Trigonometry-Aware Update for LaMPSite

• Trigonometry-Aware Update is a deep learning framework that fuses transformer-based protein embeddings with GNN ligand representations to predict binding sites directly from sequence data.
• It employs a modular design with a trigonometry module enforcing geometric constraints, achieving competitive success rates compared to methods dependent on detailed 3D structures.
• The framework also distinguishes itself by paralleling a separate application in LPMO active-site characterization, illustrating its dual relevance in computational biochemistry.

LaMPSite refers to two distinct but unrelated technical topics in protein biochemistry and machine learning: (1) a transformer-based computational framework for predicting ligand binding sites from protein sequence and ligand graph, and (2) the designation of a specific copper-oxygen coordination motif central to the function of lytic polysaccharide monooxygenases (LPMOs). Both are pivotal for modern approaches to protein function annotation and mechanistic enzymology, respectively.

1. Sequence-Based Ligand Binding Site Prediction with LaMPSite

LaMPSite is a computational method enabling 3D ligand binding site prediction directly from protein sequence and a ligand molecular graph, eschewing the need for experimentally derived protein structures. This approach addresses critical limitations in traditional ligand binding site prediction pipelines, which typically require high-resolution holo protein structures. Fewer than half of UniProt proteins currently have high-confidence structure annotations, and many drug discovery campaigns center on uncharacterized or structurally unresolved targets. LaMPSite circumvents this bottleneck by leveraging emergent structural features encoded in large-scale protein LLMs and by fusing those with detailed representations of ligands via graph neural networks (GNNs) (Zhang et al., 2023).

2. Computational Architecture and Data Flow

The LaMPSite workflow is modular, comprising the following main components:

• Protein Representation via LLM:
  • Input amino-acid sequence (length $n_p$) is processed using the ESM-2 transformer.
  • Yields a residue embedding matrix $\mathbf{H}^p \in \mathbb{R}^{n_p \times d}$ (with $d=1280$) and a predicted contact map $\mathbf{C}^p \in \mathbb{R}^{n_p \times n_p}$ via the contact prediction head.
• Ligand Encoding via GNN:
  • Ligands entered as heavy-atom 2D molecular graphs.
  • A 4-layer GNN produces atom-level embeddings $\mathbf{H}^l \in \mathbb{R}^{n_l \times d}$ (with hidden dimension 128).
  • A single conformer generated by RDKit provides the ligand distance map $\mathbf{D}^l \in \mathbb{R}^{n_l \times n_l}$, representing atomic pairwise distances.
• Interaction Embedding and Trigonometry Module:
  • Construct initial interaction tensor $$\mathbf{Z}^{(0)} \in \mathbb{R}^{n_p \times n_l \times d}$$ by Hadamard product: $$\mathbf{z}^{(0)}_{ij} = \mathbf{h}^p_i \odot \mathbf{h}^l_j$$.
  • The trigonometry module, inspired by Evoformer, iteratively updates $\mathbf{Z}$ by propagating messages weighted by predicted intra-protein contacts ($\mathbf{C}^p$) and ligand atom distances ($\mathbf{H}^p \in \mathbb{R}^{n_p \times d}$0). This approach enforces geometric constraints (triangle inequalities, steric clash avoidance).
• Scoring, Pooling, and Clustering:
  • After $\mathbf{H}^p \in \mathbb{R}^{n_p \times d}$1 trigonometry blocks, merge representations: $\mathbf{H}^p \in \mathbb{R}^{n_p \times d}$2.
  • Apply linear projection and mean-pooling over ligand atoms to yield per-residue scores $\mathbf{H}^p \in \mathbb{R}^{n_p \times d}$3; high $\mathbf{H}^p \in \mathbb{R}^{n_p \times d}$4 indicates binding likelihood.
  • Normalized $\mathbf{H}^p \in \mathbb{R}^{n_p \times d}$5 (thresholded at 0.63) identifies candidate binding residues; candidates are clustered based on the predicted contact map, and clusters are ranked by the sum of squared residue scores.

