Frequently Asked Questions (FAQ)

Installation & Setup

Q: Do I absolutely need an NVIDIA GPU to run ProMaya? A: No. ProMaya can run on CPU-only machines. However, inference will be significantly slower, and we highly recommend an NVIDIA GPU (16GB+ VRAM) for large-scale screening or for training new models.

Q: The Conda environment resolution is taking forever. What can I do? A: Use mamba instead of conda, as it has a much faster dependency solver. If you run the Nextflow pipeline with -profile conda, it automatically uses Mamba behind the scenes. Alternatively, you can use our pre-built Docker/Singularity images to bypass local environment building entirely.

Execution & Nextflow

Q: Why does the pipeline re-run tasks I already completed when I use -resume? A: Nextflow's -resume relies on caching task inputs. If any input to a task changes (e.g., you modify config.yaml, update a PDB file, or change the command-line arguments like --threshold), Nextflow invalidates the cache and re-runs the task. Ensure all inputs remain identical to utilize the cache.

Q: I get a Process requirement exceeds available CPUs error. A: The pipeline requests specific CPU/Memory resources per task (e.g., 4 CPUs and 24GB RAM for PREDICT_PPI). If your local machine or cluster node has fewer resources, you must override these requests in a custom nextflow.config file.

Q: Why are no PyMOL images generated in results/predictions/.../structure/? A: Generating rendered PNGs requires PyMOL to be installed and available in the execution environment's PATH. If PyMOL is missing, ProMaya gracefully degrades: it will skip the PNG generation but will still output the .pml script. You can open this .pml script manually in your local PyMOL installation.

Model Interpretability

Q: What does --gradcam true actually do? A: It enables the computation of Graph Grad-CAM attributions. This involves performing a backward pass through the network to determine which atoms, residues, and surface points most heavily influenced the model's prediction. It generates TSV data and visualizations (heatmaps, PyMOL scripts) for these attributions.

Q: Why is the pipeline so much slower when --gradcam true is enabled? A: Computing attributions requires a backward pass (gradient computation), which is computationally expensive. Additionally, generating the 2D heatmaps and 3D structure visualizations adds significant I/O and processing overhead. We recommend running a fast initial screen with --gradcam false, and then re-running only the high-probability hits with --gradcam true.