Analytical Approaches for Organic Compound Properties

The report of Leistra on reported versus estimated boiling points and vapor pressures for pesticides provides an interesting case study in mathematical underpinnings for the phenomena investigated. Reports of chemical properties are available from many online databases such as ChemSpider, PubChem, and others. But these aggregate other sources of data to get chemical properties. In some cases, they are from a manufacturer. Of course, manufacturers often report values without a method or explanation of how it was determined.

The article makes a number of assumptions, the most important of which is that reported values are based on experimental observations. That’s a reasonable assumption here. But Leistra uses several estimation techniques to determine if numerical models for the boiling point or vapor pressure come close. Overall, he finds that they are at least of the correct order of magnitude. Coming from a numerical background, I find that order of magnitude to insufficient for many results. But, like everything, the quality of the result depends on the application.

Reported and estimated values can each have a place in the modeling of chemical properties. If a value correct within an order of magnitude is sufficient for the application of the data, then it does not matter which value is used. Applications requiring greater precision can use experimentation to find the practical values. Further, estimated values can be used to provide ballpark estimates which can then be verified through experimentation. When experimentation is not possible, the estimate needs to be sufficient. But these can cross-check each other.

What catches my attention the most is that Leistra includes several linear or log-linear models for estimating the vapor pressure. While linear models can be quite effective, some semi-statistical models such as artificial neural networks or random forests may be able to provide superior results for estimating chemical properties. The unfortunate downside is that chemists may not like the inability to understand what the model is doing. It seems I am not the first to think of this as the use of neural networks to estimate the boiling point of alkanes was investigated as far back as 1994. Though there are a great many alternative approaches that could be used.