Table 1 The performance of four different generators with different number of neurons in hidden layers for pre-training and fine-tuning processes
From: DrugEx v3: scaffold-constrained drug design with graph transformer-based reinforcement learning
Methods | Hidden Neurons | Pre-trained Model | Fine-tuned Model | ||||||
|---|---|---|---|---|---|---|---|---|---|
Validity | Accuracy | Novelty | Uniqueness | Validity | Accuracy | Novelty | Uniqueness | ||
Graph Transformer | 512 | 100.0% | 99.3% | 99.9% | 99.4% | 100.0% | 99.2% | 68.9% | 82.9% |
Sequential Transformer | 128 | 91.8% | 62.4% | 90.2% | 92.5% | 94.5% | 80.5% | 8.6% | 24.3% |
256 | 94.2% | 69.3% | 89.3% | 91.4% | 98.8% | 89.5% | 9.2% | 26.6% | |
512 | 96.7% | 74.0% | 89.1% | 91.8% | 99.3% | 92.7% | 8.9% | 28.9% | |
1024 | 97.1% | 77.9% | 89.5% | 91.4% | 99.4% | 94.3% | 8.2% | 32.9% | |
LSTM-BASE | 128 | 87.1% | 38.7% | 83.2% | 84.0% | 85.2% | 53.1% | 9.9% | 26.8% |
256 | 91.4% | 48.8% | 89.0% | 91.2% | 94.5% | 75.8% | 5.8% | 21.2% | |
512 | 93.9% | 52.4% | 84.3% | 89.1% | 98.7% | 81.6% | 3.9% | 19.2% | |
1024 | 95.7% | 57.0% | 79.6% | 87.5% | 99.6% | 90.2% | 2.1% | 18.1% | |
LSTM + ATTN | 128 | 89.8% | 57.0% | 84.2% | 85.0% | 85.2% | 64.8% | 14.2% | 27.8% |
256 | 92.6% | 68.4% | 87.1% | 89.5% | 94.9% | 80.5% | 8.9% | 22.4% | |
512 | 94.3% | 72.8% | 85.3% | 89.7& | 96.9% | 85.9% | 6.3% | 20.7% | |
1024 | 96.0% | 75.0% | 80.7% | 89.4% | 99.1% | 92.9% | 4.2% | 20.2% | |