Tutorial 3.5 Prediction parameters

Introduction

LIT-AlphaFold allows to easily modify many of the parameters used for structure prediction during calculation. In this tutorial all the available prediction parameters are described.

Paramters

The parameters listed below are used both for monomer and multimer prediction unless it is specified otherwise.

python $LITAF/run_multimer_jobs.py --mode custom \
--output_path test_structure \
--monomer_objects_dir test_features_mmseqs2 \
--input_file monomer_simple.in \
--data_dir $AFDATA \
--num_cycle_mono=5 \
--max_seqs=8 \
--max_extra_seqs=16 \
--allow_resume=True \
--num_predictions_per_model=1 \
--dropout=True \
--cluster_profile=True \
--save_all=False \
--model_to_relax=all \

Number of cycles

The arguments num_cycle_mono and num_cycle_multi describe the number of recycle steps used during calculation. Increasing the number of recycles can lead to better structures.

For multimer predictions there is an early stopping conditions, which interrupts the calculation once the structure has converged to a minima.

Number of sequences

Number of MSA sequences and MSA extra sequences used during calculations, it overrides the default values. Subsampling of the MSA has been used to obtain a more diverse conformational ensembles.

Allow resume

Allows AlphaFold to continue the calculations from the latest saved result in case a previous calculation failed.

Number of prediction per model

AlphaFold and AlphaFold-Multimer are formed by five models, each one generating a different structure. Currently it is suggested for monomers to perform calculations once, while for multimers it is suggested to repeat them five times.

Multiple articles have shown that repeating calculations a large number of times can be beneficial, both for conformational sampling, and to obtain better multimer structures.

Dropout

Dropout is a technique used for the training of deep learning networks. It conists in randomly dropping some of the nodes to force the model to learn redundant solutions for the objective function. By applying dropout at inference time the network weights are perturbed, leading to more diverse outputs. This approach has been applied to obtain a better conformational sampling for multimer structure prediction.

Cluster profile

See description in ColabFold

Save all

AlphaFold and AlphaFold-Multimer generate a large number of outputs. To save space by default certain results are not saved to disk. By enabling --save_all additional information are saved: distogram, masked_msa, experimentally resolved weights.

Model to relax

Indicates for which model minimization should be performed.

Conclusion

In this tutorial you have learned how to change parameters for structure prediction using LIT-AF.