OpenPhase: Condition-Aware Exploration of Multicomponent Biosystem Phase-Separating Behavior

ICLR 2026 Conference Submission7835 Authors

Keywords: AI for Biology, Deep learning, Protein Liquid-liquid phase separation, Protein design

TL;DR: Condition-Aware prediction for Biomolecular Phase Behavior

Abstract:

Liquid-liquid phase separation (LLPS) is a fundamental biophysical process
in which biomolecules, such as proteins, DNA and RNA, demix from solution
to form distinct liquid phases. Crucially, experimental conditions, such as
salt, temperature, pH and concentration, profoundly influence LLPS
properties and dynamics, often determining whether phase separation occurs
and modulating the propensity of the resulting biomolecular condensates.
While numerous machine learning methods have been developed for protein
phase behavior prediction, their capabilities are frequently constrained by
the inherent complexity of biosystems and the vast variability of environmental
conditions. Here, we introduce \textbf{OpenPhase}, the first condition-aware
platform for system-level exploration of biomolecular phase behavior.
OpenPhase uniquely provides well-structured and ready-to-use datasets of
experimentally verified phase outcomes coupled with corresponding
conditions. We also formalize three canonical tasks (1) condition-aware phase
outcome prediction, (2) condition inference of LLPS, and (3) conditional phase-separating
system design. These tasks well articulate the interplay between system
components, environmental conditions, and emergent phase properties. For each
task, we propose novel solutions and benchmark them against strong
baseline models. OpenPhase also includes a user-friendly API to facilitate
\textit{in-silico} development and evaluation of novel machine learning
methods for complex biosystem phase behavior modeling.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 7835