search

wyji001 / Point-Cloud

20 stars 6 forks source link

Question about limitations #2

Open linminhtoo opened 2 years ago

linminhtoo commented 2 years ago

Dear authors,

Interesting work. I also think point clouds & 3d geometries must be learnt in order to achieve the next tier of performance for protein property prediction (be it binding affinity or binding site prediction or others). I chanced across your work yesterday and according to the test results in your paper on the CASF 2016 core set, your PointTransformer model seems to perform very well relative to many other published methods.

Given such strong performance, I am interested to try this method out. Are there any limitations that we should take note of ? Is the training fast ?

wyji001 commented 2 years ago

First, much larger data sets are required to perform comparable to current best machine learning approaches. The performance of our single model in CASF2016 is similar to that of the convolutional neural network such as AK-Score or DeepAtom. The integrated model has some advantages in performance, but it still does not surpass the machine learning model using less data. Inference time will also increase dramatically. At the same time, although the point cloud method has a great advantage in the pre-processing speed, the inference speed of Pointtransformer is slightly slower than that of 3d convolutional neural network. At present, the point cloud based method is superior only in the prediction of larger proteins with ligand.

wyji001 commented 2 years ago

Compared with 3D convolutional neural network, more video memory（GPU） is required for training and prediction.Although the total parameters are lower.