Abstract: In drug discovery, determining the binding affinity and functional effects of small-molecule ligands on proteins is critical. Current computational methods can predict these protein-ligand interaction properties but often lose accuracy without high-resolution protein structures and falter in predicting functional effects. We introduce PSICHIC (PhySIcoCHemICal graph neural network), a framework uniquely incorporating physicochemical constraints to decode interaction fingerprints directly from sequence data alone. This enables PSICHIC to attain first-of-its-kind emergent capabilities in deciphering mechanisms underlying protein-ligand interactions, achieving state-of-the-art accuracy and interpretability. Trained on identical protein-ligand pairs without structural data, PSICHIC matched and even surpassed leading structure-based methods in binding affinity prediction. In a library screening for adenosine A1 receptor agonists, PSICHIC discerned functional effects effectively, ranking the sole novel agonist within the top three. PSICHIC’s interpretable fingerprints identified protein residues and ligand atoms involved in interactions. We foresee PSICHIC reshaping virtual screening and deepening our understanding of protein-ligand interactions.
Introduction: The Challenge of Protein-Ligand Interaction Prediction
Predicting how a small molecule (ligand) interacts with a protein is a fundamental challenge in drug discovery. Researchers need to determine not just if a molecule binds, but its binding affinity (potency) and its functional effect—whether it activates (agonist) or inhibits (antagonist) a target.
Existing computational tools generally fall into two categories. Structure-based methods rely on 3D atomic coordinates of the protein-ligand complex. While physically grounded, they are computationally intensive and highly sensitive to errors in experimental or predicted (in silico) structures. Sequence-based methods, which use 1D amino acid sequences and chemical strings (SMILES), are faster and more scalable but often struggle with "overfitting"—learning statistical correlations rather than true chemical principles. PSICHIC (PhySIcoCHemICal graph neural network) addresses these limitations by embedding physical constraints directly into a sequence-based architecture, achieving the groundedness of structural models with the speed of sequence models.
The PSICHIC Framework: Decoding Interaction Fingerprints
PSICHIC represents both the protein and the ligand as graphs. For the ligand, atoms are nodes and covalent bonds are edges. For the protein, residues are nodes, and edges are defined by a predicted contact map generated by the ESM2 protein language model.
The core of the architecture consists of three iterative Physicochemical Graph Convolutional (PGC) Layers:
- Intramolecular Forces: The model uses Principal Neighbourhood Aggregation (PNA) to capture internal structural features within the ligand and protein independently.
- Physicochemical Constraints: This layer groups ligand atoms into functional groups and protein residues into clustered regions. By forcing the model to interact through these chemically meaningful clusters rather than individual, dispersed atoms, the authors effectively regularize the model with "chemical common sense."
- Intermolecular Forces: The model aggregates ligand functional groups into a "ligand ball" and employs cross-attention to calculate interaction strengths between this ball and specific protein regions.
The final output is an interaction fingerprint—a vector that represents the binding mechanism. Because this fingerprint is derived from attention scores, it is inherently interpretable, allowing researchers to visualize which atoms and residues drive the interaction.
Experimental Evaluation and Benchmarking
PSICHIC was evaluated on a massive scale, including the PDBBind gold-standard affinity datasets and binary interaction datasets like BioSNAP and BindingDB. On the PDBBind v2020 benchmark, PSICHIC achieved an RMSE of 1.314, outperforming all other sequence-based methods and matching the top-performing structure-based models.
A standout feature of PSICHIC is its robustness. Unlike structure-based models, which show significant performance degradation as the resolution of the protein structure decreases, PSICHIC’s accuracy remains stable. It also maintains high performance when using predicted structures from AlphaFold2 or ESMFold, whereas traditional methods often falter when the input structure is not perfect.
Emergent Interpretability and Real-World Application
Beyond affinity, PSICHIC excels at predicting functional effects. Trained on a curated dataset of ~3 million samples, it achieved a 0.96 accuracy in classifying agonists, antagonists, and non-binders. In a "blind" virtual screen of the MIPS library (30,282 compounds) for adenosine $A_1$ receptor agonists, PSICHIC ranked the only novel agonist (MIPS1) in the top three. In contrast, the leading structure-based method, TankBind, ranked it at 20,804.
Furthermore, PSICHIC’s fingerprints provide "emergent structural awareness." Without ever seeing a 3D structure, the model correctly identified critical residues in the $A_1$ receptor's orthosteric binding site and extracellular loop 2. This ability to pinpoint pharmacophores and binding residues makes PSICHIC a powerful tool for mechanistic understanding and lead optimization.
Conclusion
PSICHIC demonstrates that high-resolution 3D structures are not always a prerequisite for accurate and interpretable protein-ligand interaction modeling. By incorporating physicochemical constraints into a graph neural network, PSICHIC offers a robust, scalable, and physically grounded alternative to traditional methods. We anticipate that PSICHIC will significantly impact virtual screening and the design of novel therapeutics by providing deep insights into the molecular basis of binding.
Summary updated