🧬 biology

Machine Learning Models Reveal the Role of Ionization-Dependent Partitioning in Condensate Formation

By applying machine learning models to four representative biomolecular condensates, this study demonstrates that the pH-dependent distribution coefficient (logD) is the dominant predictor of small-molecule partitioning, establishing ionization-coupled hydrophobicity as a key mechanistic driver of condensate localization.

Original authors: Masoumeh Ozmaian, Seyyed Saeed Vaezzadeh

Published 2026-09-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Masoumeh Ozmaian, Seyyed Saeed Vaezzadeh

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Inside every living cell, the space is not empty; it is a bustling, crowded city of molecules. To keep order in this chaos, cells build temporary, membrane-less compartments called biomolecular condensates. Think of these as distinct neighborhoods where specific groups of molecules gather together, separating from the rest of the cellular fluid to perform specialized tasks. These neighborhoods form through a process where molecules stick to one another, creating a dense liquid phase that floats within the cell. This organization is vital for life, but it can also go wrong. When these liquid compartments turn into solid, sticky clumps, they are linked to serious diseases like Alzheimer's and Parkinson's. Understanding how small molecules decide to enter these condensates is a key puzzle for scientists, as it could help design new drugs that either stabilize these compartments or prevent them from turning solid.

For a long time, researchers believed that the main factor determining whether a small molecule would enter a condensate was simply how oily or water-loving it was. They thought that if a molecule was sufficiently "oily," it would naturally drift into the dense, oily interior of the condensate, much like oil separates from water. However, a new study by Masoumeh Ozmaian and Seyyed Saeed Vaezzadeh suggests this picture is incomplete. The researchers used advanced computer models to analyze how small molecules interact with four different types of condensates found in cells. They discovered that the molecule's behavior changes depending on its electrical charge, which shifts based on the acidity of its environment. This electrical state, combined with how oily the molecule is, creates a more accurate picture of where the molecule will go.

The team turned to machine learning, a type of computer program that learns patterns from data, to solve this problem. They fed the computer information about thousands of small molecules, including their size, shape, and chemical properties. The computer was trained to predict how strongly each molecule would stick to the four different condensate types. Initially, the researchers tested the computer using only the standard measure of oiliness, which assumes the molecule is neutral and uncharged. The computer did a decent job, confirming that oiliness and how well a molecule dissolves in water are indeed important. But the predictions were not perfect.

The breakthrough came when the researchers added a specific new piece of information to the computer's training: a measure of how oily the molecule is at the specific acidity found inside human cells. Unlike the standard measure, this new value accounts for the fact that many molecules gain or lose electrical charges in the body. When the computer learned to use this charge-sensitive measure, its predictions became significantly more accurate. In fact, this single factor became the most powerful clue the computer had for guessing where a molecule would end up. The study showed that for certain condensates, including this charge-sensitive measure improved the model's ability to predict outcomes by a noticeable margin, making it a far better guide than oiliness alone.

To ensure their findings were robust, the researchers also tested whether the three-dimensional shape of the molecules mattered. They built complex models that considered the exact 3D geometry of every molecule, including how it twists and turns. Surprisingly, adding these detailed shape descriptions did not make the computer any better at predicting the results. The simpler, two-dimensional chemical properties, when combined with the charge-sensitive oiliness measure, were enough to capture the essential rules of the game. This suggests that the primary drivers of these interactions are the broad chemical forces of attraction and repulsion, rather than a perfect lock-and-key fit of shapes.

The researchers also built a simpler tool to classify molecules as either "sticking" or "not sticking" to condensates. Even with just the charge-sensitive oiliness measure, this simple tool performed remarkably well, correctly identifying most molecules. When they combined this with other chemical details, the tool became even more reliable. The results consistently pointed to one conclusion: the electrical charge of a molecule, which changes with the environment, is a central switch that controls whether it enters a condensate. This finding offers a clear, practical path forward for scientists designing new medicines. Instead of just looking for oily molecules, they can now tune the electrical charge of a drug candidate to ensure it finds its way into the specific cellular neighborhoods where it is needed.

This work does not just refine a theory; it provides a concrete, data-driven framework for understanding the invisible rules that govern the microscopic world inside our cells. By showing that the electrical state of a molecule is just as important as its oiliness, the study reveals a hidden layer of control in how cells organize themselves. It suggests that the key to unlocking the behavior of these dynamic compartments lies in understanding how molecules change their nature in response to their surroundings. For the future of drug design, this means that the most effective way to target these cellular neighborhoods may be to carefully balance a molecule's charge and its ability to dissolve, ensuring it arrives exactly where it is needed to keep the cell healthy.

