Extended Data Table 2:

Hyperparameters of the GCNs of the first method (the Chemprop models).

Parameter	Antibiotic Model	Toxicity Model
Batch size	20	15
Depth	5	6
Dropout	0	0.3
ffn Hidden size	1200	1700
ffn num layers	3	3
Final learning rate	5.40 × 10−6	9.81 × 10−6
Hidden Size	1200	1700
Init learning rate	0.00011	4.98 × 10−7
Max learning rate	0.00015	3.78 × 10−5
Warmup epochs	14	1.30 × 101
Mode	regression	classification