LaMPSite: Protein-Ligand Binding Prediction

• LaMPSite is a computational method that predicts protein–ligand binding sites using protein language models (ESM-2) and GNN-derived ligand representations.
• It integrates sequence embeddings with geometric interactions through a trigonometry-aware module to infer binding site clusters efficiently.
• Performance benchmarks show that LaMPSite rivals traditional 3D-structure approaches, enabling rapid, structure-independent predictions in drug discovery.

LaMPSite is a computational method for protein–ligand binding site prediction that leverages protein LLMs (PLMs) and graph neural networks (GNNs) to infer the three-dimensional localization of ligand binding sites from protein sequence and ligand graph alone, without requiring any 3D protein structural information. LaMPSite combines embeddings and geometric knowledge extracted from large pretrained protein LLMs (specifically, ESM-2) with ligand representations from GNNs and integrates them through a trigonometry-aware interaction network. This approach achieves performance competitive with structure-based state-of-the-art methods, even rivaling traditional 3D-structure-dependent pipelines, and thus addresses critical gaps in protein functional annotation and drug discovery scenarios where high-quality experimental or model structures are unavailable (Zhang et al., 2023).

1. Motivation and Problem Formulation

Protein–ligand binding site prediction is a central problem in structural bioinformatics, informing both functional annotation and rational drug design. The predominant set of computational methods depend heavily on high-resolution protein “holo” structures (complexed with ligand) and include geometry-based (e.g. Fpocket), probe-based, template-based, and 3D convolutional neural network (3D-CNN) techniques. However, such structures are missing for over half of all UniProt proteins, and cryptic or allosteric sites are systematically inaccessible to structure-based inference, limiting both functional discovery and target expansion for drug design. LaMPSite circumvents these limitations, predicting binding site residues and clusters from sequence and ligand chemical graph without reference to or requirement for experimental or model 3D structure (Zhang et al., 2023).

2. Model Architecture and Representational Flow

LaMPSite’s architecture proceeds through the following major computational stages:

(i) Protein Representation via ESM-2:

• The protein sequence (length $n_p$) is processed by a pretrained ESM-2 LLM. Outputs are
  • Residue-level embeddings $\mathbf{H}^p \in \mathbb{R}^{n_p \times d}$ ($d = 1280$)
  • A pairwise contact map $\mathbf{C}^p \in \mathbb{R}^{n_p \times n_p}$ from the unsupervised contact prediction head.

(ii) Ligand Graph Representation:

• Ligand is specified as a 2D molecular graph (heavy atoms only), with a node–edge topology.
• A 4-layer GNN yields atom embeddings $\mathbf{H}^l \in \mathbb{R}^{n_l \times d}$.
• An RDKit-generated conformer provides a distance map $\mathbf{D}^l \in \mathbb{R}^{n_l \times n_l}$.

(iii) Protein–Ligand Interaction Embedding:

• Initial interaction tensor $$\mathbf{Z}^{(0)} \in \mathbb{R}^{n_p \times n_l \times d}$$ formed via elementwise product: $$\mathbf{z}^{(0)}_{ij} = \mathbf{h}_i^p \odot \mathbf{h}_j^l$$.

(iv) Iterative Trigonometry-Aware Interaction Module:

• $T$ trigonometry blocks operate on $\mathbf{Z}$ to propagate information under geometric constraints enforced by the protein contact map ($\mathbf{H}^p \in \mathbb{R}^{n_p \times d}$0) and ligand distance map ($\mathbf{H}^p \in \mathbb{R}^{n_p \times d}$1). These operate analogously to Evoformer modules in AlphaFold [Lu et al. 2022].
• Each block updates via:

$\mathbf{H}^p \in \mathbb{R}^{n_p \times d}$2

where attention weights $\mathbf{H}^p \in \mathbb{R}^{n_p \times d}$3 are modulated by $\mathbf{H}^p \in \mathbb{R}^{n_p \times d}$4, $\mathbf{H}^p \in \mathbb{R}^{n_p \times d}$5.

(v) Merging and Pooling:

• Final interaction tensor: $\mathbf{H}^p \in \mathbb{R}^{n_p \times d}$6.
• Linear projection and mean-pooling over atoms yield per-residue scores

$\mathbf{H}^p \in \mathbb{R}^{n_p \times d}$7

• Higher $\mathbf{H}^p \in \mathbb{R}^{n_p \times d}$8 implies greater likelihood that residue $\mathbf{H}^p \in \mathbb{R}^{n_p \times d}$9 is in the binding site.