Technical Summary: Machine Learning Models Reveal the Role of Ionization-Dependent Partitioning in Condensate Formation

Problem Statement
Biomolecular condensates, formed via liquid–liquid phase separation (LLPS), organize cellular components through non-covalent interactions. While previous research has established that hydrophobicity and solubility are dominant factors in small-molecule partitioning into these condensates, the specific role of ionization states remains incompletely understood. Unlike traditional drug targets with defined binding pockets, condensates engage guest molecules through nonspecific, multivalent interactions in a dynamic environment. Existing high-throughput data suggests that physicochemical properties drive partitioning, but it is unclear whether static measures of lipophilicity (logP) or solubility (logS) are sufficient, or if pH-dependent effective lipophilicity (logD) plays a more critical mechanistic role in determining molecular localization within the distinct, low-dielectric environment of condensates.

Methodology
The study employed a machine learning (ML) framework to analyze experimental partitioning data across four representative condensate systems: cGAS-DNA, SUMO-SIM, SH3-PRM, and DHH1. The dataset was derived from previous high-throughput studies (specifically Ambadi Thody et al.).

Feature Engineering: Molecular descriptors were generated using RDKit. The baseline set included 208 two-dimensional (2D) physicochemical descriptors (constitutional, topological, electronic, and surface area properties). Three specific empirical descriptors were added: ESOL_LogS (estimated aqueous solubility), AromaticProportion, and SlogP_VSA_High. Crucially, the study incorporated logD (calculated at pH 7.4 using ADMETlab 3.0) to account for ionization-dependent partitioning. Additionally, Extended Connectivity Fingerprints (ECFP4) were included.
Model Architecture: Regularized XGBoost (gradient-boosted decision trees) models were trained for both regression (predicting mean partition coefficients) and binary classification (distinguishing partitioning vs. non-partitioning molecules).
Hyperparameter Optimization: A randomized search with five-fold cross-validation was used to optimize hyperparameters (e.g., tree depth, learning rate, regularization terms) to prevent overfitting.
Interpretability: SHapley Additive exPlanations (SHAP) were utilized to quantify feature importance and understand the contribution of individual descriptors to model predictions.
Comparative Analysis: The study compared models trained with and without logD. Furthermore, a separate experiment tested the inclusion of 3D shape-based descriptors (generated via ETKDG conformers and MMFF94 minimization) to determine if molecular geometry provided additional predictive power beyond 2D features.

Key Results

Dominance of logD: When logD was included in the feature set, it consistently emerged as the single most influential descriptor across all four condensate systems, surpassing logP and logS in SHAP importance rankings.
Predictive Performance: The inclusion of logD significantly improved model performance. For the DHH1 and SH3-PRM systems, the coefficient of determination ( $R^2$ ) increased by approximately 0.06–0.07, with corresponding reductions in Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE). These improvements were statistically significant ( $p < 0.01$ ) across 20 independent train-test splits.
Correlation with Partitioning: A clear positive correlation was observed between mean log partition coefficient (log PC) and logD, indicating that molecules with higher effective lipophilicity at physiological pH partition more strongly into condensates.
Limited Value of 3D Descriptors: The addition of 3D geometric descriptors (e.g., moments of inertia, radius of gyration, 3D polar surface area) did not improve predictive performance over the 2D descriptor set. While these features offered mechanistic interpretability regarding molecular shape and polarity, they were not necessary for accurate prediction.
Classification Performance: In binary classification tasks, the full XGBoost model achieved an accuracy of 0.86 and an F1 score of 0.71. Notably, a single-feature baseline using only logD achieved an accuracy of 0.80 and F1 of 0.65, demonstrating that ionization-dependent partitioning encapsulates the majority of the discriminative signal for condensate affinity.

Significance and Claims
The authors claim that this study establishes logD as a mechanistic link connecting ionization, hydrophobicity, and small-molecule partitioning in biomolecular condensates. The findings suggest that the protonation state of small molecules directly modulates their desolvation penalty and transfer free energy within the reduced dielectric environment of condensates, making pH-dependent lipophilicity a more thermodynamically relevant descriptor than neutral-state logP.

The paper posits that 2D physicochemical features combined with logD are sufficient to capture the main determinants of phase separation behavior, rendering complex 3D structural descriptors unnecessary for predictive modeling in this context. Consequently, the study offers a predictive framework for understanding small-molecule behavior in dynamic condensate environments and suggests that tuning ionization equilibria (e.g., pKa) could serve as a practical strategy for the rational design of chemical modulators targeting condensate-associated pathways. The work reinforces the utility of interpretable machine learning models built on physically motivated descriptors to uncover thermodynamic contributions in complex biological systems.

Technical Summary: Machine Learning Models Reveal the Role of Ionization-Dependent Partitioning in Condensate Formation

More like this