# Pointwise prediction of protein diffusive properties using machine learning

## TL;DR

This paper introduces M3, a machine learning method that accurately predicts protein diffusion properties from noisy data, outperforming traditional methods.

## Contribution

M3 is a novel machine learning approach for pointwise inference of protein diffusive properties with high accuracy and minimal expert tuning.

## Key findings

- M3 achieves high state accuracies (>90%) for inferring protein diffusion properties.
- The method uses LSTM cells to reduce mean absolute errors in diffusion coefficient and anomalous exponent predictions.
- M3 successfully detects changepoints in protein behavior and finished in the Top 5 of the Anomalous Diffusion Challenge 2024.

## Abstract

The understanding of cellular mechanisms benefits substantially from accurate determination of protein diffusive properties. Prior work in this field primarily focuses on traditional methods, such as mean square displacements, for calculation of protein diffusion coefficients and biological states. This proves difficult and error-prone for proteins undergoing heterogeneous behaviour, particularly in complex environments, limiting the exploration of new biological behaviours. The importance of determining protein diffusion coefficients, anomalous exponents, and biological behaviours led to the Anomalous Diffusion Challenge 2024, exploring machine learning methods to infer these variables in heterogeneous trajectories with time-dependent changepoints. In response to the challenge, we present M3, a machine learning method for pointwise inference of diffusive coefficients, anomalous exponents, and states along noisy heterogenous protein trajectories. M3 makes use of long short-term memory cells to achieve small mean absolute errors for the diffusion coefficient and anomalous exponent alongside high state accuracies (>90%). Subsequently, we implement changepoint detection to determine timepoints at which protein behaviour changes. M3 removes the need for expert fine-tuning required in most conventional statistical methods while being computationally inexpensive to train. The model finished in the Top 5 of the Anomalous Diffusive Challenge 2024, with small improvements made since challenge closure.