DDMuffin

Missense mutations in drug targets are a leading cause of therapeutic failure: one substitution in a kinase active site can abolish inhibitor binding and drive resistance. Predicting how far a mutation shifts binding free energy (ΔΔG) has historically been handled by methods separate from those predicting absolute protein-ligand affinity. DDMuffin, posted as a preprint in July 2025 by the Biosig group at the University of Queensland and the Baker Heart and Diabetes Institute, tackles both tasks inside one framework.

Given a protein-ligand complex structure, it predicts either the wild-type binding affinity (as pKi) or the ΔΔG produced by a point mutation, fusing three molecular graphs — a residue graph annotated with protein language model embeddings, a ligand atom graph annotated with pharmacophore types, and an explicit interaction graph — through graph convolutional networks. Transfer learning links the two tasks: pre-training on a large corpus of complexes with measured affinities, then fine-tuning on a much smaller mutation-specific set.

Evaluation rigor is central. Because widely used affinity benchmarks share similar complexes across training and test splits, inflating accuracy, DDMuffin is trained and assessed on the leak-proof LP-PDBBind partitioning and validated on external sets curated to exclude that overlap. Unlike structure-free predictors such as AQAffinity, it requires a complex structure and returns residue- and atom-level attributions with each prediction.

#Key Features

• Absolute and relative affinity in one model: The same network predicts a compound's baseline binding strength and how that strength shifts under a target mutation.
• Three-graph molecular representation: Residue nodes carry 1,024-dimensional ProtT5 embeddings, ligand atoms carry pharmacophoric types, and an interaction graph encodes 14 classes of interatomic contact from Arpeggio, weighted by disruption energy.
• Leak-controlled splits: Folds cluster complexes on MMseqs2 protein identity (50%) and Morgan-fingerprint ligand similarity, so no test protein is homologous to a training protein and no ligand scaffold is shared.
• Residue- and atom-level interpretability: Graph convolution weights identify which pocket residues and ligand fragments drive a prediction, rendered as 2D interaction networks and importance-coloured 3D structures.
• Hosted inference with pocket-wide scans: The web server and REST API run alanine scanning across every residue within 5 Å of the ligand, saturation mutagenesis heatmaps, and mutation lists of up to 500 variants.

#Technical Details

The pre-training network, ProteinLigandGCN, integrates RDKit and fpocket physicochemical descriptors, 768-dimensional MoLFormer molecular embeddings, graph-based signatures from cutoff scanning over eight pharmacophore categories at 0.5–10 Å, and GCNConv layers applied separately to the three graphs, concatenated into a fully connected head. Protein graph edges connect residue pairs whose Cα atoms lie within 10 Å, with node features from ProtT5 in the ProtTrans suite. Pre-training runs four epochs on the LP-PDBBind Clean Level 1 subset (pIC50, pKd, and pKi measurements). The fine-tuning network, ProtLigMutGCN, trains 55 epochs on the Platinum mutation set (n = 796, nearly ten times smaller), adding mutation descriptors and a dual wild-type/mutant path that converts predicted pKi to ΔG at 298.15 K before differencing.

On the LP-PDBBind test set DDMuffin reaches Pearson r = 0.54 with RMSE = 1.48 kcal mol⁻¹, the lowest error among AutoDock Vina, RF-Score, InteractionGraphNet, and DeepDTA retrained on the same split, though RF-Score attains r = 0.58. Excluding the 10% highest-residual predictions raises r to 0.66 on test and 0.70 in cross-validation. External sets give r = 0.38 (BDB2020+), 0.56 (EGFR), and 0.69 (SARS-CoV-2 Mpro). ΔΔG is harder: non-redundant 10-fold cross-validation on Platinum yields r = 0.20 and RMSE = 1.78 kcal mol⁻¹, while across eight ABL1 tyrosine kinase inhibitors DDMuffin averages Spearman ρ = 0.39, ahead of PremPLI (0.32), FEP+ (0.37), and Prime (0.38). Complexes without experimental structures were modelled with HelixFold3.

#Applications

The intended users are medicinal chemists and translational researchers tackling mutation-driven drug resistance. Saturation mutagenesis over a binding pocket flags which substitutions a target may acquire under inhibitor pressure, guiding the design of resilient compounds; alanine scanning of ABL1 complexes reproduced the distinct resistance landscapes of type I and type II inhibitors — hinge and gatekeeper hotspots for dasatinib, G-loop and DFG motif for nilotinib. The affinity mode supports lead prioritization and repurposing screens without local GPU infrastructure.

#Impact

DDMuffin extends the Biosig group's line of hosted mutation-effect predictors into joint affinity and ΔΔG modelling. Its principal contribution is methodological: mutation-specific fine-tuning proves necessary, since subtracting independent wild-type and mutant affinity predictions yields near-zero correlation. The authors are candid about the limits — ΔΔG accuracy varies sharply across validation schemes, collapsing under leave-one-ligand-out for common cofactors; lyases, ligases, and membrane transport proteins are over-represented among outliers; and predictions read as consensus affinity trends rather than exact binding constants. The work remains a preprint awaiting peer review, and no source code or weights have been released — the model is usable only through the free public web server and its API, though all training and benchmark datasets are downloadable from the site.