(vi) Clustering and Ranking:

• Thresholded, normalized scores ($d = 1280$0) select candidate residues.
• Single-linkage clustering (distance by $d = 1280$1) produces residue clusters.
• Clusters ranked by $d = 1280$2. The top-$d = 1280$3 clusters are reported.

3. Mathematical Formulation and Training Objective

The LaMPSite model’s loss and evaluation metrics are defined as follows:

• The binary cross-entropy loss with an L2 penalty is applied to per-residue predictions:

$d = 1280$4

where $d = 1280$5 indicates ground-truth binding (distance $d = 1280$6 Å).

• The primary evaluation metric is the DCA (Distance from predicted pocket Center to nearest ligand Atom) Success Rate:

$d = 1280$7

• Standard classification metrics (precision, recall, F1) are also reported if framed as per-residue binary classification.

4. Benchmarking, Ablations, and Implementation

Datasets:

• Training: scPDB v2017 (post-processing yields 16,270 protein–ligand pairs)
• Testing: COACH420 benchmark set (291 proteins, 359 ligands), with strict exclusion of homologs and chemical overlap

Baselines:

• Geometry-based: Fpocket
• 3D-CNN: DeepSite, Kalasanty, DeepPocket
• ML-based: P2Rank

Hyperparameters and Implementation:

• ESM-2-650M for protein embeddings
• 4-layer ligand GNN (hidden dim = 128)
• Interaction hidden dim = 32, with 2 trigonometry blocks
• Optimization: Adam (learning rate $d = 1280$8), batch size 8, early stopping (patience 4)
• Inference speed: ~0.2 s per query (V100 GPU, AMP enabled)
• Sequences longer than 850 residues truncated

Performance:

• On COACH420 (DCA < 4 Å, top-$d = 1280$9):
  • LaMPSite: 66.02%
  • LaMPSite\textsuperscript{holo} (experimental contact map): 67.96%
  • Fpocket: ~40–50%
  • DeepSite, Kalasanty: lower than LaMPSite
  • DeepPocket: 67.96%
  • P2Rank: 68.24%

Ablation Study (Top-$\mathbf{C}^p \in \mathbb{R}^{n_p \times n_p}$0 SR):

Ablation	Success Rate (%)
Full LaMPSite	66.02
–Clustering	65.18
–Merge $\mathbf{C}^p \in \mathbb{R}^{n_p \times n_p}$1	63.50
–Interaction module	62.40

5. Practical Implications and Limitations

LaMPSite’s ability to predict protein–ligand binding sites using only sequence and ligand graph is especially relevant for:

• Proteins lacking reliable experimental or predicted structures (situations where less than 50% of UniProt entries have coverage)
• Discovery of cryptic, allosteric, or induced sites that are not captured in available “apo” structures
• Rapid prioritization in structure-less proteomes and in early-stage drug discovery, where speed and minimal input requirements are paramount

Key limitations include:

• Tendency to predict a single binding site cluster per query; refinement for multi-site detection is required
• Model is restricted to single-chain protein inputs; multi-chain complexes are not yet addressed
• Sequence truncation at 850 residues due to GPU memory, hindering application to very large proteins
• Cryptic (apo→holo) site identification was not explicitly benchmarked
• Expanded benchmarks (e.g., membrane proteins, unconventional ligands) are needed for broader generalization

A plausible implication is the extension of the LaMPSite framework to multi-domain and multi-chain systems as well as incorporation of uncertainty estimates or cryptic pocket detection strategies.

6. Context within the Field and Future Outlook

LaMPSite illustrates that a purely sequence- and graph-based approach, powered by the latent geometric knowledge embedded in large protein LLMs and trigonometry-informed interaction networks, can achieve parity with established structure-dependent methods for site localization. Given that structure-independent binding site prediction addresses a major gap in current functional proteomics, LaMPSite establishes a new computational paradigm for the systematic annotation and targeting of “dark” proteins. The approach may catalyze further advances in druggability assessment, theoretical pocketome mapping, and low-overhead virtual screening (Zhang et al., 2023).

Future work includes multi-site prediction, handling multi-chain complexes, scaling to longer sequences, and systematic validation in scenarios involving cryptic or allosteric binding events. Additional benchmarking against newly emerged datasets or experimental results—especially in cases where structure-guided pipelines are infeasible—will be necessary to fully assess generality.