DADL Lab

Topic: NLP, Bioinformatics, Computational Biology

The Protein Language Model project focuses on learning the underlying “language” of proteins by treating amino acid sequences as structured text. Using large-scale pretraining on the UniRef90 dataset, the model applies BPE tokenization and a Transformer-based architecture (similar to BERT) to capture contextual and evolutionary dependencies among residues. By learning rich sequence representations, the pretrained model serves as a foundation for diverse downstream biological tasks. These downstream tasks span multiple dimensions of protein understanding: Protein Structure Prediction – including secondary structure classification, disorder region detection, remote homology recognition, and fold class prediction. Post-Translational Modifications – addressing signal peptide identification, major PTM recognition, and neuropeptide cleavage site prediction. Biophysical Properties – modeling protein fluorescence and stability to support rational protein design. Inverse Protein Folding – generating amino acid sequences that align with desired 3D structures, enabling structure-conditioned protein design. By unifying these tasks under a shared pretrained framework, the Protein Language Model aims to establish a comprehensive foundation for protein sequence understanding, function prediction, and design — bridging the gap between bioinformatics and artificial intelligence.

Topic: NLP, Computer Vision

Team: Mustakin Alam

Recent advancements in computer vision and Natural Language Processing (NLP) have transformed how machines interpret visual and textual information. While traditional image processing techniques rely on pixel-based analysis and feature extraction, and NLP leverages self-attention mechanisms for contextual understanding, the potential of viewing images as a language remains largely unexplored. In this work as a proof-of-concept, we introduce a novel "Vision Language" framework, transforming image pixels into sequences of alphanumeric characters, effectively treating images as "text". This unique representation enables the direct application of NLP techniques to computer vision tasks, such as classification and segmentation, by utilizing text-based models for image analysis. We demonstrate this approach by adapting text classification methods for image classification on benchmark datasets. Furthermore, we construct a domain-specific medical language from the MedMNIST dataset, spanning 11 medical imaging sub-datasets, illustrating how a unified language can effectively generalize across distinct datasets. We create a broad, generalized language from ImageNet, leveraging approximately 1.3 million images to show that a language derived from extensive image data can simplify and unify image processing across tasks and datasets. Our findings reveal a promising cross-disciplinary approach that bridges computer vision and NLP.

Topic: Deep Learning

Team: Mustakin Alam

Deep learning delivers remarkable performance but demands vast computational resources and massive datasets, making training costly and inefficient. To address this, we propose a lightweight, architecture-agnostic method that accelerates training by dynamically removing data batches that contribute less to learning. It tracks the stability of internal activations across epochs and prunes batches whose activation variance shows minimal change, signaling diminishing learning utility. Unlike prior data reduction or distillation methods, it adds no auxiliary networks or extra computation, achieving significant reductions in data usage and training time while maintaining or even improving model accuracy.