DD-03B: Target-level analysis of ligand-protein dissociation pathways reveals ligand-specific exit pathways weakly coupled to binding affinity

arXiv:2604.06761 · physics.comp-ph, physics.chem-ph · Submitted 2026-04-08 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.

Kai: We're looking at the paper: "DD-03B: Target-level analysis of ligand-protein dissociation pathways reveals ligand-specific exit pathways weakly coupled to binding affinity".

Mira: Here's the gist of the paper. Understanding the kinetics of drug-protein interactions is paramount for drug design, yet the field lacks large-scale, dynamic data to move beyond static structural analysis.

Kai: Here, we present DD-03B, a massively scalable database providing dynamic, all-atom dissociation trajectories for a broad set of ligandprotein complexes.

Mira: For these systems—which possess experimental binding affinities (kd) but typically lack measured koff rates—we computed and assigned dissociation rate constants through trajectory reweighting.

Kai: Our analysis reveals that protein-ligand complexes can be categorised into three mechanistic types (pathway-dominant, open-pocket, and entropy-pocket systems), each requiring distinct strategies for accurate kinetic characterisation.

Paper summary: Mira: The paper "DD-03B: Target-level analysis of ligand-protein dissociation pathways reveals ligand-specific exit pathways weakly coupled to binding affinity" introduces DD-03B as this massively scalable database providing dynamic, all atom dissociation trajectories for a broad set of ligand protein complexes <ref:2604.06761#pg1,massively scalable database providing dynamic, all atom dissociation trajectories for a broad>.

Kai: They point out that understanding these kinetics is critical for drug design, but the field lacked large scale dynamic data to go beyond just static structural analysis.

Mira: They created DD-03B by applying a validated computational pipeline to nineteen thousand thirty-seven ligand protein complexes sourced from PDBbind plus version two thousand twenty R1, resulting in approximately zero point three billion simulation frames.

Lev: That expansion is substantial compared to their first phase database, DD-13M, which only covered hundreds of complexes and yielded around six hundred eighty complexes initially.

Kai: The key difference is that DD-03B lacks experimentally measured dissociation rate constants or koff rates; instead, they provide the experimental binding affinities like kd or IC50 for these systems <ref:2604.06761#pg1>.

Mira: This means the prediction challenge shifts; we’re not just matching structures anymore, we have to predict kinetics from thermodynamics.

Lev: The structural coverage is also much wider than before, with the maximum ligand size increasing to three hundred eighty atoms and the maximum protein size expanding to seventy-three thousand seven hundred seventeen atoms.

Kai: So they've scaled up the data significantly while keeping the all atom detail, which is crucial for detailed dynamics.

Mira: They also categorized these systems into five distinct types based on modelling success and trajectory acquisition, including Manual Setup and Single Pathway systems.

Lev: The simulation protocol used an automated high throughput molecular dynamics pipeline called SPONGE to generate these trajectories.

Conclusion: Kai: Looking at the title of this paper, "DD-03B: Target-level analysis of ligand-protein dissociation pathways reveals ligand-specific exit pathways weakly coupled to binding affinity," it suggests a very specific finding <ref:2604.06761#pg1>.

Mira: It suggests that the different ways a drug leaves a pocket—the exit pathways—are not just random; they are specifically tied to the drug itself, and these different ways aren't strongly linked to how tightly it binds, which is what we measure with kd or IC50.

Lev: So what this means for us is that when we train generative AI models to predict dissociation kinetics, we need to know which mechanistic type of system we are looking at before we even start predicting the rate constant.

Kai: The implication here is that the quality and scope of our training data directly determine the accuracy of the kinetic predictions those models can make for new drugs.

Mira: They’re building a foundation, DD-03B, which is intended to be continuously augmented with new trajectories to train next generation generative AI models to predict and optimize drug protein dissociation kinetics <ref:2604.06761#pg1>.

Lev: It sets up a framework where optimal sampling strategies will need to align with those three mechanistic perspectives they identified: explicit pathway, open pocket, and entropy pocket.

Kai: So the ultimate goal is using this massive resource, DD-03B, to develop and benchmark the next generation of predictive models for drug protein dissociation kinetics <ref:2604.06761#pg1>.

