🤖 machine learning

TriGlue: a Biology-Inspired Generative Model for Generating Molecular Glue-Induced Ternary Complex

This paper introduces TriGlue, a biology-inspired generative framework that addresses the computational challenge of designing molecular glue degraders by decomposing the problem into interface estimation and interface-conditioned ternary complex generation to simultaneously predict protein-protein interactions and generate novel molecular glues.

Original authors: Yuliang Yan, Shuo Yan, Haochun Tang, Yiqin Sun, Enyan Dai

Published 2026-07-27

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Yuliang Yan, Shuo Yan, Haochun Tang, Yiqin Sun, Enyan Dai

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the human body as a bustling city where proteins are the workers, the buildings, and the traffic lights. Sometimes, a specific worker (a "target protein") goes rogue, causing chaos like a disease. To fix this, scientists want to build a tiny, custom-made tool that can quietly sneak into the city, find the rogue worker, and convince a cleanup crew (the cell's natural waste disposal system) to take it out. This is the world of Targeted Protein Degradation.

Usually, to get the cleanup crew to grab the rogue worker, you need a giant, complex machine (a drug) that physically bridges the two. But sometimes, nature uses a much simpler trick: a "molecular glue." Think of this glue as a tiny, sticky note. It doesn't need to be a giant machine; it just needs to stick to the cleanup crew and the rogue worker at the same time, remodeling their surfaces so they stick together naturally. Once they are glued, the cleanup crew tags the rogue worker and destroys it. The problem is, finding or inventing these tiny, perfect sticky notes is incredibly hard. Scientists have only found a few hundred in history, mostly by accident or through years of expensive, messy trial-and-error in a lab. They are like needles in a haystack, but the haystack is made of invisible, shifting shapes.

This is where a new team of researchers steps in with a digital solution called TriGlue. Instead of guessing in a lab, they built a computer brain inspired by biology to design these molecular glues from scratch. The paper suggests that by teaching the computer to understand how proteins move and stick together, they can generate new, valid molecular glues and predict exactly how they will assemble the "glued" trio (the cleanup crew, the glue, and the target). The results from their simulations show that their method is better at creating these complex structures than previous attempts, offering a promising new way to speed up the discovery of these powerful drugs.

The Problem: The Invisible Puzzle

Designing a molecular glue is like trying to solve a three-dimensional puzzle where you don't have the picture on the box, and the pieces are constantly moving. In a normal drug design, scientists usually look at a single protein with a known "pocket" (a hole) and try to fit a drug into it. But molecular glues are different. They work by creating a new surface where two proteins that normally ignore each other suddenly stick together.

The challenge is that before the glue is added, the two proteins are floating separately. You don't know where they will touch, or what shape the glue needs to be to hold them there. It's a chicken-and-egg problem: you need the glue to know where the proteins will meet, but you need to know where they meet to design the glue. Traditional computer methods couldn't handle this because they were built to look at one protein at a time, not three moving parts at once.

The Solution: TriGlue's Two-Step Dance

The authors propose a framework called TriGlue (short for "Ternary Glue," meaning three-part glue). They break the impossible problem down into two manageable steps, mimicking how nature might do it.

Step 1: Guessing the Meeting Spot (Interface Estimation)
First, the computer looks at the two separate proteins (the cleanup crew and the target) and tries to guess where they might touch if a glue were present. Since the actual meeting spot is invisible, TriGlue uses a clever trick. It imagines a "virtual interface"—a set of invisible points floating between the two proteins. To make sure these points make geometric sense, the model wraps them in a mathematical shape called an ellipsoid (think of a slightly squashed ball or an egg shape). This acts like a 3D guidepost, telling the computer, "The glue will likely be somewhere inside this egg-shaped zone." This step doesn't create the glue yet; it just sets the stage.

Step 2: Building the Glue and Moving the Proteins (Complex Generation)
Once the "egg-shaped" meeting zone is defined, the second part of TriGlue kicks in. This is a generative engine that does two things simultaneously:

It designs the glue: It creates a new, tiny molecule that fits perfectly into that virtual zone.
It moves the proteins: It calculates exactly how to rotate and slide the two proteins so they snap together around the new glue.

