A chemical-screening model can look convincing on familiar structures yet struggle with a new collection of compounds. Alaa M. Elsayad and Khaled A. Elsayad tested their ready-biodegradability workflow on an external group of 670 compounds and reported 86.6% accuracy. The result comes with a useful second measure: it correctly identified 75.4% of the readily biodegradable substances.[1]
The workflow classified substances as readily biodegradable or not readily biodegradable, using 837 training records and a 218-record internal test before the external evaluation. The researchers compared generalised linear models, random forests, gradient boosting and neural networks. Each predicted a defined biodegradability endpoint from molecular structure, rather than observing how a substance broke down in the environment.[1]
Turning molecular structure into predictors
The approach is known as a quantitative structure–activity relationship, or QSAR. A molecular structure is translated into numbers describing such features as surface properties, connectivity and local atomic patterns. The team used RDKit to calculate 119 conventional descriptors and 1,024 Morgan fingerprint bits; filtering reduced the final feature set to 311. H2O AutoML and KNIME provided modelling and workflow components.[1]
The authors report recalculating the gradient-boosting results from deposited external predictions. The area under the receiver-operating-characteristic curve was 0.911, with a paired-bootstrap 95% interval of 0.886–0.935. This measures how well scores distinguish the two classes across thresholds. Accuracy and recall describe performance at the selected classification threshold.[1]
Gradient boosting on the external validation set
| Measure | Reported result |
|---|---|
| External compounds | 670 |
| ROC–AUC | 0.911 |
| 95% bootstrap interval for AUC | 0.886–0.935 |
| Accuracy | 86.6% |
| Recall for readily biodegradable compounds | 75.4% |
| Matthews correlation coefficient | 0.668 |
Recall of 75.4% means that about one-quarter of the truly readily biodegradable compounds in this sample were missed. That could cause a screening programme to overlook useful candidates. The opposite error—labelling a non-readily-biodegradable substance as readily biodegradable—raises a different concern. Choosing a threshold therefore requires knowing how the prediction will be used, rather than optimising a single headline percentage.[1]
What an explanation adds
The authors examined feature importance with several approaches, including permutation-based methods and local explanations. Some signals were shared across methods; others changed rank depending on how importance was measured. A fingerprint bit encodes a structural pattern, so its importance alone does not establish a chemical mechanism.[1]
The authors also assess the model’s applicability domain: how closely a proposed chemical resembles the structures on which the workflow was developed. That matters when moving from this external dataset to different chemical families. The immediate use is to prioritise experimental attention. A structure-based prediction does not itself measure degradation in a laboratory assay or establish what a substance will do in a particular environment.[1]
Sources & further reading
- Externally validated and explainable QSAR for ready biodegradability: KNIME–H2O AutoML with RDKit descriptors and Morgan fingerprints Scientific Reports · 25 Oct 2026 · Accessed 26 Oct 2026