We offer an alternative approach to toxicological risk assessment of new chemicals. We combine Operations Research techniques with those from Machine Learning to tackle the decision-making process. More specifically, we use Markov decision processes and Bayesian networks to derive the optimal cost-sensitive time-efficient Integrated Testing Strategies for chemical hazard classification under minimal expected cost in a mathematically rigorous fashion. We develop Bayesian networks which outperform state-of-the-art mechanistic causal models previously reported. More specifically, these models exhibit accuracy of 90% and sensitivity and specificity of 93% and 84%, respectively. Moreover, the inferred Bayesian networks are of considerably simpler structure as they comprise only the permeation coefficient, octanol/water coefficient, and TIMES software compared to their counterparts already in print, which comprise 15 descriptors. We use these simplified causal models to study the effect of varying misclassification costs on the nature of the optimal policy by means of sensitivity analysis. We note such analysis was previously computationally infeasible due to the fact that the variables which comprised the mechanistic model were categorical assuming a large number of possible values. We find that a variety of optimal policies can emerge subject to different misclassification costs assumed. Theoretical modeling framework developed is illustrated on the concrete example of hazard classification of skin allergens of previously unknown toxicological characteristics via integrating data obtained from in silico assays alone thus contributing to the literature of toxicological decision making based on nonanimal tests.