[ADDA]Gene expression -> ATAC-seq (gene score / bin)

[marvinquiet]

• Dataset: 941 features selected on 10X GE, then 10X GE -> AML PBMC_D11T1 gene score (original prediction by Seurat: Acc: 0.647 ARI: 0.409 MacroF1: 0.486, by MLP: Acc: 0.543 ARI: 0.169 MacroF1: 0.410)
• Tao has tried the pytorch implemented version but it seems having some problem in the author's implementation. I thus tried the original tensorflow version published by the author.

Experiments	Accuracy	ARI	MacroF1
Training Evaluation (no ADDA)	0.9997	0.9989	0.9998
Test Evaluation (no ADDA)	0.4473	0.0906	0.3454
Training Evaluation (ADDA, 2000 epoch)	0.9303	0.8352	0.9199
Test Evaluation (ADDA, 2000 epoch)	0.6232	0.3785	0.4336

Check discriminator performance by distinguishing between gene expression and gene score.

[marvinquiet]

Check input before discriminator

I used epoch 2000 to train again and get Acc: 0.619, ARI: 0.381, MacroF1: 0.386. There do seem to be a lot of performance variations as Tao mentioned. Overall, training 2,000 epochs seem to achieve the best performance.

Epoch 0	Epoch 500

Epoch 1000	Epoch 1500

Final Epoch: