Questions about the epitope prediction experiment

[bbjy]

Thank you for the good work. However, I have some questions about the article as follows:

1. How many times of augmentation when you make the epitope prediction on the EpiPred dataset?
2. Since I cannot get the result as the paper showed on EpiPred dataset, so would you please tell me "The results showed in the article is the best ones or the ones averaged over many times of results"?

Thanks again!

[FTD007]

Hi @bbjy
I am so sorry for the inconvenience.

1. since DBD and EpiPred is small dataset, the aug number is 50. Just in case it would be helpful for you, look at those SE3 network which is rotation invariant so augmentation is not necessary.
2. result in table 3 is the average of the EpiPred test set and there is no augmentation for test set just to be clear. The following figure is the best one, 75% one and the median one.
   Sorry again for the inconvenience.

[bbjy]

Hi @bbjy I am so sorry for the inconvenience.

1. since DBD and EpiPred is small dataset, the aug number is 50. Just in case it would be helpful for you, look at those SE3 network which is rotation invariant so augmentation is not necessary.
2. result in table 3 is the average of the EpiPred test set and there is no augmentation for test set just to be clear. The following figure is the best one, 75% one and the median one.
   Sorry again for the inconvenience.

Hi @FTD007. Thank you so much for your reply! I still have some confusion:

1. Do the "SE3 network" means "SE3-Nets: Learning Rigid Body Motion using Deep Neural Networks"? Or others?
2. If the augmentation is not necessary, why the experiment in the Table 3 "PInet [EpiPred & Aug]" is much better than "PInet [EpiPred]"?
3. And the result in table 3 is the average of the EpiPred test set for one time of experiment or the average of many times of experiments?
4. " The following figure" in your answer means the Figure 4 in the paper? Or which one?

Sorry for trouble you, but your work is really good and I want to understand it well. Thanks again!

[FTD007]

1. se3:
   https://arxiv.org/abs/2006.10503
   I believe this would be next improvement using point cloud input but I dont really have time to implement it in PInet recently.
2. as I mentioned above it is not in current PInet setup yet, so for PInet and PInet alpha the augmentation is necessary.
3. I might misunderstand some part but there is not multi times of experiments, let me know if some part is still not clear
4. fig 5

[bbjy]

1. se3:
   https://arxiv.org/abs/2006.10503
   I believe this would be next improvement using point cloud input but I dont really have time to implement it in PInet recently.
2. as I mentioned above it is not in current PInet setup yet, so for PInet and PInet alpha the augmentation is necessary.
3. I might misunderstand some part but there is not multi times of experiments, let me know if some part is still not clear
4. fig 5

Thank you for your quick kindly reply! I got it.
But the article I accessed (https://academic.oup.com/bioinformatics/article-abstract/37/17/2580/6162157?redirectedFrom=fulltext) doesn't present the Figure 5, although you metioned "Figure 5" in the article.