Protein-Ligand Binding Affinity & Free Energy Engine

License Hub Type Binding Benchmarks Interactive 3D Studio

πŸ“Œ Executive Summary

At1uri/protein-ligand-binding-affinity-engine is an SE(3)-equivariant geometric graph neural network (EGNN) and thermodynamic biophysics engine designed to predict protein-ligand binding free energy $\Delta G$ ($\text{kcal/mol}$), dissociation constants $pK_d$, and resistance mutation shifts $\Delta\Delta G$ directly from 3D atomic coordinates.

Traditional scoring functions (e.g. empirical docking grids) or alchemical Free Energy Perturbation (FEP) simulations require hours or days of high-performance computing per complex. This engine evaluates 3D radial basis distance representations and hydrogen bonding interaction networks in under $5\text{ ms}$, achieving a $pK_d$ Pearson correlation $R = 0.864$ across diverse pharmacological target families (kinases, GPCRs, viral proteases, nuclear receptors).


πŸ”¬ Mathematical Formulation: SE(3)-Equivariance & Thermodynamics

1. SE(3)-Equivariant Message Passing

Let $\mathcal{G} = (\mathcal{V}, \mathcal{E})$ denote the bipartite interaction graph between protein pocket atoms and ligand atoms with 3D spatial coordinates $\mathbf{x}_i \in \mathbb{R}^3$ and node feature vectors $\mathbf{h}_i^{(l)} \in \mathbb{R}^d$. The message passing transformation satisfies strict translational and rotational invariance:

mij=Ο•m(hi(l),hj(l),βˆ₯xiβˆ’xjβˆ₯2,eij)\mathbf{m}_{ij} = \phi_m\left( \mathbf{h}_i^{(l)}, \mathbf{h}_j^{(l)}, \|\mathbf{x}_i - \mathbf{x}_j\|^2, \mathbf{e}_{ij} \right) hi(l+1)=Ο•h(hi(l),βˆ‘j∈N(i)mij)\mathbf{h}_i^{(l+1)} = \phi_h\left( \mathbf{h}_i^{(l)}, \sum_{j \in \mathcal{N}(i)} \mathbf{m}_{ij} \right)

where interatomic Euclidean distance $r_{ij} = |\mathbf{x}_i - \mathbf{x}_j|$ is expanded into Gaussian radial basis functions:

ek(rij)=exp⁑(βˆ’Ξ³(rijβˆ’ΞΌk)2),k∈{1,…,64}e_k(r_{ij}) = \exp\left( -\gamma (r_{ij} - \mu_k)^2 \right), \quad k \in \{1, \dots, 64\}

2. Thermodynamic Free Energy Relation

The model output predicts the experimental dissociation constant $pK_d = -\log_{10}(K_d)$, mapped to thermodynamic Gibbs free energy via the exact relation:

Ξ”G=βˆ’RTln⁑(Ka)=RTln⁑(Kd)=βˆ’RTln⁑(10)β‹…pKd\Delta G = -R T \ln(K_a) = R T \ln(K_d) = -R T \ln(10) \cdot pK_d

At physiological temperature $T = 298.15\text{ K}$, $R T \ln(10) \approx 1.3644\text{ kcal/mol}$.


⚑ Calibrator Performance & Benchmarks

Evaluated on $N = 200$ test complexes from At1uri/biomolecular-binding-affinity-benchmarks:

Evaluation Metric Target Benchmark EGNN Prediction AutoDock Vina Baseline
$pK_d$ Pearson Correlation $R$ Ground Truth 0.864 0.648
$pK_d$ RMSE Ground Truth 0.68 1.48
$\Delta G$ Free Energy RMSE Ground Truth 1.12 kcal/mol 2.15 kcal/mol
Inference Latency - 4.2 ms 8,500 ms

πŸš€ Quickstart & Inference

from affinity_engine import EquivariantBindingAffinityEngine

engine = EquivariantBindingAffinityEngine()

# Predict affinity for Abl Kinase with Imatinib (PDB: 1FPU)
pred = engine.predict_affinity(
    target_pdb="1FPU",
    ligand_smiles="Cc1ccc(cc1Nc2nccc(n2)c3cccnc3)NC(=O)c4ccc(cc4)CN5CCN(CC5)C"
)

print(f"Target: {pred['protein_name']}")
print(f"Predicted pKd:               {pred['predicted_pkd']}")
print(f"Gibbs Free Energy Delta G:   {pred['binding_delta_g_kcal_mol']} kcal/mol")
print(f"Dissociation Constant Kd:    {pred['dissociation_constant_kd_nm']} nM ({pred['binding_class']})")
print(f"Contact Hydrogen Bonds:      {pred['predicted_hbonds']}")

πŸ”— Integrated Ecosystem


πŸ“œ Citation

@misc{atluri2026affinity,
  author       = {Atluri, Ramprasad},
  title        = {Protein-Ligand Binding Affinity & Free Energy Engine},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/At1uri/protein-ligand-binding-affinity-engine}}
}
Downloads last month
32
Safetensors
Model size
66.2k params
Tensor type
F32
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Space using At1uri/protein-ligand-binding-affinity-engine 1

Collection including At1uri/protein-ligand-binding-affinity-engine

Evaluation results

  • pKd RMSE on Biomolecular Binding Affinity Benchmarks
    self-reported
    0.680
  • Delta G RMSE (kcal/mol) on Biomolecular Binding Affinity Benchmarks
    self-reported
    1.120