ExplainBind: Explainable Physicochemical Determinants of Protein-Ligand Binding via Non-Covalent Interactions

ExplainBind is a novel, structure-free AI framework that predicts protein-ligand binding likelihood, pinpoints specific binding residues, and decodes non-covalent interaction patterns to provide mechanistic insights for drug discovery, outperforming existing black-box models across diverse targets and successfully identifying both inhibitors and activators with distinct functional mechanisms.

The Gist

Imagine you are trying to figure out why a specific key fits perfectly into a specific lock. In the world of biology, the "lock" is a protein, and the "key" is a drug molecule. When they click together, it can stop a disease or fix a broken process in the body.

For a long time, scientists have used computer programs (AI) to predict if a key will fit a lock. But these programs have been like magic boxes: they can tell you "Yes, it fits!" or "No, it doesn't," but they can't explain why. They don't tell you which tiny part of the key touched which part of the lock, or whether the connection was held together by a magnetic pull, a sticky glue, or a gentle handshake.

Enter "ExplainBind." Think of this new tool as a super-sleuth that doesn't just guess the answer; it shows you the detective work.

Here is what makes ExplainBind special, using simple analogies:

• It Doesn't Need a Blueprint: Usually, to understand how a key fits a lock, you need a perfect 3D model (a blueprint) of the lock. ExplainBind is like a detective who can solve the mystery just by looking at the names and descriptions of the key and lock, without needing the 3D blueprint.
• It Finds the Exact Contact Points: Instead of just saying "the key fits in the lock," ExplainBind points to the exact teeth on the key and the exact grooves in the lock that are touching. It identifies the specific "residues" (the tiny building blocks) that do the heavy lifting.
• It Explains the "Glue": It breaks down the invisible forces holding them together. It tells you if the bond is strong because of an electrical attraction, a hydrogen bond (like a tiny Velcro strip), or something else. It translates the complex physics into a clear story.

How did it do in the real world?

The researchers tested this sleuth on two different "locks" it had never seen before:

1. The Blood Pressure Lock (ACE): They asked ExplainBind to find the best keys to stop a specific enzyme involved in blood pressure. It didn't just pick the winners; it explained why one key was stronger than another by showing the different patterns of how they stuck together.
2. The Metabolism Lock (L2HGDH): They asked it to find keys that could either turn this enzyme "off" (inhibitors) or "on" (activators). ExplainBind found both types and explained the difference: the "off" keys and the "on" keys used completely different patterns of interaction to achieve their opposite goals.

In short, ExplainBind moves drug discovery from a game of "blind guessing" to a game of "understanding the rules." It helps scientists see exactly how a drug works, not just that it works.

Technical Summary

Technical Summary of ExplainBind

Problem Statement
Protein-ligand binding is a fundamental process governing enzymatic catalysis, metabolic regulation, and therapeutic modulation, making its accurate prediction a cornerstone of drug discovery. Despite the proliferation of AI-driven methods in this domain, a critical limitation persists: existing approaches largely function as "black-box" predictors. These models typically fail to resolve two essential mechanistic questions: which specific amino acid residues mediate the binding event, and which non-covalent forces drive molecular recognition. Furthermore, many current methods rely heavily on three-dimensional structural inputs, which may not always be available or may introduce noise, limiting their applicability to novel sequences and scaffolds.

Methodology
To address these gaps, the authors present ExplainBind, an interaction-aware framework designed to operate without requiring three-dimensional structural inputs. The methodology is distinguished by three core capabilities:

1. Binding Likelihood Prediction: It accurately predicts the probability of binding between proteins and ligands.
2. Residue-Level Localization: Unlike prior models that often identify coarse "pocket-level" regions, ExplainBind pinpoints the specific binding residues involved in the interaction.
3. Interaction Decoding: The framework decodes the underlying patterns of non-covalent interactions, providing a transparent view of the physicochemical determinants driving affinity.

The model was trained and evaluated across a diverse dataset encompassing proteins and ligands ranging from closely related sequences to highly novel structures, ensuring robustness against distribution shifts.

Key Contributions

• Interpretability: The primary contribution is shifting the paradigm from opaque prediction to mechanistic explanation, explicitly mapping binding to specific residues and non-covalent interaction types.
• Structure-Free Operation: The framework achieves high performance without the prerequisite of 3D structural data, broadening its utility to targets where structural information is scarce.
• Granular Resolution: It moves beyond general pocket identification to residue-specific localization, offering a finer resolution of molecular recognition events.

Results
ExplainBind demonstrated consistent superiority over representative baseline models across various protein-ligand pairs. Its performance was validated in two specific application scenarios involving unseen targets:

• Angiotensin-Converting Enzyme (ACE): The model successfully ranked potent ACE inhibitors and elucidated potency differences by analyzing affinity-stratified interaction landscapes.
• L-2-hydroxyglutarate dehydrogenase (L2HGDH): In a prospective discovery setting, ExplainBind identified both inhibitors and activators of L2HGDH. Crucially, it provided mechanistic rationales for their divergent functional outcomes by revealing distinct interaction profiles, thereby explaining how different binding modes lead to opposing biological effects.

Significance
The paper positions ExplainBind as a broadly applicable tool for mechanistically informed drug discovery. By bridging the gap between predictive accuracy and physical interpretability, the framework enables researchers to not only predict binding but to understand the "why" and "how" of molecular recognition. This capability to rationalize functional outcomes through specific non-covalent interaction patterns represents a significant step forward in developing AI tools that support, rather than obscure, the chemical intuition required for effective therapeutic design.