The authors use a technique called Flow Matching. Imagine a river flowing from a chaotic, messy state (random noise) into a calm, organized lake (the perfect drug structure). The model learns the "current" of this river, guiding the random noise step-by-step until it becomes a valid molecular glue and a stable protein complex.

What They Found

The team tested TriGlue on a massive dataset of known protein complexes (called TernaryDB), which contains over 22,000 examples of how these three-part structures look in real life. They compared their results against other top-tier AI methods.

Better Placement: When the computer generated a new molecular glue, TriGlue placed it in the correct spot more accurately than the other methods. The average distance between the generated glue and the "perfect" spot was about 5.22 Angstroms (a unit of measurement for atoms), which was the lowest error among the competitors.
Stronger Sticking: The generated molecules showed strong signs of being able to bind tightly to the proteins, with high scores in predicted binding affinity.
Better Protein Docking: TriGlue was also better at figuring out how to move the two proteins into the right position to hold the glue. It achieved a "success rate" (where the proteins dock correctly) of 35.2%, beating the next best method which only reached 30.8%.
The "Egg" Matters: The researchers also tested what happened if they removed the "ellipsoid" guide (the egg shape) or the step that corrects the protein positions. Without these, the model struggled significantly. The success rate of docking dropped, and the molecules were placed less accurately. This suggests that the "egg-shaped" guide is crucial for helping the computer understand the geometry of the problem.

The Takeaway

TriGlue suggests that by breaking down the complex problem of molecular glue design into "guessing the meeting spot" and "building the glue while moving the pieces," we can create better drug candidates. The paper doesn't claim to have cured a disease or found a new drug yet; it shows that in computer simulations, this biology-inspired approach can generate chemically valid molecules and plausible protein structures better than current methods. It's a new tool in the toolbox, offering a way to explore the vast, invisible world of molecular glues with more speed and precision than ever before.

Technical Summary: TriGlue

Problem Statement

The paper addresses the computational design of Molecular Glue Degraders (MGDs), a class of therapeutics that induce targeted protein degradation by forming a ternary complex between an E3 ubiquitin ligase, a target protein (POI), and a small molecule "glue." Unlike conventional structure-based drug design, which typically assumes a predefined binding pocket on a single protein, molecular glue design is governed by an unknown protein-protein interface.

The core challenge is a ternary complex generation problem that requires the simultaneous optimization of three tightly coupled components:

The small molecule (ligand): A novel drug-like structure.
The protein-protein interface: The specific geometric region where the ligand mediates the interaction between the two proteins.
The rigid-body transformation: The rotation and translation required to assemble the unbound target protein and the bound receptor into a coherent ternary complex.

Existing AI methods are largely unsuitable because they focus on binary protein-ligand binding or predict structures for known ligands, failing to address the de novo design of the ligand and the complex assembly simultaneously.

Methodology: TriGlue

The authors propose TriGlue, a biology-inspired generative framework that decomposes the complex generation task into two coupled stages: Interface Estimation and Interface-Conditioned Complex Generation. This factorization aligns the computational process with the biological mechanism where the glue-mediated interface guides complex assembly.

1. Biology-Inspired Factorization

The problem is formulated as a conditional distribution $p(C | R, \tilde{T})$ , where $C$ is the ternary complex, $R$ is the bound receptor, and $\tilde{T}$ is the unbound target. The authors factorize this into:
$p(C | R, \tilde{T}) = p(I_R, I_{\tilde{T}} | R, \tilde{T}) \cdot p((R, M, T) | I_R, I_{\tilde{T}}, R, \tilde{T})$
where $I_R$ and $I_{\tilde{T}}$ represent the sets of interface residues.

2. Interface Estimation Module

Since the interface is unknown for unbound monomers, TriGlue infers a Global Virtual Interface and aligns it with geometric priors.

