โจทย์ ปรนัย
What is the primary application of machine learning (ML) in the field of drug discovery?

Machine learning and artificial intelligence have to ability to predict bioactivity and physical property of compounds proving very useful to discovering new drugs.

The predictions of bioactivity and physical properties are some of the most important applications of machine learning (ML) and artificial intelligence (AI) in drug discovery.

โจทย์ ปรนัย
What is QSAR in the context of drug discovery?

An abbreviation for Quantitative Structure-Activity and Property Relationships

This field is broadly known as quantitative structure-activity and property relationships (QSAR, QSPR) and is a necessary component of numerous drug discovery projects

โจทย์ ปรนัย
Why is data preparation critical in the model life cycle of ML for drug discovery?

Although pharmaceutical industry is becoming progressively more active in fundamental ML and AI research, the focus on the final model application remains the central pillar.

ML is used to make better decisions faster and to accelerate the design-make-test-analyze (DMTA) cycle of novel molecular entities [5]. Although pharmaceutical industry is becoming progressively more active in fundamental ML and AI research, the focus on the final model application remains the central pillar. Academic studies typically focus on pushing the boundaries of ML in drug discovery, e.g., by borrowing inspiration from other fields such as natural language processing (NLP) [8] or geometric deep learning (DL) [9].

โจทย์ ปรนัย
What is a major challenge when using public data sets in drug discovery ML models?

Public data set are generally smaller than in-house data set in pharmaceutical companies. In-house data set always contain more compound measurement than public data set. Because of this there is need to combine lots of public data together to increase data size. Merging data sources implies major efforts and bears the risk of biases, redundancies, and error accumulation, and poses challenges both in academic and industrial settings.

To increase data set size, public data are generally pooled from multiple sources [29]), which in turn increases heterogeneity. Merging data sources implies major efforts and bears the risk of biases, redundancies, and error accumulation [30], and poses challenges both in academic and industrial settings. In industry, assay protocols are standardized and typically include multiple measurements per compound, giving rise to more homogeneous and consistent data sets. However, bringing diverse data sources together might also constitute a major effort due to legacy systems, change of protocols over time or different conventions in annotations (e.g. units or molecule identifiers). Given these challenges with experimental data, curation and homogenization are crucial steps for successful ML applications [11]. Discussions with experimentalists may help detecting outliers, properly combining data from diverse sources, or defining other modelling aspects. Moreover, having replicates to analyze experimental variability and error across the measurement range also provides information about the maximum accuracy that a ML model can achieve and facilitates interpretation of model outputs [31], [32].

โจทย์ ปรนัย
How does time-split data validation benefit the predictive performance of ML models in drug discovery?

Discussions with experimentalists may help detecting outliers, properly combining data from diverse sources, or defining other modelling aspects. Moreover, having replicates to analyze experimental variability and error across the measurement range also provides information about the maximum accuracy that a ML model can achieve and facilitates interpretation of model outputs

To increase data set size, public data are generally pooled from multiple sources [29]), which in turn increases heterogeneity. Merging data sources implies major efforts and bears the risk of biases, redundancies, and error accumulation [30], and poses challenges both in academic and industrial settings. In industry, assay protocols are standardized and typically include multiple measurements per compound, giving rise to more homogeneous and consistent data sets. However, bringing diverse data sources together might also constitute a major effort due to legacy systems, change of protocols over time or different conventions in annotations (e.g. units or molecule identifiers). Given these challenges with experimental data, curation and homogenization are crucial steps for successful ML applications

โจทย์ ปรนัย
Which aspect is NOT a direct benefit of AI as perceived by dermatologists for improving the melanoma diagnosis process?

AI can diagnose with high consistency and accuracy because of the image processing function

We decided to choose one single area in medicine in order to achieve high consistency. Dermatological diagnosis is a particularly suitable area of study as it makes use of image processing aspect of AI, which is particularly well-developed. Specifically, we focus on the process of diagnosing melanoma; this provides a useful basis for comparing the participants' accounts.

โจทย์ ปรนัย
What is the primary reason dermatologists want scientific proof of AI's validity for diagnosis?

The medical experts wanted to see scientific proof to see the AI validity and how it function before implementing them.

The other aspect of trusting AI is also something we expected: explainability. However, our interviewees did not think about explainability in a trivial way. Before a widespread routine implementation of AI, these medical experts want to see scientific proof of its validity, and they all wanted to get a broad range of detailed information about the design, operation, learning, and adaptive capabilities of AI in their domain

โจทย์ ปรนัย
Which of the following is identified as a challenge introduced by AI in the diagnostic process?

