Unsupervised ECNet

[eric-jm-lang]

Hello,
In the ECNet paper, you built an unsupervised ECNet model that does not require DMS data for training. It uses the predicted probability of an amino acid at a position as a proxy for fitness.
Is there a specific code for this unsupervised model? Or is it a question of using the current ECNet code to generate an unsupervised model by using a different input for --train?
Could you please provide more details on how to build such an unsupervised model?
Many thanks in advance

[NOforgetQY]

+1

[TernencezzZ]

I have the same confusion!

[meehljd]

Same here!

[meehljd]

Maybe one could hack a training file with a bunch of neutral mutations

mutation    score
M1M         1.0
F12F;L30L  1.0
G89G         1.0

[meehljd]

Looks like it worked. Tested with separate test file with random mutations. Need to still validate with experimental data.

Prediction from Training File with neutral mutations:

mutation score prediction
D36D;G142G 1.00000000 1.04933691
E145E;S128S 1.00000000 1.04933691
L19L;N152N 1.00000000 1.04933691
E237E;P12P 1.00000000 1.04933691

Prediction from Test File with random mutations:

mutation score prediction
A9D;T27L 1.00000000 0.51061106
A124A;I3T 1.00000000 0.98425829
V258L;A211L 1.00000000 -0.28957328
A276R;K252E 1.00000000 1.15801334
E175E;F14A 1.00000000 1.18147123

[meehljd]

Ignore my previous naive attempt. I re-read the paper and recalled @luoyunan used homologous sequences to train a bidirectional model on masked amino acid residues. I reviewed the ECNet and Dataset classes. The provided model can only process mutation-feature paired TSV files for training. Training on homologous sequences must be in a different code base.

[eric-jm-lang]

Thank you for your input @meehljd. Hope @luoyunan can provide more information on how to do this.