UFold icon indicating copy to clipboard operation
UFold copied to clipboard

Training datasets

Open andrewjjung47 opened this issue 2 years ago • 11 comments

Hello,

I am wondering if the training datasets used in the publication (RNAStralign, TR0, and fine-tuning dataset for bpRNA-new) are available? I was able to find the evaluation datasets from Google Drive (https://drive.google.com/drive/folders/143TJDgGqYOhumayiDx7w4XWLDVEOSZ2X) but do not see the training sets.

Also, I am wondering which datasets were used to train ufold_train.pt and ufold_train_alldata.pt?

Thank you so much in advance!

andrewjjung47 avatar May 20 '22 14:05 andrewjjung47

Hi there,

The training dataset we used in the publication can be retrieved from mxfold2 paper(RNAStralign and TR0), you can download that data from its repository, as for the fine-tuning dataset for bpRNA-new, we kept the fasta file after mutating and CD-Hit. You can download that from the google drive link.

As for the ufold_train.pt and ufold_train_alldata.pt, the difference between these two files is the training data. We use all the data including training and testing datasets to train the latter model for future prediction on other sequences. And we also kept the pre-trained model only for training (ufold_train.pt).

Thanks.

sperfu avatar May 24 '22 11:05 sperfu

Thank you for your response!

andrewjjung47 avatar Jun 07 '22 07:06 andrewjjung47

Hello there, I have another quick question regarding exactly which datasets were used to train ufold_train.pt.

Was it trained only on TR0 or RNAStralign & TR0? Also, was the provided ufold_train.pt also trained on the fine-tuning dataset for bpRNA-new?

andrewjjung47 avatar Jun 07 '22 08:06 andrewjjung47

Hello there, I have another quick question regarding exactly which datasets were used to train ufold_train.pt.

Was it trained only on TR0 or RNAStralign & TR0? Also, was the provided ufold_train.pt also trained on the fine-tuning dataset for bpRNA-new?

Same question

CatIIIIIIII avatar Oct 19 '22 15:10 CatIIIIIIII

Hello there, I have another quick question regarding exactly which datasets were used to train ufold_train.pt.

Was it trained only on TR0 or RNAStralign & TR0? Also, was the provided ufold_train.pt also trained on the fine-tuning dataset for bpRNA-new?

I doubt that the author mistoke testing data for training since I can only reach up to 0.5 F1s with TR0 for training and TS0 for testing under same settings with this repo

CatIIIIIIII avatar Oct 19 '22 15:10 CatIIIIIIII

Hi there,

As illustrated in our paper, we trained the model on TR0 & RNAStralign & augmented mutated dataset for ufold_train.pt. TS0 is only used for training(the dataset we used is the same as MXfold2 and SPOTRNA). So we don't think we mistoke testing data for training. And ufold_train.pt is not trained on the fine-tuning dataset for bpRNA-new dataset, it is designed for fine-tuning for PDB dataset.

Thanks

sperfu avatar Oct 20 '22 08:10 sperfu

Hi there,

As illustrated in our paper, we trained the model on TR0 & RNAStralign & augmented mutated dataset for ufold_train.pt. TS0 is only used for training(the dataset we used is the same as MXfold2 and SPOTRNA). So we don't think we mistoke testing data for training. And ufold_train.pt is not trained on the fine-tuning dataset for bpRNA-new dataset, it is designed for fine-tuning for PDB dataset.

Thanks

Hello,

Thank you for your response. Maybe I need to set training epoches larger for better results.

CatIIIIIIII avatar Oct 21 '22 14:10 CatIIIIIIII

Hi there,

As illustrated in our paper, we trained the model on TR0 & RNAStralign & augmented mutated dataset for ufold_train.pt. TS0 is only used for training(the dataset we used is the same as MXfold2 and SPOTRNA). So we don't think we mistoke testing data for training. And ufold_train.pt is not trained on the fine-tuning dataset for bpRNA-new dataset, it is designed for fine-tuning for PDB dataset.

Thanks

Hi,

Could you share your augmented mutated dataset pickle file?

CatIIIIIIII avatar Oct 28 '22 04:10 CatIIIIIIII

Hi all @CatIIIIIIII @sperfu @andrewjjung47 !

I would like to ask if it's possible to share with us the mutated sequence of bpRNA-new that you generated (synthetic/augmented data).

That would be very helpful!

Thanks in advance. N.

black0017 avatar Jun 01 '23 11:06 black0017

@CatIIIIIIII @sperfu @andrewjjung47

The PDB train set (669 structures), as detailed in the supplementary material, would be very helpful if you could indicate which file/folder it corresponds to.

image

black0017 avatar Jun 01 '23 11:06 black0017

@CatIIIIIIII @sperfu @andrewjjung47 Update:

PDB train data

hi, I have found the 669 PDB files here: https://bprna.cgrb.oregonstate.edu/index.html Nonetheless, the .fasta files are just the sequences, there are no contact maps.

Mutated data

merge_allbpnew_mutate.cdhit.fa probably contains the mutated sequences but the targets (contact maps) are again missing.

black0017 avatar Jun 09 '23 08:06 black0017