AI and ML lacks the ability in sense-making, context consideration and decision making under uncertainty meaning that there must be a human working with an AI to incorporate the AI into the industry.

However, such ML faces challenges in explainability including sense-making, consideration of context, and decision-making under uncertainty, making it necessary to incorporate human expertise for usable intelligence.

โจทย์ ปรนัย
How do dermatologists perceive the final responsibility for a medical diagnosis when AI is used?

The responsibility still lies with the physician, however AI may be accountable for its own action and decision in the near future.

Most of the interviewed physicians expressed a positive attitude towards an AI in medicine, but every single one of them confirmed that, at the end of the day, it is the physician who must take responsibility and make the final decision about a diagnosis, based on a value judgment. Only one participant speculated that perhaps AI will be able to take responsibility someday, but the rest firmly rejected even a remote future possibility:

โจทย์ ปรนัย
What is essential for dermatologists to effectively utilize AI in diagnosing melanoma?

A number of dermatologist were interviewed on how participants could use AI to diagnose melanoma mostly questioning about decision making with the influence of AI.

In total 17 dermatologists were interviewed in two rounds; nine in the first and eight in the second. An outline interview protocol was set up for the first round of interviews, focusing on how the participants use or could use AI-generated predictions when diagnosing melanoma, and how that would influence their decision-making process (judgment) about melanoma. The second round of interviews commenced five months later, following the analysis of the interviews from the first round, therefore the interview protocol, albeit loosely, centered around In total 17 dermatologists were interviewed in two rounds; nine in the first and eight in the second. An outline interview protocol was set up for the first round of interviews, focusing on how the participants use or could use AI-generated predictions when diagnosing melanoma, and how that would influence their decision-making process (judgment) about melanoma. The second round of interviews commenced five months later, following the analysis of the interviews from the first round, therefore the interview protocol, albeit loosely, centered around the initial themes. In this second round we probed what we learned from the first round, aiming for high consistency, digging deeper trying to unpack further richness.

โจทย์ ปรนัย
What factor is crucial for the accuracy of a machine learning model in predicting drug efficacy?

A leaning model is named so because it requires input of information in order to predict outcome and give consistent and accurate prediction. It cannot do so if the is of low quality or if the data is irrelevant.

To increase data set size, public data are generally pooled from multiple sources [29]), which in turn increases heterogeneity. Merging data sources implies major efforts and bears the risk of biases, redundancies, and error accumulation [30], and poses challenges both in academic and industrial settings. In industry, assay protocols are standardized and typically include multiple measurements per compound, giving rise to more homogeneous and consistent data sets. However, bringing diverse data sources together might also constitute a major effort due to legacy systems, change of protocols over time or different conventions in annotations (e.g. units or molecule identifiers). Given these challenges with experimental data, curation and homogenization are crucial steps for successful ML applications [11].

โจทย์ ปรนัย
How does collaboration between academia and industry contribute to advancements in drug discovery using machine learning?

Naturally there are lots of overlap between academic and industry in each of the field. There are interesting thought on how academia and industry can work together. As both have different specialization but with the same goal, if collaboration is done can be highly beneficial as they could combine there diverse expertise and resources to accelerate discoveries.

Industry and academia have been considered in our analysis up to this point as two separated worlds. Naturally, there is a lot of research overlap and both parties benefit from one another. In the following discussion about academia-industry collaborations, we focus on two central aspects: (i) how do we handle reproducibility of publications and the FAIR guidelines when working with proprietary data, and (ii) how can collaborations, including researchers’ education, be funded. Interesting thoughts on how academia and industry can work together in drug discovery in general are also provided in the review by Tralau-Stewart and co-workers [114].

โจทย์ ปรนัย
What is a significant challenge in deploying machine learning models for drug discovery in a real-world scenario?

As the real world already has people performing jobs; doctors diagnosing, pharmacist predicting outcome of drugs, and many others. To implement AI means that people has to either be relocated into different function in the system or has to have to change some aspect of their work.

AI has to fill in spot. There are currently human working those spot. If AI is implemented AI will fill those spot. Human loses spot.

โจทย์ ปรนัย
Why is model validation using a time-split approach considered effective in the context of drug discovery?

Splits are essentially evaluation of model's performance in a scenario. The evaluation mimics how ML will perform in practice, however this typically gives optimistic view of the prospective performance while this is preferred the only way to really know how it will function in the real world is the use it in the real world.

