· 4 min read · Gaia Lab
4,400 protein complexes with Boltz-2: predicting how two proteins bind
An open diffusion model, from the AlphaFold 3 family, applied to thousands of protein pairs to predict the structure of the complex and the confidence of the binding. Fifteen jobs, two campaigns and a lesson on what each pair costs when the complexes grow.
Fourth instalment of Cluster X-ray and the line with the fewest jobs: fifteen. Also one of those that reserves the most GPU, because each job runs for days.
The question #
Knowing whether two proteins interact, and how, is one of the central questions of structural biology: it explains signalling pathways, drug targets and the effect of mutations. Determining it experimentally costs months per pair. The line uses structure prediction to estimate, at the scale of thousands of pairs, the structure of the complex that two proteins form and the confidence that the binding is real.
How it is approached #
The model is Boltz-2, an open diffusion model for biomolecular structure, successor to Boltz-1 and comparable to AlphaFold 3. The workflow is the same in both campaigns:
- Resolve the sequences of each pair (from a local cache, a shared FASTA or live UniProt) and write one description per unique physical pair.
- Predict each complex with Boltz-2, using the project’s remote multiple-sequence-alignment server and generating three diffusion samples per pair to estimate variability.
- Build the results table per row of the experimental design, with the model’s confidence scores for each pair.
The first campaign covered 2,534 pairs; the second, 1,907, of which some 1,650 were new. About 4,400 predictions in total. The experimental design is generated separately and does not change inside the jobs: pairs already predicted are skipped, so a campaign can be resumed as many times as needed without losing anything.
What is learned #
Besides the table of complexes and confidences, the line left behind a useful measurement for anyone planning a campaign like this: what a pair costs. The first estimate, written before measuring, assumed two minutes per pair on a 48 GB card. The observed rate over the first 2,533 pairs was 4 to 5 minutes (median 240 s, mean 304 s). And in the tail of the second campaign, on reaching the largest complexes, the mean rose to 21 minutes per pair. A factor of ten between the initial hypothesis and the reality of the large complexes, which the resumable design absorbed without losing any work.
With a single sequence of predictions per card and the alignments computed elsewhere, the GPU spends a good part of its time waiting. Computing the alignments locally or predicting several pairs in parallel per card are the two obvious levers if the line grows.
On the cluster #
In the fifth instalment , LiDAR point clouds and individual trees.
Figures from Slurm accounting (reserved capacity, not measured usage) and the archived sbatch files. Anonymised post: no identifiable users, paths, emails or project names. Quotations are from the script comments.