ProMaya

Two protein surfaces can be geometrically complementary and still refuse to bind. What separates a real interface from a merely well-fitting one is how tightly its atoms are packed: hydrophobic cores, stacked aromatics and buried polar networks generate enough van der Waals attraction to pay for the entropy lost on association. Sequence models never see that packing, and residue-level graphs average it away.

ProMaya makes it an explicit input. Developed at CSIR-Institute of Himalayan Bioresource Technology with AcSIR, it computes a Gaussian-smoothed local surface mass density (LSMD) for every heavy atom and carries that scalar upward through residue and surface representations. Those geometric channels are fused with per-residue embeddings from a frozen protein language model, so evolutionary signal and physical packing are read jointly rather than concatenated at the end.

It sits alongside X-PAIR and FlashPPI, which answer the same yes/no question from sequence, and PUMBA, which scores docked poses. ProMaya answers only whether a pair interacts — it is not a structure, interface geometry or affinity predictor.

#Key Features

• Mass density as a first-class feature: LSMD is computed per heavy atom, mean-aggregated to residues and interpolated onto surface points, giving one physically interpretable scalar that ties all three scales together.
• Modalities aligned before fusion: rather than encoding four inputs independently and concatenating once, ProMaya runs four staged rounds of masked bidirectional cross-attention, each enforcing a physical consistency constraint through an auxiliary loss.
• Disorder treated as signal, not noise: IUPred2A long-mode scores enter as a residue feature and again as a soft supervisory target, so flexible segments that bind through short linear motifs are represented rather than discarded.
• Attribution at three scales: a node-level reformulation of Grad-CAM reports which atoms, residues and surface patches drove a prediction.
• Deployment without assembly: a containerized Nextflow DSL2 pipeline and a hosted web server both return interaction scores, residue annotations and network files.

#Technical Details

Each protein is encoded in four modalities: an atomic graph with an 8 Å cutoff and 41-dimensional nodes (element and hybridization one-hots, Gasteiger partial charge, van der Waals radius, LSMD); a residue graph with a 10 Å Cα cutoff and 59-dimensional nodes (DSSP secondary structure, relative SASA, torsions, IUPred2A disorder, a PSI-BLAST PSSM with an explicit no-MSA flag); a 1,024-point MSMS surface cloud carrying curvature, Poisson-Boltzmann electrostatics and interpolated LSMD, encoded by a three-layer PointNet++; and 1,024-dimensional per-residue embeddings from the frozen ProtT5-XL-UniRef50 encoder of ProtTrans. A heterogeneous graph transformer with 8 attention heads and 4-8 layers encodes each protein, cross-attention compares the pair, and the resulting 128-dimensional interaction signature is classified by an isotonically calibrated gradient-boosted ensemble fit after the network is frozen. The trainable stack is 46M parameters.

Training drew 46,706 binary complexes from PDBbind 2021 and DIPS-Plus across six organisms, filtered to 3.0 Å resolution and ≥200 Ų buried interface area and deduplicated at 40% sequence identity with CD-HIT. Negatives came in five classes - localization-incompatible, random, interface-ablated, context-aware hard negatives and low-scoring docking decoys - trained against with a focal-loss term. The authors report 95.7% accuracy, AUROC 0.958 and AUPRC 0.823 on their held-out test set against eight prior tools, and MCC ≥ 0.92 with F1 ≥ 0.96 on four external benchmarks (human SHS27k, 1,472 SARS-CoV-2 host-pathogen pairs, 5,700 mouse pairs, 715 maize pairs). On the subset with no detectable homology to training complexes they report 94.3% accuracy against 36.2% for an HHsearch template baseline, and 95.9% under a cold-start split where neither partner was seen in training.

#Applications

The intended use is proteome-scale screening in organisms with no curated interactome. The authors demonstrate this on the Himalayan medicinal plant Picrorhiza kurrooa, reconstructing a cold-responsive iridoid metabolon from RNA-seq differential expression plus predicted structures, with attribution pointing at the hydrophobic contacts that stabilize it. Host-pathogen work follows the same path, and single-pair inference takes 0.5-1.8 seconds on a workstation GPU or 8.7-37.5 seconds on a 16-core CPU, so triaging candidates before co-immunoprecipitation needs no cluster access.

#Impact

PPI benchmarking is unusually easy to inflate: accuracy on randomly paired negatives and on test proteins homologous to the training set overstates what a predictor will do on genuinely novel partners. ProMaya's contribution is as much in confronting that as in the architecture - the five negative classes, the identity clustering and the cold-start splits are the paper's own answer to it, and the reported margins are largest where those controls bite hardest. The numbers are self-reported from a preprint that has not yet been peer reviewed, and accuracy is a generous framing for a task this sensitive to class balance. The release is partial: the trained checkpoint, a Docker image and the Nextflow orchestration layer are archived on Zenodo under CC BY 4.0 alongside a public web server, but the model code itself ships only as compiled bytecode, and neither the 88,740-pair training corpus nor the four external benchmark splits are distributed — so the released artifacts let you rerun ProMaya, not independently check its numbers.