Table 2 F1-weighted and average accuracy scores of each MoA on the pre-training and holdout compound sets for different self-supervised models. On the pretraining set, downstream evaluation was performed on 175 compounds recovered as phenotypically active at 10 μM 20h by best performing model (PA + DINO + Barlow + XB).

Topoisomerase inhibitor

19

0.69

0.69

0.76

0.76

0.76

9

0.75

0.75

0.74

0.78

0.78

JAK inhibitor

19

0.30

0.29

0.45

0.43

0.45

10

0.29

0.47

0.56

0.59

0.67

Retinoid receptor agonist

17

0.19

0.27

0.88

0.82

0.85

9

0.33

0.50

0.88

0.88

0.88

HSP inhibitor

19

0.88

0.88

0.88

0.86

0.86

9

0.71

0.71

0.71

0.71

0.71

Protein synthesis inhibitor

14

0.32

0.32

0.30

0.30

0.29

6

0.40

0.20

0.50

0.33

0.50

HDAC inhibitor

25

0.77

0.77

0.94

0.94

0.94

7

0.44

0.44

0.60

0.60

0.60

Tubulin polymerization inhibitor

17

0.76

0.78

0.82

0.82

0.82

9

0.84

0.84

0.84

0.84

0.84

Aurora kinase inhibitor

18

0.53

0.56

0.74

0.74

0.80

7

0.33

0.38

0.67

0.63

0.63

ATPase inhibitor

12

0.25

0.27

0.25

0.35

0.32

7

0.00

0.22

0.46

0.57

0.53

PARP inhibitor

15

0.18

0.20

0.29

0.26

0.40

8

0.00

0.00

0.67

0.67

0.67

F1-weighted

0.52

0.53

0.67

0.66

0.68

0.43

0.47

0.67

0.67

0.69

Accuracy

0.45

0.46

0.62

0.62

0.64

0.37

0.41

0.62

0.62

0.64