Table 3 Results for the zero-shot image-to-image molecule classification task

From: CLOOME: contrastive learning unlocks bioimaging databases for queries with chemical structures

Split

Method

Accuracy [%]

  

Top-1

95% CI

Top-5

95% CI

Top-10

95% CI

Random

CLOOME

18.4

[18.1, 18.8]

41.5

[41.1, 42.0]

56.2

[55.8, 56.7]

 

GapNet

0.361

[0.307, 0.422]

1.07

[0.973, 1.17]

1.78

[1.66, 1.91]

 

CellProfiler

2.68

[2.53, 2.84]

8.24

[7.99, 8.51]

12.6

[12.3, 12.9]

Scaffold

CLOOME

22.3

[21.9, 22.9]

50.8

[50.2, 51.4]

67.1

[66.6, 67.7]

 

GapNet

0.517

[0.437, 0.608]

1.62

[1.47, 1.77]

2.56

[2.38, 2.75]

 

CellProfiler

3.64

[3.42, 3.86]

10.4

[10.0, 10.7]

15.6

[15.1, 16.0]

  1. Given a molecule-perturbed microscopy image, the image corresponding to the matched molecule must be selected from a set of candidates. Top-1, top-5, and top-10 accuracy in percentage are shown for a hold-out test set, along with the upper and lower limits for a 95% confidence interval (CI) (n = 43,778 for the random split and n = 28,248 for the scaffold split) on the resulting proportion. The best method in each category is marked in bold.
Back to article page