Table 2 Comparison on the drug response prediction performance of different data representations and prediction models.

From: Converting tabular data into images for deep learning with convolutional neural networks

Dataset

Prediction model

Data representation

R2

P-value

CTRP

LightGBM

Tabular data

0.825 (0.003)

8.19E−20

Random forest

0.786 (0.003)

5.97E−26

tDNN

0.834 (0.004)

7.90E−18

sDNN

0.832 (0.005)

1.09E−16

CNN

IGTD images

0.856 (0.003)

 

REFINED images

0.855 (0.003)

8.77E−01

DeepInsight images

0.846 (0.004)

7.02E−10

GDSC

LightGBM

Tabular data

0.718 (0.006)

2.06E−13

Random forest

0.682 (0.006)

4.53E−19

tDNN

0.734 (0.009)

1.79E−03

sDNN

0.723 (0.008)

6.04E−10

CNN

IGTD images

0.74 (0.006)

 

REFINED images

0.739 (0.007)

5.93E−01

DeepInsight images

0.731 (0.008)

2.96E−06

  1. In the R2 column, the number before parenthesis is the average R2 across 20 cross-validation trials, and the number in the parenthesis is the standard deviation. Bold indicates the highest average R2 obtained on each dataset. P-value is obtained via the two-tail pairwise t-test to compare the performance of CNNs trained on IGTD images with those of other combinations of prediction models and data representations.
Back to article page