Random splits typically give a too optimistic view of the prospective performance of a model [70], [71], [72], while time splits have emerged as a preferred approach in industry to evaluate the prospective model performance [71]. Such evaluation mimics how a ML model will be used in practice, i.e. to predict compounds that have not been synthesized or measured (new chemical matter under exploration).

โจทย์ ปรนัย
In the context of drug discovery, what is the primary benefit of using machine learning models that can quantify prediction uncertainty?

As AI has been inputted useful information it can use those information to predict data and outcome better than a human can. Helping to divert risks and assist in decision making. Moreover it can learn from past mistakes as it is called Machine learning after all.

The predictions of bioactivity and physical properties are some of the most important applications of machine learning (ML) and artificial intelligence (AI) in drug discovery. This field is broadly known as quantitative structure-activity and property relationships (QSAR, QSPR) and is a necessary component of numerous drug discovery projects (for an overview of QSAR and its history see Tyrchan et al. [1]).

โจทย์ ปรนัย
Considering the ethical implications of AI in medical diagnosis, which factor is most crucial for ensuring responsible AI use?

People are afraid of the unknown by making the knowledge on how AI function more tangible and transparent to the public will lower the distress among the population and overtime gain trust from the majority it all goes well.

In the paper there are various interview on how human needs scientific proof on how the AI work in order to trust it to diagnose patients and assist them in decision making.

โจทย์ ปรนัย
In the context of human-AI collaboration in medical diagnosis, what is the primary challenge in integrating AI into clinical practice according to the article?

AI has to work on their explainability including sense-making, consideration of context, and decision-making under uncertainty, making it necessary to incorporate human expertise for usable intelligence.

Today's ML relying on statistical learning algorithms, large datasets, and available computational capacity [108], which should, in principle, enable evidence-based decision-making [109] across various domains by replicating statistical frequencies from previous data and improving it based on new data [110]. However, such ML faces challenges in explainability including sense-making, consideration of context, and decision-making under uncertainty, making it necessary to incorporate human expertise for usable intelligence [[111], [112], [113]].

โจทย์ ปรนัย
What do dermatologists view as the potential benefit of AI in improving melanoma diagnosis?

According to participants A interview he states: “I can see its clear benefit, that compared to a human, AI can handle big volume of data, and if it could scan and analyze the whole body of a patient and point out that might have a risk for melanoma, that could be a great support and save time for us, physicians.” And another participant indicate that AI could be used as pre-screening to save time and money as he state: “AI could point out those that differ from the rules and may bring any risk of melanoma. With that, it could save time for the dermatologist and money for the patient.”

12 participants were very keen on using AI in their diagnostic work. They emphasized data processing power and speed of AI, like Participant B, saying: “I can see its clear benefit, that compared to a human, AI can handle big volume of data, and if it could scan and analyze the whole body of a patient and point out that might have a risk for melanoma, that could be a great support and save time for us, physicians.” Time saving for physicians was a leitmotif, one suggestion was that AI could provide a kind of pre-screening: “AI could point out those that differ from the rules and may bring any risk of melanoma. With that, it could save time for the dermatologist and money for the patient.”

โจทย์ ปรนัย
Why is scientific proof of AI's validity crucial for its acceptance among dermatologists?

They, just like any human being would not trust a machine that they do not know the function, thinking process and capability just like you wouldn't trust a predatorial animal as you don't know their full capability and thought. The researcher says that if he at least know how it is designed and who designed it he would be able to trust in the machine more.

Before a widespread routine implementation of AI, these medical experts want to see scientific proof of its validity, and they all wanted to get a broad range of detailed information about the design, operation, learning, and adaptive capabilities of AI in their domain: “I doubt I could trust entirely and would use 100% of what the AI proposes, but if I knew how the AI tool has been designed and who did participate in the design, that could increase my trust.”

โจทย์ ปรนัย
Regarding the final responsibility for medical diagnoses, how do dermatologists perceive their role when AI is used?

Presently, the physician has to be responsible of the AI decision and action. In this case physician want to take responsibility as they believe that this is the right thing to do considering they are the one allowing for AI to assist them.

However, in our case, it seems that it can be simply resolved: medical doctors want to take responsibility, and based on our data, we believe that this is not because they are worried about their jobs; they genuinely believe that this is the right thing to do. Additional implications of the concept of responsibility relate to AI design, specifically collaborative AI design; we address this in the final part of the discussion.