Predicting Chemical Biodegradability Using Molecular Structure Data Through Machine Learning Classification

Whether a chemical breaks down in the environment or stays there for decades is, in the end, a fact about its molecular structure. The practical difficulty is that establishing it experimentally is slow and expensive: a ready biodegradability assay runs for 28 days per compound. If structure determines fate, then descriptors computed from structure alone ought to carry enough information to predict that fate in advance — which is the premise of Quantitative Structure–Activity Relationship modelling.

This project tests that premise on 1055 chemicals. Each is described by 41 numerical descriptors — electronic, topological and steric properties derived from the molecular graph — and labelled by experiment as readily biodegradable or not. The work is a classification problem, but the more interesting part is the exploratory analysis that precedes it: finding out which descriptors are carrying real information and which are simply restating each other.

§Research question

Can molecular descriptors computed from structure alone classify a chemical as readily or not readily biodegradable, and which descriptors carry that information?

§Theoretical background

A QSAR model is a mapping from the structural and physicochemical properties of a compound to a property or activity of interest. Developing one follows a fixed sequence: assemble compounds with known outcomes, compute molecular descriptors, fit a statistical or machine-learning model relating descriptors to outcome, validate it on data held out from fitting, interpret which features matter, and only then apply it to new molecules.

The outcome here is binary. Under the OECD convention a substance counts as readily biodegradable (RB) when it reaches roughly 60% biodegradation within 28 days; anything slower or requiring special conditions is not readily biodegradable (NRB). So the task is binary classification: given a descriptor vector $\mathbf{d} \in \mathbb{R}^{41}$, predict a label in $\{\mathrm{RB}, \mathrm{NRB}\}$.

The descriptors themselves are a mixture of counts and graph-spectral quantities — the number of heavy atoms, the percentage of carbon atoms, the frequency of C–N pairs at a given topological distance, alongside eigenvalue quantities such as SpMax_L (leading eigenvalue of the Laplace matrix) and SM6_L (spectral moment of order 6). Several are therefore different functions of the same molecular graph, which matters when their correlations are examined.

§Methodology

Data. Experimental biodegradation values for 1055 chemicals were taken from the QSAR biodegradation dataset in the UC Irvine Machine Learning Repository: 41 molecular descriptors and one experimental class per compound.

Exploratory data analysis. Descriptive statistics — minimum, maximum, mean and standard deviation — were computed for every variable. Univariate histograms were examined across all 41 descriptors and the class variable, showing a mixture of approximately normal, right-skewed, multimodal and sparse distributions. A correlation matrix was computed over all descriptors and then over a focused subset of 15, and pairwise relationships were visualised with seaborn pair plots and joint regression plots, separated by class.

Modelling. The data were partitioned 60% training, 20% validation and 20% testing. Five classifiers were then fitted and compared: logistic regression, a decision tree, a support vector machine, a random forest, and k-nearest neighbours.

Tools. Python for the modelling and plotting, with Google Sheets and Excel used alongside for tabulation.

§Results and observations

Descriptor correlations. The strongest relationship found was between SpMax_A and SM6_L at $r = 0.9226$ — the leading eigenvalue of the adjacency matrix against the sixth spectral moment of the Laplace matrix, which is to say two summaries of the same graph spectrum tracking each other closely. A further strong pair was SM6_B(m) and SpMax_B(m) at $r = 0.9174$, again two functions of one Burden matrix. The most negative relationship reported was between SpPosA_B(p) and Mi at $r = -0.7870$.

Classification. Of the five models tested, the support vector machine was the most accurate at 86.7%, with random forest close behind at 86.3%.

Model	Reported accuracy
Support vector machine	86.7%
Random forest	86.3%
Logistic regression	see note below
Decision tree	not reported individually
k-nearest neighbours	not reported individually

The manuscript’s overall conclusion is a methodological one rather than a performance claim: that a QSAR model’s usefulness depends on the quality of the training data and on whether the chosen descriptors actually capture the chemistry that governs the outcome.

§Challenges and limitations

• Redundant descriptors. The highest correlations are between pairs that are different spectral summaries of the same molecular graph. Highly collinear features inflate apparent information and destabilise coefficient estimates in linear models.
• Accuracy alone is the wrong measure. Only accuracy is reported. On a task where the two classes are not balanced, accuracy can be high while the minority class is badly handled; precision, recall, F1 and a confusion matrix per class would be needed to know whether the models are useful.
• A validation split that is not used. Twenty percent of the data was set aside for validation, but no hyperparameter selection or model-selection procedure using it is described.
• No feature selection was carried out. The correlation analysis identifies redundancy but the models appear to have been fitted on the full descriptor set regardless.
• Dataset-bounded conclusions. Every result is a statement about this one curated dataset. Generalisation to chemicals outside its applicability domain is not established, and the manuscript notes the OECD compound data as a larger resource not yet used.