Introduction

Small-angle X-ray scattering (SAXS) experiments are widely used for the characterization of biological macromolecules in solution. SAXS patterns contain information on the size and shape of dissolved particles in nanometer resolution. Here we propose a novel method for primary SAXS data analysis based on the application of artificial neural networks (NN). Trained on synthetic SAXS data, the feedforward neural networks are able to reliably predict molecular weight and maximum intraparticle distance (Dmax) directly from the experimental data. The method is applicable to data from monodisperse solutions of folded proteins, intrinsically disordered proteins, and nucleic acids. Extensive tests on synthetic SAXS data generated in various angular ranges with varying levels of noise demonstrated a higher accuracy and better robustness of the NN approach compared to the existing methods.

How to cite

If you use gnnom for your research please cite: D.Molodenskiy, D.Svergun, A.Kikhney (2022) Artificial neural networks for solution scattering data analysis, Structure DOI

How to apply NN

To apply the NN you need to use a command similar to this:

python apply_nn.py p mw /path/to/my/datafile 1.0 2.8 --units=nanometer --n=10000

Here p stands for type of particles (proteins); mw - what parameter you want to predict (mw or dmax); 1.0 and 2.8 are the I(0) and Rg values, accordingly; units are the angular units (inverse angstroms or inverse nanometers) and n is the number of repetitions for resampling. The latter parameter is usually set to about 10 000. The following command

python apply_nn.py -h

will print the help menu and exit.

How to train a new NN

To train a NN you need the training, validation and test sets. To run the script you can use e.g. the following command

python makemodel_scalar.py /path/to/data /path/to/crysol/logfiles 500 mw 

Here /path/to/data implies the structure similar to this one. The root directory should contain the subfolders with different simulated concentrations used for training and the list of these folders is hard-coded into the makemodel_scalar.py script as the "folders" variable. If you are using your own training set, please change this variable in the code. The next parameter is /path/to/crysol/logfiles - the path to the CRYSOL log files; 500 is the number of epochs for training; and mw is the desired parameter. It is also possible to re-use old weights if you want to train your model up using previously generated *.h5 file. You can also use a pickle file to quickly re-read the input data.
The command

python makemodel_scalar.py -h

will print the help menu and exit.

Training/validation/test sets

The original data sets were generated from PDB, PED and NDB databases that were used for training of the NNs and benchmarking it against other methods. The data are available here

Web service

If you don't want to modify the code and just want to try it out as is on your SAXS data, this approach is also implemented as a web service. Please make sure that you select the proper macromolecule type and (in case you are using WAXS data) the angular units!