andysenner/aptamergnn-modulea

AptamerGNN Module A

Structure-conditioned generation of candidate aptamer sequences from protein–nucleic-acid co-crystal structures.

What it does

You upload a PDB containing a protein bound to a nucleic acid (DNA or RNA). The pipeline:

1. 1.Identifies the binding pocket — protein atoms within 8 Å of the bound oligo
2. 2.Encodes pocket geometry — graph neural network produces per-atom embeddings
3. 3.Generates candidate sequences — transformer decoder samples sequences conditioned on pocket
4. 4.Folds and ranks — ViennaRNA secondary-structure prediction with G-quadruplex motif detection
5. 5.Visualizes — 2D structure rendering for each candidate

What it's for

• —Computational triage of aptamer candidates for targets where structural data exists
• —Generating starting points for SELEX library design
• —Exploring how a known target's pocket constrains aptamer sequence space

What it isn't

• —A replacement for SELEX. Module A produces candidates, not guaranteed binders.
• —Suitable for bare apo proteins. The model needs a bound oligo to identify the pocket. Future versions will integrate fpocket / P2Rank for apo-protein support.
• —An affinity predictor. For predicted Kd/pKd of a specific protein-aptamer pair, use the AptamerGNN V1 Space.

Methodology

Module A combines:

• —A 33k-parameter pretrained graph encoder (GATv2 over heterogeneous protein-nucleic-acid graphs)
• —A 1M-parameter Transformer decoder with cross-attention to the encoder
• —A 30-token vocabulary covering canonical DNA/RNA bases and common modified nucleotides

Validation metrics (test set, 58 held-out pockets):

• —Reconstruction accuracy: 0.66
• —Cross-target specificity ratio: 1.63 (between-pocket vs within-pocket sequence distance)
• —Thrombin-motif specificity: 21.6× (anchor vs non-anchor pockets)
• —Exact reproduction of TBA1 (GGTTGGTGTGGTTGG) from thrombin pockets: 8 / 20 trials

Full methodology and validation in the Module A paper.

Companion tool

AptamerGNN V1 predicts binding affinity (pKd) for a known protein-aptamer pair. Workflow: generate candidates here, model the chosen one in 3D via AlphaFold Server, then score the predicted complex with V1.

Built with

• —PyTorch Geometric
• —BioPython
• —ViennaRNA
• —forgi
• —Gradio

Andrew Senner, Subcellular Products LLC, 2026.