural networks use opaque and difficult-to-interpret decision-making processes based on arcane ML techniques, so they come to be viewed (or denigrated) as black boxes. Typical practising medicinal chemists don’t know their Random Forest from their Gradient Boosted Trees, so they disdain black boxes and regard them with suspicion. Medicinal chemists who can’t understand and explain the scientific rationale for the underlying AI/ML decision-making processes won’t trust the tools, which will then be left as the sole domain of the computational chemists.
  • Difficulty integrating AI/ML tools into existing drug discovery workflows. If the AI/ML tools can’t be seamlessly integrated into existing drug discovery workflows and processes, and used as routine steps in discovery procedures, they will largely be ignored, and uptake will be hindered. Similarly, if the tools insist on the chemists learning a different and disconnected user interface, or they need access to separate files that may require reformatting or cleaning, they will tend to be unused by medicinal chemists and will remain within the computational chemistry group. 

Overcoming the Concerns

If organizations accept the premise that medicinal chemists who use AI/ML will be more effective than those that don’t, they will need help to overcome the objections described above to ensure widespread and effective uptake of the tools. 

Provision of high-quality data in sufficient volume for reliably predicting important, high-performance descriptors and models; access to interpretable and explainable underlying ML decision-making rationales; and user-friendly AI/ML-based tools integrated into drug discovery workflows and pipelines are all possible now, and a second paper in this series will discuss how to mitigate the concerns and achieve these objectives.

Read More

 

 

There is a lot of current chatter about how AI/ML tools like ChatGPT will soon take over several areas of human endeavour: so should medicinal chemists be worried that they too will soon be replaced by algorithms?  

The use of AI/ML in drug discovery and design is a burgeoning area, with > 200 startups active in the space and > $10B investments raised, and a claimed first AI-developed drug recently entered Phase II clinical trials - so why are some medicinal chemists slow to jump on the band wagon? 

Part of this may be due to their experience of living through multiple previous Gartner Hype Cycles relating to other drug design phases (e.g. computer-aided drug design, rational drug design, combinatorial chemistry, high throughput screening, etc.), and their natural scepticism about anything new and over-hyped like AI/ML.

But putting the scepticism aside, do medicinal chemists have bona fide issues and concerns about the use of AI/ML in drug design and development? This paper focuses on ML-based models rather than other AI tools (e.g. robotics, etc.) and discusses several areas where chemists’ concerns may be justified. A follow-on article will describe ways to mitigate and overcome these issues. 

Higher Level Concerns

There are some genuine higher-level concerns about the use of AI/ML tools in drug design that are beyond most medicinal chemists’ pay grade to solve, and they are mentioned here for completeness:

  • High cost and infrastructure requirements. AI/ML techniques in drug design and development may require significant computational resources and infrastructure; and training complex models and handling large datasets can be computationally intensive and expensive. To address these computational challenges, organizations increasingly turn to cloud-based services to try to mitigate the expense: but this in turn raises other issues. See https://chemaxon.com/blog/chemistry-in-the-cloud for a discussion of these.
  • Regulatory and ethical considerations. Regulatory agencies may require transparent and explainable models, which can be challenging with complex ML algorithms; and there may be ethical considerations when using AI to make decisions that can impact human health, such as predicting toxicity or efficacy. 

Medicinal Chemists’ Concerns

As long ago as 2018, a technical director at a medchem consultancy suggested that "AI won’t replace medicinal chemists, but medicinal chemists who use AI will replace those who don’t" – so greater interest in and use of the tools might be expected if chemists’ concerns can be overcome. These fall into three main areas, each of which is an impediment to successful and productive uptake of AI/ML tools.

  • Lack of quality and availability of suitable data. AI and ML algorithms rely on large amounts of high-quality data for successful model training and validation. There is a scarcity of data itself, and the data that is available is often in disparate formats and of insufficient quality to derive high quality predictions and models. These issues, coupled with a lack of suitable pre-processing tools for data and chemical structure cleaning and standardization can lead to the generation of inaccurate or misleading models and predictions, which in turn diminish the value of the tools for medicinal chemists.
  • Lack of interpretability and transparency of the underlying decision-making.  Many AI and ML models such as deep learning neural networks use opaque and difficult-to-interpret decision-making processes based on arcane ML techniques, so they come to be viewed (or denigrated) as black boxes. Typical practising medicinal chemists don’t know their Random Forest from their Gradient Boosted Trees, so they disdain black boxes and regard them with suspicion. Medicinal chemists who can’t understand and explain the scientific rationale for the underlying AI/ML decision-making processes won’t trust the tools, which will then be left as the sole domain of the computational chemists.
  • Difficulty integrating AI/ML tools into existing drug discovery workflows. If the AI/ML tools can’t be seamlessly integrated into existing drug discovery workflows and processes, and used as routine steps in discovery procedures, they will largely be ignored, and uptake will be hindered. Similarly, if the tools insist on the chemists learning a different and disconnected user interface, or they need access to separate files that may require reformatting or cleaning, they will tend to be unused by medicinal chemists and will remain within the computational chemistry group. 

Overcoming the Concerns

If organizations accept the premise that medicinal chemists who use AI/ML will be more effective than those that don’t, they will need help to overcome the objections described above to ensure widespread and effective uptake of the tools. 

Provision of high-quality data in sufficient volume for reliably predicting important, high-performance descriptors and models; access to interpretable and explainable underlying ML decision-making rationales; and user-friendly AI/ML-based tools integrated into drug discovery workflows and pipelines are all possible now, and a second paper in this series will discuss how to mitigate the concerns and achieve these objectives.

Read More