In-silio Protein Design Pipeline

This repository contains the in-silico protein design and evaluation pipeline that we used for assessing Genie 2. We set this up separately from Genie 2 repository to facilitate assessments of different structure-based protein diffusion models. The pipeline consists of:

• assessment on designability through self-consistency pipeline
• assessment on secondary diversity through P-SEA algorithm
• assessment on tertiary diversity through hierarchical clustering
• assessment on novelty

Set up

Assume the environment has a cuda-compatiable PyTorch installed and Python <= 3.9. For example, on our own machine, the environment is created and initialized by running.

python3.9 -m venv insilico_pipeline_venv
source insilico_pipeline_venv/bin/activate
module load cuda11.8
pip install torch==2.0.1+cu118 torchvision==0.15.2+cu118 --index-url https://download.pytorch.org/whl/cu118

The setup process consists of three parts:

• Set up the pipeline package and additional packages (TMscore and TMalign) by running

  bash scripts/setup/setup.sh
  

• Set up an inverse folding model and its dependencies by running

  bash scripts/setup/inverse_folds/setup_[INVERSE_FOLD_MODEL_NAME].sh
  

  Our current pipeline supports ProteinMPNN (proteinmpnn) and we intend to extend this to include more inverse folding models.

• Set up a folding model and its dependencies by running

  bash scripts/setup/folds/setup_[FOLD_MODEL_NAME].sh
  

  Our current pipeline supports ESMFold (esmfold) and we intend to extend this to include more folding models.

Additional notes

When setting up the environment for ESMFold, we install OpenFold v1.0.1 to ensure compatibility. One known issue for this OpenFold installation is its compatibility with deepspeed. This raises AttributeError: module 'deepspeed.utils' has no attribute 'is_initialized' when running the pipeline and could be fixed by replacing all occurences of deepspeed.utils.is_initialized() with deepspeed.comm.comm.is_initialized().

Pipelines

Our design package consists of three separate pipelines:

• standard pipeline for assessing designability and secondary diversity
• diversity pipeline for assessing tertiary diversity
• novelty pipeline for assessing novelty with respect to a reference dataset.

Standard pipeline (pipeline/standard)

Evaluate a set of generated structures by running

python pipeline/standard/evaluate.py --version [VERSION] --rootdir [ROOTDIR]

Our standard pipeline currently supports evaluation of structures from unconditional generation (by setting version to unconditional) and motif scaffolding (scaffold). For both modes, we assume that the root directory contains a folder named pdbs, which contains the PDB files of generated structures to be evaluated. For motif scaffolding, we additionally assume that the root directory contains a folder named motif_pdbs, which contains the PDB files of the corresponding motif structures (with the same filename as the generated structure and residue index aligned). Note that for motif scaffolding, we also support evaluations of multiple problems at the same time. This means that the root directory could contain a list of subdirectories, each of which consists of a pdbs and motif_pdbs folder detailed above. When evaluating multiple motif scaffolding problems, our pipeline supports distribution of tasks across multiple GPUS by adding the following flags --num_devices [NUM_GPUS] --num_processes [NUM_GPUS].

Evaluation results are stored in the root directory, which contains:

• a directory named designs, where each PDB file stores the fold model predicted structure that is most similar to the corresponding generated structure;

• a csv file named info.csv, which contains evaluation statistics for the set of generated structures. Information on columns is summarized in the table below.

  Column	Description
  domain	Name of generated structure
  seqlen	Sequence length of generated structure
  scRMSD	RMSD between the generated structure and the most similar structure predicted by the specified fold model
  scTM	TM score between the generated structure and the most similar structure predicted by the specified fold model
  pLDDT	Local confidence from the specified fold model, averaged across all residues
  pAE	Confidence from the specified fold model in the relative position of two residues, averaged across all residue-residue pairs
  generated_pct_helix	Percentage of helices in the generated structure
  generated_pct_strand	Percentage of strands in the generated structure
  generated_pct_ss	Percentage of helices and strands in the generated structure
  generated_pct_left_helix	Percentage of left-handed helices in the generated structure
  designed_pct_helix	Percentage of helices in the most similar structure predicted by the specific fold model
  designed_pct_strand	Percentage of strands in the most similar structure predicted by the specific fold model
  designed_pct_ss	Percentage of helices and strands in the most similar structure predicted by the specific fold model
  designed_pct_left_helix	Percentage of left-handed helices in the most similar structure predicted by the specific fold model

Note that for secondary structure evaluations, we use the P-SEA algorithm, which allows us to predict secondary structures based on Ca atoms only.

Diversity pipeline (pipeline/diversity)

Assume that a set of generated structure is assessed by the above standard pipeline. Evaluate this set of generated structures on tertiary diversity by running

python pipeline/diversity/evaluate.py --rootdir [ROOTDIR] --num_cpus [NUM_CPUS]

Our default value of num_cpus is 1. We found this can be slow, so we recommend setting the value to the number of physical cores or the number of processes you want to run in parallel.

Results are stored by updating info.csv in the root directory to include

Column	Description
single_cluster_idx	Index of cluster that the generated structure belongs (hierarchically clustered via single linkage)
complete_cluster_idx	Index of cluster that the generated structure belongs (hierarchically clustered via complete linkage)
average_cluster_idx	Index of cluster that the generated structure belongs (hierarchically clustered via average linkage)

Note that for hierarchical clustering, we use TMalign to compute pairwise TM scores among all generated structures and a TM score threshold of 0.6 in the clustering process.

Novelty pipeline (pipeline/novelty)

Assume that a set of generated structure is assessed by the above standard pipeline. Evaluate this set of generated structures on novelty by running

python pipeline/novelty/evaluate.py --rootdir [ROOTDIR] --dataset [DATASET] --datadir [DATADIR] --num_cpus [NUM_CPUS]

where DATASET is the name of the reference dataset and DATADIR is the directory for the reference dataset (with each reference structure stored in a PDB format). Results are stored by updating info.csv in the root directory to include

Column	Description
max_[DATASET]_name	Name of structure in the dataset that is most similar to the generated structure
max_[DATASET]_tm	TM score between the generated structure and the most similar structure in the dataset

Examples

In the examples directory, we provide three examples (together with their correponding outputs) to demonstrate the input and output to our evaluation pipeline. Examples include:

• unconditional: evaluation of a unconditionally generated structure
• scaffold_single: evaluation of a conditionally generated structure, whose generation is conditioned on a single functional motif
• scaffold_multi: evaluation of a conditionally generated structure, whose generation is conditioned on multiple functional motifs

Profiling

Unconditional generation

Assume that the standard (designability) and diversity pipelines are run. To show the evaluation metrics on the set of generated structures, run

python scripts/analysis/profile_unconditional.py --rootdir [ROOTDIR]

This reports designability, diversity and F1 score on the set of generated structures. It also reports PDB novelty and/or AFDB novelty, provided that the corresponding novelty pipeline is run. Details on these evaluation metrics are found in the Genie 2 paper.

Motif scaffolding

Assume that the standard (designability) and diversity pipelines are run. To show the evaluation metrics on the set of generated structures, run

python scripts/analysis/profile_scaffold.py --rootdir [ROOTDIR]

This reports the number of solved motif scaffolding problems and the total number of unique clusters, aggregated across all problems. Details on these evaluation metrics are found in the Genie 2 paper. Here, we assume that the root directory contains a set of subdirectories, where each subdirectory starts with a prefix of motif= and contains inputs and outputs for a motif scaffolding problem (check out examples/scaffold_single and examples/scaffold_multi for detailed examples).