![]() |
International Journal of Scientific Research and Engineering Development( International Peer Reviewed Open Access Journal ) ISSN [ Online ] : 2581 - 7175 |

DNA Sequence-Based Disease Classification with Caduceus: A Bi-Directional Equivariant Transformer with GAN Augmentation and SHAP Explainability
π Paper Information
| π Paper Title | DNA Sequence-Based Disease Classification with Caduceus: A Bi-Directional Equivariant Transformer with GAN Augmentation and SHAP Explainability |
| π€ Authors | Prof. Roopa B, Vaishakh M P, Yashvardhan Chougule, Rokkam Akash, Yogeesh G |
| π Published Issue | Volume 9 Issue 6 |
| π Year of Publication | 2026 |
| π Unique Identification Number | IJSRED-V9I6P63 |
π Abstract
Genomic information is increasingly used for diagnosis, preventive medicine and personalized treatment planning. However, converting raw DNA sequences into accurate, interpretable disease predictions remains difficult because genomic data is high-dimensional, sparse and shaped by long-range, non-linear interactions. Disease datasets are also small and imbalanced, particularly for rare diseases, and many deep learning models behave as black boxes that clinicians hesitate to trust. This paper presents a DNA sequence-based disease classification system built on Caduceus, a bi-directional, reverse-complement-equivariant long-range DNA language model. The pretrained Caduceus backbone is fine-tuned with a custom three-layer classification head for both single-disease (binary) and multi-disease (multi-label) prediction. A SeqGAN-style generative module, with an LSTM generator, a CNN discriminator and a rollout mechanism, produces synthetic DNA sequences to reduce class imbalance. Class-weighted binary cross-entropy and threshold tuning further support rare-disease learning. SHAP-based explanations identify the nucleotide positions that influenced each prediction, and a web interface returns predictions, confidence scores, visual explanations and a downloadable report. On the evaluated dataset (sequences of 98β104 bp), Caduceus achieved an F1-Micro of 0.6349 on 15-label classification and 0.5905 on 100-label classification, outperforming DNABERT (0.5408) and a CNNβBiLSTM baseline (0.1168), and reached an F1-Micro of 0.9094 on binary classification.
π How to Cite
Prof. Roopa B, Vaishakh M P, Yashvardhan Chougule, Rokkam Akash, Yogeesh G, "DNA Sequence-Based Disease Classification with Caduceus: A Bi-Directional Equivariant Transformer with GAN Augmentation and SHAP Explainability" International Journal of Scientific Research and Engineering Development, V9(6): Page(457-467) November - December 2026. ISSN: 2581-7175. www.ijsred.com. Published by Scientific and Academic Research Publishing.
π Other Details