Maodong Li, Dechin Chen, Zhijun Pan, Zhe Wang, §Yi Isaac Yang

Institute of Systems and Physical Biology, Shenzhen Bay Laboratory

physics.comp-ph, physics.chem-ph

Submitted: 2026-04-08

Updated: 2026-11-13

Comments: 26 pages, 6 figures

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 77/100

The gist: Understanding the kinetics of drug-protein interactions is paramount for drug design, yet the field lacks large-scale, dynamic data to move beyond static structural analysis.

Key concepts

DD-03B
A large, scalable database containing dynamic simulations of ligand-protein dissociation trajectories for over 19,000 complexes. It provides experimental binding affinities (kd) but allows researchers to calculate dissociation rates (koff) through trajectory reweighting.
Mechanistic Classification
The process of grouping protein-ligand complexes into distinct types based on how the ligand leaves the pocket. These types—pathway-dominant, open-pocket, and entropy-pocket—dictate different strategies needed to accurately study the dissociation kinetics.
Dissociation Rate Constants (koff)
A kinetic parameter that measures how fast a drug molecule leaves its binding site on a protein. While often hard to measure directly, this paper calculates estimated koff values from dynamic simulations, which is crucial for understanding drug efficacy.
AI Model Training Signal
The rich data in DD-03B (trajectories, pathways) is used to train generative AI models. These models learn the complex relationship between a drug's structure and its kinetic behavior (koff/kd), enabling better prediction of how drugs will interact with proteins.

Terminology

Summary

Understanding the kinetics of drug-protein interactions is paramount for drug design, yet the field lacks large-scale, dynamic data to move beyond static structural analysis. p. 0 Here, we present DD-03B, a massively scalable database providing dynamic, all-atom dissociation trajectories for a broad set of ligandprotein complexes. p. 0 For these systems—which possess experimental binding affinities (kd) but typically lack measured koff rates—we computed and assigned dissociation rate constants through trajectory reweighting. p. 0 Our analysis reveals that protein-ligand complexes can be categorised into three mechanistic types (pathway-dominant, open-pocket, and entropy-pocket systems), each requiring distinct strategies for accurate kinetic characterisation. p. 0 Together with our previously released DD-13M, DD-03B forms the core of the expandable Dissociation Dynamic Database (DDD) project, which will be continuously augmented with new trajectories. p. 0 This large-scale, publicly available resource establishes a critical foundation for training and benchmarking next-generation generative AI models to predict and optimise drug-protein dissociation kinetics. p. 0

Database Construction Scale

The DD-03B database was constructed by applying a validated workflow to all 19,037 freely available complex structures from PDBbind+v2020R1, yielding approximately 0.3 billion simulation frames. p. 0 p. 0 This second-phase database was generated by applying our validated workflow to all 19,037 freely available complex structures from PDBbind+v2020R1, yielding approximately 0. p. 23 billion simulation frames. Compared to our first-phase DD‑13M database, the DD-03B database represents a 28-fold expansion in the number of complexes (from 680 to 19,037). p. 3 A key distinction is that DD‑03B lacks experimentally measured dissociation rate constants (koff); instead, it provides commonly available experimental binding affinities (kd or IC50), thereby presenting a distinct prediction challenge. p. 3 The structural coverage is substantially broader, with the maximum ligand size increasing from 322 to 380 atoms and the maximum protein size expanding from 31,855 to 73,176 atoms. p. 3

Simulation Protocol Details

Molecular dynamics simulations were performed using an automated high-throughput molecular dynamics (MD) simulation pipeline based on SPONGE, which accepts standardised input and performs automated system construction, simulation, and termination once ligand escape is detected. Initial docked complexes were prepared using the Python package XPONGE for protein modeling with the AMBER FF14SB14 force field and the ligand with the AMBER GAFF15 force field. p. 3 The simulations employed a 500 ps NVT equilibration at 300 K and a 500 ps NPT equilibration at 1 bar, with heavy atoms restrained during this process. p. 4 Production MetaD-MD simulations were conducted with restraints applied to protein C-α atoms, using the Cartesian coordinates of the ligand centre of mass as the collective variable (CV). p. 4 An adaptive termination criterion was implemented where simulations were terminated immediately when the ligand centre of mass reached the protein solvent‑accessible surface. p. 4