Virtual Interface Points: The model maps each protein to a fixed set of $K$ virtual interface points using attention-weighted aggregation of residue features (encoded via EGNN).
Ellipsoid-Parameterized Alignment: To ensure geometric meaningfulness, the distribution of these virtual points is constrained to match a ground-truth ellipsoid representation of the physical interface. This is achieved by minimizing the discrepancy between the first-order (centroid) and second-order (covariance) moments of the predicted points and the target ellipsoid. This provides a geometrically constrained structural prior.

3. Interface-Conditioned Ternary Flow Network

The second stage jointly generates the molecular glue and the rigid-body transformation using a Flow Matching framework conditioned on the estimated interface. The generation is decomposed into three independent flows:

Rigid Transformation Flow ( $p(R, t | \cdot)$ ): Predicts the rotation and translation to dock the target protein. It utilizes a Differentiable Kabsch Model to generate an initial transformation based on the virtual interfaces, followed by an iterative refinement block that predicts residual updates in the Lie algebra space ($SO(3)$ and $\mathbb{R}^3$ ).
Coordinate Flow ( $p(x | \cdot)$ ): Generates the 3D coordinates of ligand atoms. It first pre-aligns the target protein using the predicted rigid transformation. It then employs a bi-level message passing network to denoise coordinates, capturing both intra-ligand geometric dependencies and inter-molecular interactions between the ligand and the two proteins.
Atom Type Flow ( $p(a | \cdot)$ ): Predicts the chemical identities of ligand atoms by modeling a continuous flow on the probability simplex.

The model uses an Invariant Point Attention (IPA) encoder to create unified latent representations of the receptor, target, and ligand, ensuring invariance to input structure while capturing interactions.

Key Contributions

Novel Problem Formulation: The paper investigates the de novo design of molecular glue-induced ternary complexes, a problem distinct from standard binary docking or single-protein design.
Biology-Inspired Framework: TriGlue decomposes the generation process into interface estimation and interface-conditioned generation, mimicking the biological mechanism of glue action.
Geometric Constraints: The introduction of an ellipsoid-parameterized interface provides a geometrically meaningful prior that guides the generative process without requiring pre-defined binding pockets.
Unified Generative Model: The use of Ternary Flow Matching allows for the simultaneous generation of the small molecule and the prediction of the rigid-body docking pose.

Experimental Results

The authors evaluated TriGlue on the TernaryDB dataset (22,303 ternary complexes), using non-overlapping splits to prevent data leakage.

Molecular Generation: TriGlue generated chemically valid molecules with high drug-likeness (QED: 0.67) and strong predicted binding affinity. It achieved the lowest average RMSD (5.22 Å) for ligand placement compared to baselines like AR, TargetDiff, and PocketXMol, even without access to binding pocket information.
Ternary Complex Reconstruction: In terms of protein docking quality (measured by DockQ), TriGlue achieved a 35.2% success rate (DockQ > 0.23), outperforming DeepTernary (30.8%) and demonstrating a more robust distribution of results with less long-tail failure.
Interface Prediction: The ellipsoid interface estimation module significantly outperformed baselines (PeSTo, ScanNet) across geometry-aware metrics (JSD, Wasserstein-2 distance, Center-L2, LogDet error), confirming the model's ability to capture high-level interface geometry.
Ablation Studies: Removing the ellipsoid prior or the interface correction module led to substantial drops in docking performance (e.g., average DockQ dropped from 0.22 to 0.15 without interface correction), validating the necessity of the geometric constraints and correction mechanisms.

Significance and Claims

The paper claims that TriGlue demonstrates the potential of biology-inspired generative modeling for accelerating molecular glue discovery. By explicitly modeling the unknown protein-protein interface and coupling ligand generation with protein docking, the framework addresses a critical gap in computational drug design. The authors assert that their approach produces chemically valid molecules and structurally plausible ternary complexes, offering a new direction for AI-driven targeted protein degradation. The work highlights that decomposing complex biological mechanisms into tractable generative sub-tasks is a viable strategy for designing complex therapeutic systems.