RAVEN: Frozen Random Graph Reservoirs with Physics-Informed Interaction Fingerprints for Protein-Ligand Binding Affinity Prediction

RAVEN (Randomized Atomistic Views with Ensemble Neural Reservoirs) is a protein–ligand binding affinity prediction framework that combines multiple independently initialized and fully frozen atomistic graph encoders with a deterministic physicochemical interaction fingerprint and heterogeneous supervised readers. The central design principle is that competitive protein–ligand affinity regression can be obtained by combining multiple independently sampled frozen graph representations with explicit physicochemical descriptors and heterogeneous supervised readers rather than through end-to-end optimization of the graph representation.

The method represents each protein–ligand complex as a directed atomistic heterogeneous graph with ligand and protein atoms as distinct node types, intramolecular covalent bonds, and directed cross-interface contacts. The graph is processed by multiple independently initialized relation-specific message-passing neural networks (MPNNs), each containing four propagation layers, with all parameters fully frozen after initialization, preserving the separation between graph representation and supervised affinity regression. Each encoder produces three 384-dimensional graph-level views (protein, cross-interface interaction, and ligand domains), concatenated into a 1152-dimensional feature block. The final configuration uses 32 ordinary encoders plus one additional asymmetric global protein–ligand branch, yielding a 37,632-dimensional raw structural reservoir.

The deterministic physicochemical fingerprint is a fixed 788-dimensional representation constructed without affinity labels, trainable parameters, or random sampling. It includes molecular composition and intramolecular topology (522 dimensions), cross-interface geometry (64 dimensions), and channel-resolved physicochemical interaction statistics (202 dimensions) covering ten interaction categories: van der Waals contact, hydrogen bonding, ionic attraction, ionic repulsion, hydrophobic contact, aromatic interaction, metal coordination, halogen bonding, cation–π interaction, and a water-mediated polar-contact proxy.

Three supervised expert families decode these representations: reservoir-only MLPs (R-MLPs) using the frozen structural representation alone, joint MLPs (J-MLPs) combining structural reservoir with physicochemical fingerprint, and ExtraTrees regressors operating directly on the deterministic fingerprint. Three independently trained experts are used within each family, producing nine candidate experts combined through validation-fitted nonnegative softmax fusion.

The model was evaluated on a similarity-isolated PDBbind 2020R1 split reconstructed using GEMS similarity resources, with 8,161 Train, 1,747 Validation, and 1,749 Test complexes, including all 285 available CASF-2016 complexes as a protected subset. The final 32-head model achieved a Pearson correlation of 0.7995, a Spearman correlation of 0.7922, and an R2 of 0.6341 on the complete PDBbind 2020R1 Test set, with RMSE and MAE of 1.2396 and 0.9379 on the dimensionless pK scale. On the protected CASF-2016 subset, Pearson (r) and Spearman (ρ) reached 0.8446 and 0.8407, respectively, while R2 reached 0.6895, with RMSE and MAE of 1.2095 and 0.9243.

Reservoir-size experiments showed that increasing the number of independent frozen encoders generally improved structure-only R-MLP performance, with Pearson correlation on PDBbind Test rising from 0.7548 at M=1 to 0.7803 at M=32 with the global branch retained. However, gains became smaller and locally nonmonotonic at larger sizes, consistent with the competition between representation expansion and supervised estimation. The global protein–ligand branch provided a directionally consistent contribution across all evaluated reservoir sizes.

Regressor-family ablations revealed that the physicochemical fingerprint was particularly well matched to tree-based readers such as ExtraTrees and XGBoost, with ExtraTrees achieving Pearson correlations of 0.7781 on PDBbind Test and 0.8196 on CASF-2016, compared to Physics-MLP's 0.7395 and 0.7726. The complete three-family ensemble (R-MLP, J-MLP, ExtraTrees) provided the most balanced performance, and expert multiplicity (three instances per family versus one) improved Pearson correlation from 0.7785 to 0.7967 on PDBbind Test and from 0.8197 to 0.8320 on CASF-2016.

Residual correlation analysis showed extremely high within-family residual correlations (approximately 0.991 for R-MLP, 0.990 for J-MLP, 0.998 for ExtraTrees), with mean R-MLP–J-MLP residual correlation of approximately 0.985, but substantially lower cross-family correlations between ExtraTrees and neural families (approximately 0.835 and 0.841). The learned fusion weights distributed across all nine experts with an effective ensemble size of approximately 7.7, with ExtraTrees receiving the largest total weight (0.493).

In comparison with representative models on CASF-2016 scoring power, RAVEN's Pearson correlation of 0.8446 exceeded AutoDock Vina (0.604), Pafnucy (0.780), PIGNet ensemble (0.761), and InteractionGraphNet (0.837), while remaining below specialized interaction-fingerprint approaches including OnionNet-2 (0.864), ECIF6::LD-GBT (0.866), and Multi-shelled ECIF (0.877).

