A Classification Problem using the Molgrid library - Protein-Ligand Dataset
Budget: $30 – $250 USD
A classification project of protein-ligand binding (pocket) sites. The attention-based CNN model is used.
***The model and training part were written by me, and they work without error***. My problem is that my model cannot learn. In a 5-class problem, the model makes fix predictions.
For example,
Ground Truth: tensor([0, 1, 2, 3, 4, 0, 1, 2, 3, 4]
but model's prediction is
Predicted tensor([4, 4, 4, 4, 4, 4, 4, 4, 4, 4] - REALLY BAD!
Gradient clipping, different optimization algorithms, changing hyperparameters, etc. methods were applied, but the problem was not solved.
Those who have experience in Bioinformatic and have worked with protein files before are preferred!
See the example in the "READ ME" file in the link below. If you don't understand this example, you don't understand my problem either. Please don't waste each other's time if you don't have experience.
https://github.com/gnina/libmolgrid
Expectation: Reaching at least 75% of training accuracy and testing the model with the test dataset I will share.
***The model and training part were written by me, and they work without error***. My problem is that my model cannot learn. In a 5-class problem, the model makes fix predictions.
For example,
Ground Truth: tensor([0, 1, 2, 3, 4, 0, 1, 2, 3, 4]
but model's prediction is
Predicted tensor([4, 4, 4, 4, 4, 4, 4, 4, 4, 4] - REALLY BAD!
Gradient clipping, different optimization algorithms, changing hyperparameters, etc. methods were applied, but the problem was not solved.
Those who have experience in Bioinformatic and have worked with protein files before are preferred!
See the example in the "READ ME" file in the link below. If you don't understand this example, you don't understand my problem either. Please don't waste each other's time if you don't have experience.
https://github.com/gnina/libmolgrid
Expectation: Reaching at least 75% of training accuracy and testing the model with the test dataset I will share.