Table 2. Scaffold diversity and novelty. Metrics were measured after fine-tuning on bioactive molecules for three targets (PPAR, PIM1, and JAK2) using 10 and 100 molecules selected with (a) high similarity and (b) low similarity. Scaffold diversity and novelty relative to the fine-tuning sets (FT) and pre-training sets (PT) are reported as the mean ± standard deviation for 10 fine-tuning molecules. For each experimental setup and each metric, the best and second best values are reported in boldface and italics, respectively.

Similarity	Target	Augmentation	10 fine-tuning molecules			100 fine-tuning molecules
			Scaffold diversity	Scaffold novelty (FT)	Scaffold novelty (PT)	Scaffold diversity	Scaffold novelty (FT)	Scaffold novelty (PT)
High	PPARδ	Enumeration	67 ± 2	67 ± 2	35 ± 1	60 ± 1	56 ± 1	42 ± 1
		Token deletion (random)	78 ± 1	78 ± 1	46 ± 1	70 ± 2	62 ± 2	41.8 ± 0.9
		Token deletion (validity)	77.0 ± 0.9	76.3 ± 0.9	45 ± 1	52 ± 2	48 ± 2	35 ± 2
		Token deletion (protected)	82 ± 1	81 ± 1	46 ± 3	64.5 ± 0.8	56.3 ± 0.8	37 ± 3
		Atom masking (random)	77 ± 3	75 ± 2	44 ± 1	61.3 ± 0.8	53 ± 1	35 ± 2
		Atom masking (funct. Group)	81 ± 1	80 ± 1	45 ± 1	63 ± 4	55 ± 4	35 ± 2
		Bioisosteric substitution	80 ± 1	80 ± 1	51 ± 3	54 ± 2	51 ± 2	35.2 ± 0.3
		Self-training	83 ± 1	83 ± 1	40.1 ± 0.9	55.3 ± 0.9	53.8 ± 0.9	25 ± 1
		No augmentation	93 ± 1	93 ± 1	52 ± 3	77 ± 1	75 ± 1	44 ± 1
	PIM1	Enumeration	93.6 ± 0.3	93.2 ± 0.2	58.2 ± 0.8	91 ± 2	88 ± 2	77 ± 1
		Token deletion (random)	94.1 ± 0.1	93.6 ± 0.1	59.9 ± 0.1	92 ± 2	85 ± 2	75 ± 2
		Token deletion (validity)	93.4 ± 0.9	92.9 ± 1.0	57 ± 2	83 ± 2	77 ± 2	71 ± 2
		Token deletion (protected)	94.4 ± 0.2	93.9 ± 0.1	60 ± 2	92 ± 2	86 ± 2	74.5 ± 0.8
		Atom masking (random)	90.5 ± 0.4	89.8 ± 0.5	56 ± 2	77.1 ± 0.9	68.3 ± 0.4	57 ± 2
		Atom masking (funct. Group)	86.1 ± 0.4	85.3 ± 0.4	51 ± 1	80.1 ± 0.6	72.6 ± 0.9	55 ± 2
		Bioisosteric substitution	92.8 ± 0.8	92.5 ± 0.8	53.5 ± 1.0	83 ± 2	81 ± 2	56.7 ± 0.5
		Self-training	86.5 ± 0.8	86.5 ± 0.8	45 ± 1	85.7 ± 0.4	84.0 ± 0.5	48 ± 1
		No augmentation	94.7 ± 0.7	94.5 ± 0.7	54.6 ± 0.6	93 ± 2	91 ± 1	56.5 ± 0.7
	JAK2	Enumeration	93.2 ± 0.4	93.2 ± 0.4	63.8 ± 0.2	87.1 ± 0.8	85 ± 1	73 ± 2
		Token deletion (random)	96.8 ± 0.9	96.3 ± 0.9	70 ± 2	91 ± 1	87.2 ± 0.4	72 ± 3
		Token deletion (validity)	95.8 ± 0.5	95.1 ± 0.4	75 ± 2	78.2 ± 0.7	76.3 ± 0.7	68 ± 1
		Token deletion (protected)	97.3 ± 0.4	96.7 ± 0.4	76.1 ± 0.6	89 ± 2	85 ± 2	72 ± 4
		Atom masking (random)	93 ± 1	92 ± 1	68 ± 1	76.8 ± 0.7	72 ± 1	55.5 ± 0.8
