Construct custom model using DreaMS

[WinterWei22]

Hi, Thanks for the great work!

I am interested in constructing a custom model based on DreaMS, but some details are confusing to me. For example, for the input of the DreaMS encoder, do we need to involve the pre-cursor peaks for the input? Because I noticed that you use the pre-cursor token in the ssl model:

class CustomModel(nn.Module):
     ...
    def forward(self, x):

        **x = self.spec_encoder(x)[:, 0, :]  # [:, 0, :] to get the precursor peak token embedding**

        x = F.sigmoid(self.lin_out(x))  # Example forward pass through the head

       return x

If the precursor peaks are needed, how can I provide them in the right way?

Thanks in advance for your help!

[roman-bushuiev]

Hi @WinterWei22 ,

Thank you for your question. We need to improve the documentation in this part.

Input x in the forward is expected to have shape (batch_size, num_peaks + 1, 2) where + 1 stands for the precursor peak and 2 stands for m/z and intensity of each peak. So, the precursor peak is indeed used and is contained within the input tensor. It corresponds to the first element across the second dimension and is prepended to the tensor of MS/MS peaks by SpectrumPreprocessor.