PBCNet2.0

Lead optimization rarely asks whether a molecule binds — it asks whether the next analog binds better. Free energy perturbation (FEP), and in particular Schrödinger's FEP+, answers that question at roughly chemical accuracy but costs hours of GPU simulation per ligand pair, which keeps it off the critical path in most medicinal chemistry campaigns. PBCNet2.0, developed at the Shanghai Institute of Materia Medica of the Chinese Academy of Sciences with collaborators across several Chinese universities, targets exactly this gap: it predicts the relative binding affinity between two congeneric ligands occupying the same pocket, directly from their docked complex structures, in a single forward pass.

PBCNet2.0 is the second generation of PBCNet, and rebuilds it along two axes. First, the training set was expanded to 8.6 million protein–ligand complex pairs derived from BindingDB (2023.12), a 14-fold increase over the original. Second, the geometric encoder was replaced: instead of hand-engineered interaction descriptors, PBCNet2.0 uses an equivariant message-passing network built on Cartesian rank-2 tensors, so that three-dimensional geometry — distances, angles, and the directionality of interactions — is represented natively rather than approximated.

The work was posted to bioRxiv in June 2025 and published in Nature Chemical Biology in 2026, accompanied by a public repository containing trained weights, inference and retraining code, and the scripts to reproduce every figure in the paper.

#Key Features

• Pairwise, not absolute: A Siamese network scores a reference ligand and a query ligand against the same protein and predicts the difference in activity (ΔpAct), which sidesteps the systematic offsets that plague absolute affinity scoring and matches how chemists actually compare analogs.
• Cartesian tensor equivariance: Atomic environments are represented as rank-2 tensors (3 × 3 matrices) that decompose into scalar, vector, and tensor irreducible components, allowing rotationally equivariant message passing to be implemented with plain matrix multiplication.
• FEP+ parity, zero-shot: On a 16-chemical-series FEP benchmark the model reaches ρ = 0.67 and R = 0.66 without any target-specific data, against FEP+ at ρ = 0.70 and R = 0.70 — a difference that is not statistically significant (p = 0.39).
• Data-efficient fine-tuning: Given a modest amount of target-specific SAR data, fine-tuned performance rises steadily with the number of ligands supplied and overtakes the FEP+ baseline.
• Emergent mutation sensitivity: Although no mutation data appears in training, the model averages ρ = 0.53 on a benchmark of eight targets with clinically observed binding-pocket mutations, a task on which Boltz-2, Glide SP, RTMScore, and GenScore all produced negative correlations.

#Technical Details

The architecture extends TensorNet-style Cartesian tensor message passing to protein–ligand complexes, adding interface edges between ligand and pocket atoms so the network reasons over intermolecular contacts rather than the ligand alone. Training pairs were built from IC50, EC50, Ki, and Kd measurements in BindingDB, with input poses generated by Glide docking and rule-based filtering, yielding 8.6 million pairs that share a protein target and a common binding mode. Evaluation used two established FEP benchmark sets against twelve baselines spanning physics-based scoring (Glide SP, MM-GB/SA, FEP+), sequence-based predictors (PSICHIC, BIND, PLAPT), structure-based scoring functions (PIGNet2, RTMScore, GenScore, OnionNet-2, PBCNet), and Boltz-2. In retrospective prioritization across eight pharmaceutically relevant targets, PBCNet2.0 reached an advantage order of 13.57 (PBCNet 11.18, MM-GB/SA 5.89), cut theoretical resource investment by 40.5%, and improved optimization efficiency by 7.2-fold. Ablating protein–ligand interaction edges degrades predictions, and weight analysis on a thymidylate synthase N229C mutant shows the model's signal collapsing exactly where the abolished hydrogen bond was — evidence that it reads interactions rather than memorizing ligand scaffolds.

#Applications

The model is aimed at structure-based lead optimization and chemical probe development, where a chemist has a validated hit and needs to rank tens or hundreds of proposed analogs. Two prospective campaigns exercised it end to end. On ENPP1, PBCNet2.0 correctly predicted that removing a fluorine orthogonal multipolar interaction with TRP322 would weaken binding, confirmed by thermal shift, enzyme assay, and SPR (Kd 35.3 nM versus 1030 nM). On ALDH1B1, it distinguished two stereoisomers of an inhibitor and, via in-silico alanine scanning of the pocket, nominated six residues as binding-critical; site-directed mutagenesis confirmed five of the six.

#Impact

PBCNet2.0 is the first machine-learning method to reach statistical parity with industrial FEP on relative binding affinity ranking while running in seconds rather than hours, which makes affinity-guided triage practical at a scale physics-based simulation cannot reach. Its mutation-prediction behavior — acquired without mutation training data — also points toward resistance forecasting, a use case where co-folding models currently fail outright. The limitations are concrete: predictions depend on static docked poses and therefore ignore entropic and conformational contributions to binding, and the pairwise formulation assumes both ligands share a binding mode, so R-group changes that flip the pose are mispredicted. Applying the model still requires a docking pipeline and pose validation upstream.