		Atom masking (funct. Group)	93.8 ± 0.6	92.7 ± 0.7	68 ± 2	81.0 ± 0.3	77.7 ± 0.9	60 ± 2
		Bioisosteric substitution	92.8 ± 0.2	92.0 ± 0.2	67.5 ± 0.2	75.1 ± 1.0	71.4 ± 1.0	59 ± 1
		Self-training	87 ± 1	87 ± 1	49 ± 2	86.0 ± 0.4	85.2 ± 0.3	51 ± 1
		No augmentation	94.8 ± 0.4	94.8 ± 0.4	58.4 ± 0.3	93 ± 1	91 ± 1	61 ± 1
Low	PPARδ	Enumeration	76.9 ± 0.3	76.6 ± 0.3	40.1 ± 0.8	86.1 ± 0.6	84.7 ± 0.8	58 ± 2
		Token deletion (random)	85.6 ± 1.0	85.0 ± 1.0	47 ± 2	81.2 ± 0.7	77 ± 1	42 ± 1
		Token deletion (validity)	91.0 ± 0.9	90.5 ± 1.0	48 ± 3	84.5 ± 0.5	82.4 ± 0.4	46.1 ± 0.9
		Token deletion (protected)	91.5 ± 0.6	91.3 ± 0.6	52 ± 2	86 ± 1	84 ± 1	46 ± 1
		Atom masking (random)	90.0 ± 0.6	89.7 ± 0.6	49 ± 1	83 ± 3	80 ± 2	42 ± 2
		Atom masking (funct. Group)	83 ± 2	82 ± 2	44 ± 2	82.3 ± 0.8	79.4 ± 0.9	44 ± 1
		Bioisosteric substitution	90.0 ± 1.0	89.7 ± 0.9	51 ± 2	91 ± 1	89 ± 1	55.0 ± 0.8
		Self-training	89 ± 1	89 ± 1	46.2 ± 0.4	71 ± 1	71 ± 1	33 ± 1
		No augmentation	94 ± 2	94 ± 2	54 ± 2	88.9 ± 0.8	88.0 ± 0.9	46 ± 3
	PIM1	Enumeration	90.0 ± 0.3	89.8 ± 0.2	47.6 ± 0.7	93.5 ± 0.3	91.6 ± 0.5	63 ± 1
		Token deletion (random)	94.6 ± 0.6	93.9 ± 0.6	55.6 ± 0.7	92.6 ± 0.1	90.0 ± 0.4	54.0 ± 1.0
		Token deletion (validity)	94.9 ± 0.2	94.8 ± 0.2	56 ± 1	94.9 ± 0.5	93.0 ± 0.7	55 ± 2
		Token deletion (protected)	95.5 ± 0.4	95.4 ± 0.4	57.4 ± 0.5	94.5 ± 0.4	92.5 ± 0.7	55 ± 1
		Atom masking (random)	96.0 ± 0.4	95.6 ± 0.3	54 ± 2	93.9 ± 0.3	91.8 ± 0.7	52.4 ± 0.9
		Atom masking (funct. Group)	94.7 ± 0.3	94.2 ± 0.3	55 ± 1	91.5 ± 0.3	88.3 ± 0.2	49.6 ± 0.4
		Bioisosteric substitution	95.2 ± 0.6	95.0 ± 0.5	55 ± 2	94.4 ± 0.6	93.4 ± 0.6	55 ± 1
		Self-training	88.6 ± 0.5	88.6 ± 0.5	46.5 ± 0.9	87.9 ± 0.9	87.5 ± 0.8	42.2 ± 0.4
		No augmentation	94.8 ± 0.9	94.7 ± 0.8	54 ± 2	94.5 ± 0.7	93.8 ± 0.6	53 ± 2
	JAK2	Enumeration	83 ± 1	83 ± 1	39.8 ± 0.3	94.3 ± 0.9	93 ± 1	66 ± 1
		Token deletion (random)	91.9 ± 0.5	91.3 ± 0.4	51 ± 1	95 ± 1	93 ± 2	59 ± 1
		Token deletion (validity)	93.1 ± 0.6	92.8 ± 0.6	51.8 ± 0.4	95.0 ± 0.8	93.5 ± 1.0	56 ± 2
		Token deletion (protected)	91.3 ± 0.7	90.6 ± 0.6	51 ± 1	96.7 ± 0.9	96 ± 1	59 ± 1
		Atom masking (random)	90.8 ± 0.4	90.3 ± 0.4	52.3 ± 0.9	93.5 ± 0.4	91.1 ± 0.5	55 ± 2
		Atom masking (funct. Group)	92.9 ± 0.3	92.4 ± 0.2	54 ± 2	94.1 ± 0.2	92.2 ± 0.5	56 ± 2
		Bioisosteric substitution	94.5 ± 0.2	94.3 ± 0.1	54.3 ± 0.9	94.7 ± 0.6	94.2 ± 0.7	57.5 ± 0.9
		Self-training	87.6 ± 0.7	87.5 ± 0.6	44.8 ± 0.7	87 ± 1	86.6 ± 0.9	41.4 ± 0.9
		No augmentation	94 ± 1	94 ± 1	53 ± 1	96 ± 1	96 ± 1	53.3 ± 0.3