biocheming / solubility_afdesign

Solubilty-aware AfDesign

This repository is based on the AfDesign binder hallucination protocol in ColabDesign, which is published by Dr. Sergey Ovchinnikov on GitHub.

Google Colab

setup

git clone https://github.com/sokrypton/af_backprop.git
pip -q install biopython dm-haiku==0.0.5 ml-collections py3Dmol
mkdir params
curl -fsSL https://storage.googleapis.com/alphafold/alphafold_params_2021-07-14.tar | tar x -C params
wget -qnc https://raw.githubusercontent.com/sokrypton/ColabFold/main/beta/colabfold.py
wget -qnc https://raw.githubusercontent.com/ohuelab/Solubility_AfDesign/main/design.py
wget -qnc https://raw.githubusercontent.com/ohuelab/Solubility_AfDesign/solubility/design_util.py
wget -qnc https://raw.githubusercontent.com/ohuelab/Solubility_AfDesign/solubility/constant.py

import numpy as np
from IPython.display import HTML
from design import mk_design_model, clear_mem
from design_util import *
import sys
sys.path.append('af_backprop')

binder hallucination

For a given protein target and protein binder length, generate/hallucinate a protein binder sequence AlphaFold thinks will bind to the target structure. To do this, we minimize PAE and maximize number of contacts at the interface and within the binder, and we maximize pLDDT of the binder.

model = mk_design_model(protocol="binder")
model.prep_inputs(pdb_filename="1YCR.pdb", chain="A", binder_len=13)
model.design_3stage(soft_iters=100, temp_iters=100, hard_iters=10)

FAQ

Can I reuse the same model without needing to recompile?

model.restart()

What are all the different design_??? methods?

• For design we provide 5 different functions:

  • design_logits() - optimize logits inputs (continious)
  • design_soft() - optimize softmax(logits) inputs (probabilities)
  • design_hard() - optimize one_hot(logits) inputs (discrete)

• For complex topologies, we find directly optimizing one_hot encoded sequence design_hard() to be very challenging. To get around this problem, we propose optimizing in 2 or 3 stages.

  • design_2stage() - soft → hard
  • design_3stage() - logits → soft → hard

What are all the different losses being optimized?

• general losses

  • pae - minimizes the predicted alignment error
  • plddt - maximizes the predicted LDDT
  • msa_ent - minimize entropy for MSA design (see example at the end of notebook)
  • pae and plddt values are between 0 and 1 (where lower is better for both)

• binder specific losses

  • pae_inter - minimize PAE interface of the proteins
  • pae_intra - minimize PAE within binder
  • con_inter - maximize number of contacts at the interface of the proteins
  • con_intra - maximize number of contacts within binder

How do I change the loss weights?

model.opt["weights"].update({"pae":0.0,"plddt":1.0})
model.opt["weights"]["pae"] = 0.0

WARNING: When setting weights be careful to use floats (instead of 1, use 1.0), otherwise this triggers recompile.

How do I disable or control dropout?

model.opt["dropout_scale"] = 1.0
model.design_???(dropout=True)

How do I control number of recycles used during design?

model = mk_design_model(num_recycles=1, recycle_mode="sample")

• num_recycles - max number of recycles to use during design (for denovo proteins we find 0 is often enough)
• recycle_model - When using more than 1 > num_recycles:
  • sample - at each iteration, randomly select number of recycles to use. (Recommended)
  • add_prev - add prediction logits (dgram, pae, plddt) across all recycles. (Most stable, but slow and requires more memory).
  • last - only use gradients from last recycle. (NOT recommended).

How do I control which model params are used during design?

By default all five models are used during optimization. If num_models > 1, then multiple params are evaluated at each iteration and the gradients/losses are averaged. Each iteration a random set of model params are used unless model_mode="fixed".

model = mk_design_model(num_models=1, model_mode="sample", model_parallel=False)

• num_models - number of model params to use at each iteration.
• model_mode:
  • sample - randomly select models params to use. (Recommended)
  • fixed - use the same model params each iteration.
• model_parallel - run model params in parallel if num_models > 1. By default, the model params are evaluated in serial, if you have access to high-end GPU, you can run all model params in parallel by enabling this flag.

How is contact defined? How do I change it?

con is defined using the distogram, where bins < 20 angstroms are summed. To change the cutoff use:

model.opt["con_cutoff"] = 8.0

For binder hallucination, can I specify the site I want to bind?

model.prep_inputs(...,hotspot="1-10,15,3")

How do I set the random seed for reproducibility?

model.restart(seed=0)

References

Kosugi T, Ohue M. Solubility-aware protein binding peptide design using AlphaFold. (under review)