Protein-Based Fish Species Identification: Dataset, Models, and Insights from Native Bangladeshi Fish

A research team has released the first curated protein-sequence dataset for nine native Bangladeshi fish species and proposes a lightweight hybrid model that achieves 79.80% classification accuracy, offering a practical alternative to far larger protein language models for resource-limited settings [1]. The study, submitted on 16 June 2026, addresses what its authors describe as a complete absence of benchmarks for identifying native Bangladeshi fish from protein sequences — a gap with direct consequences for food security and biodiversity monitoring in the region [1]. The dataset comprises 2,845 high-quality protein sequences spanning nine species, and the paper establishes the first classification baselines by systematically testing seven architectural paradigms [2]. The novel architecture, called MotifCNN-Transformer with Terminal-Aware Positional-Encoding (MotifCNN-Transformer+TA-PE), reached 79.80% accuracy with a macro-F1 of 0.80 [1]. A fine-tuned version of the protein language model ProtBERT scored higher at 83.04%, but the 3.24-percentage-point margin was statistically insignificant under McNemar’s test, with a p-value of 0.1120 [2]. The proposed model outperformed ProtBERT on six of the nine individual species classes [1]. Beyond raw accuracy, the researchers emphasize deployment practicality. ProtBERT contains 420 million parameters and requires dual 16 GB GPUs for inference, whereas the MotifCNN-Transformer+TA-PE is roughly 42 times smaller, approximately five times faster, supports a 16-times larger batch size, and can run without GPU acceleration [2]. The authors argue this makes it suited for use in rural Bangladesh, where computational infrastructure is thin [1]. The work also examines how phylogenetic relationships influence sequence similarity, laying groundwork for fisheries management, food authentication, and conservation across South Asia’s protein-dependent economy [2]. Bangladesh’s scientific capacity has drawn international notice in related fields. In 2026, BRAC University dean Mahbubul Alam Majumdar received the Spirit of Abdus Salam Award for strengthening the country’s scientific and educational foundations, and was also awarded the Ekushey Padak in education [4]. The country’s broader development challenges — including infrastructure strain documented in Rohingya refugee camps and efforts to improve economic complexity — underscore the value of tools that function outside well-equipped laboratories [5][6]. The fish-identification study does not claim to surpass large language models on accuracy alone; instead it offers a statistically comparable, deployable alternative for regions where high-end computing remains scarce [1][2].