MolJam is presented as an open-source, multidimensional framework to provide quantitative assessment of molecular dataset quality and to clarify how dataset quality relates to downstream machine learning performance in cheminformatics and bioinformatics. The authors emphasize that dataset quality is rarely assessed systematically and that the relationship between dataset properties and model outcomes remains poorly understood. MolJam is designed to standardize diagnosis of dataset limitations and to enable quantitative evaluation of curation decisions.
MolJam assesses dataset quality across five explicit dimensions: structural integrity, data quality, experimental information quality, chemical space coverage, and data distribution. Within these dimensions the framework reports 12 standardized metrics (the source lists 12 metrics but does not enumerate them in this summary). The approach aims to capture both representation-level issues (for example, molecular structure completeness) and dataset-level properties (for example, coverage and label distributions) to provide a multidimensional view of dataset fitness for machine learning tasks.
The authors applied MolJam to a collection of benchmark datasets: 11 datasets from MoleculeNet and eight datasets derived from ChEMBL. This systematic application revealed widespread and heterogeneous quality issues across commonly used molecular benchmarks. The reporting emphasizes comparing benchmark quality and uncovering dataset-specific limitations that may influence reproducibility and model comparisons.
Using MolJam, the study identified several recurring problems:
These problems were heterogeneous across datasets: not every benchmark showed the same issues or severity, highlighting the need for dataset-specific diagnosis rather than assuming uniform quality across widely used collections.
To test whether improving MolJam quality metrics yields better model performance, the authors refined two datasets: ESOL and Lipophilicity. Refinement increased the MolJam quality scores for these datasets. However, the changes produced mixed effects on predictive performance: improvements in dataset cleanliness did not uniformly translate to better model metrics. The authors attribute this to a competing influence of reduced dataset size after curation — that is, removing lower-quality entries can reduce training data and thus offset gains from improved data quality.
The study included controlled ablation experiments to dissect the relative contributions of dataset quality and data quantity to model performance. Results demonstrated that both factors matter: neither maximum cleanliness nor maximum size alone guarantees optimal predictive performance. In a notable finding, retaining molecules with incomplete stereochemical information sometimes outperformed removing them, when the additional data quantity compensated for the loss in structural fidelity. This observation supports the conclusion that curation decisions require trade-off assessment rather than one-size-fits-all rules.
The authors argue that molecular dataset curation cannot be reduced to maximizing data cleanliness. Instead, curation must balance multiple dimensions of data quality against information loss due to removal or modification of entries. MolJam is proposed as a standardized diagnostic and comparative tool to:
These capabilities are intended to inform reproducible model development and fair benchmark comparisons by making dataset strengths and weaknesses explicit.
MolJam is described as an open-source framework. The preprint reporting this work was posted August 25, 2026. The authors declared no competing interests. For correspondence, the source lists rz24{at}columbia.edu. The preprint provides the framework and the dataset analyses as the basis for the conclusions summarized here.