Table 1. Comparison of Loss, Cosine Similarity, and Mean Average Precision (mAP) from Neural Networks Trained with Different Chemical Descriptorsa.
| model | classification levels | loss (SD) | cosine similarity (SD) | mAP (SD) |
|---|---|---|---|---|
| MFs (binary) | Pathway | 0.0197 (0.0004) | 0.9863 (0.0004) | 0.9932 (0.0003)** |
| Superclass | 0.0050 (0.0000) | 0.9642 (0.0003) | 0.9423 (0.0010) | |
| Class | 0.0012 (0.0000) | 0.9314 (0.0002) | 0.8734 (0.0018) | |
| CMFs (Integer) | Pathway | 0.0211 (0.0016) | 0.9849 (0.0013) | 0.9920 (0.0004) |
| Superclass | 0.0046 (0.0001)** | 0.9682 (0.0009)** | 0.9515 (0.0030)** | |
| Class | 0.0010 (0.0000)*** | 0.9377 (0.0005)*** | 0.8951 (0.0022)*** |
a
Each model was optimized based on results from the validation set (n = 11777) and evaluated by using the test set (n = 14721). The results are the average values from five runs of each model. There was no significant difference in the pathway loss or cosine similarity between the two models, so neither is bolded. *Significant at p < 0.05; **significant at p < 0.005; ***significant at p < 0.001. SD = standard deviation. MFs = Morgan fingerprints. CMFs = counted Morgan fingerprints.