The paper concludes that "useful protein–ligand representations do not require every component to be optimized jointly: randomized structural projections, explicit physicochemical information, and heterogeneous regression models can instead be combined to exploit complementary strengths." The authors note that the experiments establish predictive viability of frozen reservoirs rather than universal superiority over end-to-end trained encoders, and suggest future work including evaluation on temporally separated or prospective datasets, pose-ranking and virtual-screening experiments, and direct comparison between frozen and end-to-end trained versions of the same encoder family.

Improvements to AI Systems:

1. Frozen-Representation Ensemble Learning: Implement a training paradigm where multiple independently initialized, fully frozen graph encoders are combined with deterministic handcrafted features and heterogeneous supervised readers, eliminating the need for end-to-end backpropagation through the representation layer. This reduces overfitting, computational cost, and catastrophic forgetting while enabling scalable parallel feature extraction.

2. Cross-Family Residual Decorrelation for Ensemble Fusion: Use validation-fitted nonnegative softmax fusion across deliberately diverse regressor families (e.g., MLPs, tree-based models) rather than relying on homogeneous deep ensembles. The paper shows that cross-family residual correlations (0.84) are much lower than within-family (0.99), so the improved system explicitly maximizes inter-family diversity to boost ensemble accuracy and robustness.

3. Physicochemical Fingerprint-Guided Tree-Based Regression: Integrate a fixed, label-free 788-dimensional physicochemical interaction fingerprint (covering 10 interaction types, cross-interface geometry, and topology) as a primary input to tree-based models like ExtraTrees or XGBoost, which outperform neural MLPs on this feature type. The improved system can use this as a fast, interpretable baseline or as a component of a hybrid pipeline.

4. Asymmetric Global Protein–Ligand Branch: Add a dedicated global branch that captures whole-complex interactions (beyond local atomistic views) alongside multiple local encoders. This branch provides consistent performance gains across all reservoir sizes, so the improved system should include such a global context module to stabilize and enhance predictions.

5. Reservoir-Size Scaling with Diminishing Returns Awareness: Dynamically tune the number of frozen encoders (M) based on validation performance, stopping when gains become nonmonotonic. The improved system can implement an early-stopping criterion for representation expansion, balancing capacity against supervised estimation noise.

6. Multi-Expert with Effective Ensemble Size Control: Train three instances per regressor family and fuse them with learned nonnegative weights, yielding an effective ensemble size of 7.7 out of 9 experts. The improved system can use this weight distribution to prune redundant experts and allocate compute to the most complementary models (e.g., ExtraTrees receiving 0.49 weight).

What the Improved AI System Can Do:

• Predict protein–ligand binding affinities (pK values) with high accuracy (Pearson r 0.80 on PDBbind, 0.84 on CASF-2016) without end-to-end training of graph encoders, making it deployable on limited hardware.

• Achieve robust generalization on similarity-isolated test sets, including protected CASF-2016 subsets, by leveraging frozen representations that avoid label leakage.

• Provide interpretable predictions via the deterministic physicochemical fingerprint, enabling analysis of which interaction types (e.g., hydrogen bonding, halogen bonding) drive binding strength.

• Scale efficiently to large compound libraries for virtual screening, as the frozen encoders can be precomputed once and reused across many ligands.

• Serve as a strong baseline or ensemble member in hybrid systems, where its complementary strengths (tree-based on fingerprints vs. neural on structural reservoirs) improve overall accuracy beyond any single model family.

• Adapt to new datasets without retraining the graph encoders—only the supervised readers and fusion weights need fitting, enabling rapid transfer to new targets or assay conditions.

Quantitative estimation of protein-ligand binding affinity from three-dimensional complex structures is a fundamental task in structure-based computational chemistry and molecular modeling. Reliable prediction remains challenging because available structure-affinity data are limited, experimentally heterogeneous, conformation-dependent, and sensitive to dataset partitioning. RAVEN (Randomized Atomistic Views with Ensemble Neural Reservoirs) utilizes a multihead reservoir of independently initialized and fully frozen atomistic graph encoders to generate diverse structural projections without end-to-end optimization of the graph representation. These projections are integrated with a deterministic physicochemical interaction fingerprint and processed by heterogeneous supervised readers, including neural and tree-based regressors, whose outputs are combined through validation-based nonnegative fusion. The random reservoir expands structural feature coverage across independent encoder realizations, whereas the explicit physicochemical descriptors and heterogeneous readers contribute complementary information and distinct inductive biases. Evaluation on a similarity-isolated PDBbind 2020R1 split reconstructed using GEMS similarity resources, together with the protected CASF-2016 subset, demonstrated strong predictive performance. The results indicate that frozen multi-view graph representations, explicit physicochemical statistics, and heterogeneous model fusion provide a robust and flexible framework for protein-ligand binding-affinity prediction.