3. Mathematical Formulation

The update mechanism in the trigonometry module can be summarized as:

$\mathbf{H}^p \in \mathbb{R}^{n_p \times d}$6

The objective function is a binary cross-entropy loss per residue, with L2 regularization:

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

where ground-truth $\mathbf{H}^p \in \mathbb{R}^{n_p \times d}$8 if the Cα is within 8 Å of any ligand atom.

Performance is mainly quantified by DCA Success Rate (fraction with pocket center within 4 Å of the nearest ligand atom), and by per-residue metrics such as precision, recall, and $\mathbf{H}^p \in \mathbb{R}^{n_p \times d}$9.

4. Training Protocols, Datasets, and Baselines

LaMPSite’s model is trained on the scPDB v2017 dataset, comprising 16,270 protein–ligand pairs after stringent redundancy and similarity filtering. The COACH420 dataset (291 structures, 359 ligands) is reserved for external evaluation, with no homologous proteins or similar ligands in the training pool. Baselines for benchmarking include geometry-based (Fpocket), 3D-CNN-based (DeepSite, Kalasanty, DeepPocket), and machine-learning (P2Rank) approaches.

Implementation uses ESM-2-650M, a 4-layer GNN with 128 hidden-dimension, and two trigonometry blocks. The optimizer is Adam (learning rate $d=1280$0, batch size 8, early stopping with 4-patience epochs). Protein sequences exceeding 850 residues are truncated.

5. Performance Evaluation

Quantitative results demonstrate LaMPSite achieves a top-$d=1280$1 DCA success rate of 66.02% (sequence only) on COACH420, comparable to models using explicit 3D protein inputs (DeepPocket: 67.96%, P2Rank: 68.24%). Using experimental contact maps (holo conditions) slightly improves LaMPSite’s rate to 67.96%. Fpocket and most neural-network-based 3D structure models perform notably worse (Fpocket: 40–50%). Ablation studies illustrate the impact of each architectural component; removal of clustering, interaction module, or the merge step decreases success by 0.8–3.6 percentage points.

Inference speed is approximately 0.2 seconds per protein–ligand pair on a V100 GPU.

6. Limitations and Future Directions

• Current design predicts a single binding site cluster per query; extension to multiple-site identification is required for allosteric and multi-pocket proteins.
• Only single-chain proteins are supported; multi-chain complexes and very long sequences present scalability challenges.
• Potential utility in detecting cryptic (“apo→holo”) pockets remains to be systematically evaluated.
• Additional benchmarking on challenging classes (cryptic sites, membrane proteins) would validate generality.
• Sequences > 850 residues could not be processed due to hardware/memory constraints.

7. LaMPSite in Enzyme Active Site Terminology

Independently, “LaMPSite” denotes the copper active site in lytic polysaccharide monooxygenases (LPMOs) (Hedegård et al., 2016):

• Cu(II), Resting State: Distorted octahedral geometry, defined primarily by three strong equatorial imidazoles (His86, His1, His1) and weakly bound axial waters and a tyrosine.
• Cu(I), Reduced State: Collapse to a four-coordinate T-shaped or pseudo-tetrahedral environment, with loss of strong axial interactions.
• Superoxide Binding: O₂⁻ binds preferentially in the equatorial plane (displacing equatorial water), stabilized by a second-sphere Gln173 H-bond; equatorial binding is energetically favored by >60 kJ/mol over axial binding.
• Frontier Orbitals: Significant mixing between Cu $d=1280$2 and O₂ $d=1280$3; spin density is delocalized (0.6 e– on Cu, 0.4 e– on superoxide).

A plausible implication is that the “LaMPSite” bottleneck described in LPMO enzymology and the LaMPSite framework in deep learning for protein–ligand interaction analysis are homophonous but conceptually discrete.

---

For sequence-based ligand binding site prediction, LaMPSite demonstrates that transformer-derived structural priors, combined with ligand GNNs and geometric message propagation, can reach or surpass the precision of methods requiring crystal structures, expanding the applicability to the proteome fraction lacking reliable 3D information (Zhang et al., 2023).