Computational tools offer a rapid, cost-saving advantage to toxicologists assessing the hazard of chemicals. The predictivity of a model for a group of structures is one aspect to be considered in a computational assessment. However, given the universe of chemicals, there are structural classes which a model will predict with a higher level of reliability than others. There may also be chemicals which are not confidently reflected in the model’s chemical space and these chemicals are typically not in the model domain (more on this topic in a later blog). It is important to assess the reliability of the model’s prediction as part of a computational assessment. The expert review is the method that allows the toxicologist to gain confidence in an assessment. I like to think of this process as being analogous to the review of experimental data through controls, statistical analyses, and the determination of false negative or positive results.

The degree to which a user can interact with the computational platform and understand how an assessment was made relates to the level of review which could be performed practically. Access to descriptors that are used in the prediction and the ability to support the use of the descriptors through analysis of the underlying data which substantiate the descriptor use is important. Parameters such as the diversity of the training set examples, the extent that any structural descriptors can be linked to a mechanism or whether there are other structural moieties that explain the activity of the training set examples are used to evaluate the assessment.

In addition to the above, here are a couple of items that I like to evaluate as part of a computational assessment.

  • Are there any potentially reactive features that are not considered by the model? This provides added information to support negative predictions in cases where the structural features of a statistical model does not consider the entire structure. In Figure 1, the entire structure is considered as a feature. This feature maps to two examples which are assessed as negative.
Figure 1. An evaluation of LS-167087 for potentially reactive features

Figure 2 shows features not considered in the analysis of LS-181651. An analysis of potentially reactive features considers the 1,3,5-triazine,2-phenyl- feature. This feature mapped to 4 examples, which are all negative for sensitization hazard. One of these examples (LS-181621) is a close chemical analog. Such reviews support a negative prediction.

Figure 2. An evaluation of LS-181651 for potentially reactive features
  • Is an alerting fragment represented in a known negative example structure and how does the chemical environment of the alerting fragment compare to the target structure? Figure 3 shows an aromatic nitro indeterminate alert which matched the target structure. A search for analogs showed that the indeterminate alert is also present in LS-188180, a known negative.  The system performs a comparison of the target and analog structures and indicates that the alerting sub-structure is within the same chemical environment in both structures. Such a review supports a negative assessment for the target structure, which is also assessed as negative by a statistical model.
Figure 3. Assessment of an analog to support an expert-rule based prediction

Expert reviews give added reliability to an assessment. Transparent models and platforms facilitate such reviews and mitigate any black-box concerns around in silico tool use.

Please send me a note at cjohnson@leadscope.com if you would like to discuss in more detail.