Md Nurul Muttakin

About

Md Nurul Muttakin completed his MSc in Computer Science at KAUST in 2023 under the supervision of Robert Hoehndorf. His research focused on machine learning methods for predicting protein function from three-dimensional structural information.

His thesis, 3D conformation-based protein function prediction, investigated how structural information should be exploited by deep learning models in light of the availability of near-experimental-accuracy structures from AlphaFold. Almost all existing structure-aware protein function predictors combine graph neural networks (GNNs) operating on 2D contact maps with rich 1D sequence features from language models such as ESM or from LSTM embeddings. Muttakin's analysis showed that these 1D features in fact obfuscate the contribution of the structural signal, and that GNNs as currently used struggle to learn structural motifs from contact graphs in isolation.

Building on this finding, the thesis explored alternative ways to extract structure-aware signals, including 2D contact maps derived from pairwise distances, 3D convolutions on molecular point clouds, and learnable graph kernels on contact graphs. The work contributes to the group's research line on protein function prediction and provides a critical re-evaluation of how much of the apparent success of recent structure-based predictors actually depends on structure rather than on co-supplied sequence features.