Drug discovery project teams must make many decisions, from choosing the best target to selection of appropriate compounds in hit finding, hit-to-lead, lead optimization, and nomination of a preclinical candidate. Making good decisions at each stage is essential to the successful outcome of a drug discovery project.
Poor choice of target or compound can lead to wasted effort due to unnecessary synthesis and screening or, worse, late-stage failure after incurring high cost. Conversely, overaggressive filtering of pipelines can lead to missed opportunities to find new therapies.
Making good decisions in drug discovery is an enormous challenge. Early screening and the widespread use of predictive modeling have dramatically increased the amount of compound-related data that is available from the earliest stages of drug discovery. Optimization of a compound requires many properties to be balanced simultaneously. This is made even more difficult by the fact that all of the sources of data have significant uncertainty.
Human beings are notoriously poor at making decisions based on complex, uncertain data, particularly where there is a lot at stake. For this reason, decision-support software tools are often used to assist with the decision-making process. However, as this tutorial will discuss, they are not necessarily sufficient to deal with all of the challenges.
Software platforms such as Optibrium’s StarDrop™ go beyond visualization to guide decisions and help scientists to objectively assess all of the available data, focusing attention on a good set of options for detailed consideration.
Decision-support tools include laboratory information management systems, databases, data-processing systems, and data-visualization packages. These collect, aggregate, and process data and, ultimately, provide engaging visualizations that help to analyze data and present findings to colleagues. Clearly, these are essential capabilities. But, are they sufficient to gain the most value from the data and drive effective decisions?
Common visualization approaches include data tables with “traffic lights” to indicate good, intermediate, or poor results (Figure 1A). But, this view is complicated when dealing with large numbers of properties and, if an ideal (all green) molecule is not present, it is difficult to select molecules visually; for example, is it better to have one red property value or three yellow?
A multidimensional plot of the same dataset is shown in Figure 1B. While visually appealing, these plots rapidly become complex and difficult to interpret when dealing with many properties. It is difficult to quickly identify high-quality compounds that meet the success criteria across multiple properties and showing the uncertainties in the data (e.g., using error bars) quickly makes the plot unreadable.
When faced with such a complicated picture, judgments tend to be made by gut instinct. This makes a consistent, objective assessment of the data difficult to achieve and hence visualization is often, in practice, used to support a decision that has already been made, rather than to drive the decision-making process itself. Gut instinct can be a useful guide, but psychologists have identified unconscious biases in decision making to which humans are all subject.
In this context, the most relevant of these is “confirmation bias,” which is the human tendency to look for evidence that confirms rather than refutes initial judgment. Confirmation bias can lead to missed opportunities, as choices may be narrowed too quickly or, conversely, wasted effort as projects are failed too late.