Multistate Protein

Protein research spans multiple connected states and representations: understanding and designing proteins requires models to connect sequence, structure, function, interactions, and design. However, most current systems and benchmarks evaluate these problems independently. Strong performance in one state does not necessarily mean that a model can preserve biological constraints, transfer useful information, or support decisions in another.

MuSProt is motivated by this gap. It investigates protein AI as a connected multistate problem rather than a collection of isolated prediction tasks. The aim is to understand how knowledge can be transferred between states, how one prediction influences subsequent decisions, and how a model can remain biologically coherent across the complete workflow.

MusBench provides the evaluation setting for this research. Beyond measuring performance on individual tasks, it is intended to examine cross-state transfer, consistency, robustness, and error propagation. Together, MuSProt and MusBench aim to make multistate protein modelling measurable and to identify where current protein AI systems succeed or break down.