GoKawiilDistributed computing
Most calculations in this study were performed on the Triton and Pegasus computing clusters at the Frost Institute for Data Science and Computing at the University of Miami. An overview of our computational pipeline is shown in Extended Data Fig. 3a, indicating the calculation phases that required distributed computing. Except for TM-alig13, we developed the algorithms for this study in our laboratory using Python, integrating structural informatics approaches from our previous work17,18,19,20,50. We submitted our compute jobs using the IBM load-sharing facility (LSF) platform and bsub commands wrapped in custom Python scripts.
Downloading the AF2 v.4 dataset
On 20 July 2023, we downloaded the AlphaFold database5 to our high-performance computing environment using the protocol provided by the Google-Deepmind GitHub page51. The download took approximately 1 week, and the compressed archive files (.tar) required 24 tebibytes (TiB) of disk space.
Preparing the AF2 database for distributed computing
The 24 TiB AF2 database contained 1,015,798 single-species archive files (.tar) of various sizes. Using a Python script, we redistributed these files into 607 equally sized 40-gibibyte (GiB) groupings of compressed mmCIF files (.cif.gz). We decompressed the mmCIF files in each 40-GiB grouping using a second Python script and distributed computing. Each 40-gigabyte (GB) grouping contained an average of 353,425 mmCIF files (.cif), occupying an average of 100-GiB on disk. The 607 uncompressed AF2 prediction files occupied collectively 60 TiB on disk. Using a Python script, we next equally distributed the 214,528,851 AF predictions of the 607 groupings into 1,000 job submission files.
Identifying superdark GPCR-like folds
As shown in Extended Data Fig. 3a, we performed an exhaustive search for 7TM folds in the 214,528,851 AF2 predictions using the prototypical GPCR rhodopsin (PDB 1F88, chain A)52 as our query structure and the program TM-align13. Using the 1,000 job submission files described above, we generated 1,000 tabular output files (TM-align option ‘-outfmt 2’), each containing pairwise alignment results for 214,528,851 AF predictions. We processed this set of 1,000 output files using a Python script, identifying 1,543,898 matches with at least 200 residues and a TM-align score of 0.5. Next, we identified and removed programmatically 82,419 matches corresponding to the obsolete UniProt database7 entries as of 1 August 2024, arriving at 1,461,479 AF2 structure matches. Because a TM-align score alone is not a robust metric for matching fold topologies, we had to develop additional structure-based algorithms to rank structural similarity and discern the correct 3D fold topology (Extended Data Fig. 3a).
Aligning hits
To generate and visualize structure-based alignments of related protein models, we developed a two-part Python workflow that automated large-scale pairwise superpositions using TM-align. The first script distributed alignment tasks across the Pegasus computing cluster by dividing the input FASTA file into evenly sized subsets (6,000 sequences per job), creating input directories, and submitting array jobs to the cluster scheduler. Each job invoked a second script, which parsed TM-align output files to extract TM-scores, root mean square deviation values and file paths for each query–target pair. For every pair, the script re-ran TM-align to generate the corresponding 3 × 3 rotation and translation matrices, applied these transformations to the atomic coordinates of the mobile structure and wrote the resulting aligned coordinates to new PDB files labelled by alignment score. This automated system enabled efficient, high-throughput structural superposition and normalization of AlphaFold-predicted models for subsequent comparative analysis and visualization.
... continue reading