Pathway Analysis and Classification

From the endpoints of 766,550 dissociation trajectories across 15,540 complexes, we projected the ligand centre-of-mass positions onto a defined reaction surface and performed clustering to identify distinct exit channels. Subsequent refinement with the nudged elastic band (NEB) method yielded 19,450 candidate minimum free energy paths (MFEPs) after filtering for short pathways, non-converging pathways, and single-visit clusters. p. 5 The final curated pathway dataset consists of 15,844 reproducible dissociation pathways from 15,540 complexes. p. 5 Complexes in the DD-03B database are systematically classified into five distinct categories based on modelling success and trajectory acquisition. p. 9 These categories include Manual Setup, Shallow Pocket, Short Pathway, Single Pathway, and Multiple Pathway. p. 9

Mechanistic Classification

A comprehensive analysis of ligand dissociation pathways and protein pocket volumes reveals that the unbinding process is governed by at least two fundamental questions: whether the process is driven primarily by ligand egress or by protein pocket reorganisation, and whether it is determined by static structure or by the accessible dynamic space. This leads to three distinct mechanistic classes: explicit pathway, open pocket, and entropy pocket. p. 0 From the ligand-centric perspective, dissociation is described by a welldefined, extended egress pathway which requires the pathway to have a significant length. p. 9 For these systems, such as Single Pathway and a subset of Multiple Pathway systems, path-based collective variable methods are applicable. p. 9 Conversely, for a substantial portion of complexes, even extensive sampling fails to yield a single representative path, suggesting a radial or multi-directional dissociation process.

AI Model Training Signal

The scale and richness of the DD-03B database presents a unique opportunity to advance generative models for dissociation dynamics. This combination of raw atomic trajectories with higher-level structural and energetic labels enables the training of next-generation models that can learn a more comprehensive representation of the dissociation process. p. 1 Models are trained for two critical, interconnected tasks: predicting the dissociation rate constant (koff) and estimating the binding affinity (kd). p. 1 The availability of end-to-end dynamic pathways provides the necessary data to learn the relationship between kinetics and thermodynamics. p. 1 The authors are currently developing an enhanced generative model based on the DD-03B dataset to address these dual objectives. p. 9

Future Directions

Optimal sampling strategies should align with the three primary mechanistic perspectives identified: explicit pathway, open pocket, and entropy pocket. Future methodological refinements will focus on unifying the calculation for these three classes of complexes within an integrated framework of SinkMeta32, by adaptively tuning the sampling precision and strategy to match the specific dissociation mechanism of each system. The DD-03B database provides a foundational resource to develop and benchmark the next generation of predictive models for drug-protein dissociation kinetics. p. 2 This large-scale, open-source dataset is available at https://aimm. p. 6szbl.ac.cn/database/ddd/, inviting the broader AI and computational biophysics community to utilise it. p. 0 The authors are currently developing such an improved generative model based on the DD-03B dataset. p. 9

How the training signal is built

The database provides four core data types: Modelled Structures, AllAtom Trajectories, Unbinding Pathways, and Binding Pocket Angiography (BPA). The Modelled Structures include complete, ready-to-run input files for the SPONGE v1. p. 74 simulation package. p. 3 All-Atom Trajectories are stored in the. p. 7h5md format, including the full atomic detail of the simulated system such as explicit solvent molecules. p. 7 Unbinding Pathways include centroid coordinates (x, y, z) of each clustered dissociation route stored in. p. 7xyz files. Binding Pocket Angiography is stored as a 4D text file (x, y, z, F) visualising the spatial probability distribution of the ligand during its escape. p. 7

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improve generative AI models to predict dissociation kinetics by training them on DD-03B, which provides complete unbinding trajectories paired with derived mechanistic descriptors, including low-dimensional pathway coordinates (x, y, z) and Binding Pocket Angiography (BPA) (x, y, z, F). This combination allows models to be trained for two critical, interconnected tasks: predicting the dissociation rate constant (koff) and estimating the binding affinity (kd).

  2. Develop a system that can quantitatively classify protein-ligand complexes into three distinct mechanistic classes—pathway-dominant, open-pocket, and entropy-pocket systems—to guide specific sampling strategies. This classification is based on whether the process is driven primarily by ligand egress or by protein pocket reorganisation.

  3. Enhance AI models to predict kinetic rates (koff and kd) with quantitative precision by adopting a tailored approach where optimal sampling strategies should align with the three primary mechanistic perspectives identified: explicit pathway, open pocket, and entropy pocket.

Sources

Related papers