Researchers at the Indian Institute of Technology (IIT), Mandi, have developed BioFASTNet (Biomedical Fragmented Attention Spectral Transformer Network), an advanced artificial intelligence framework that enables rapid, accurate and automated interpretation of Fourier Transform Infrared (FTIR) spectra. The innovation aims to simplify molecular analysis by eliminating the need for extensive preprocessing and expert-driven interpretation, paving the way for faster biomedical diagnostics and research.
Study Details and Research Team
The study was carried out by Durgesh Ameta, Hardik Sharma, Dr Praful Hambarde, and Prof Laxmidhar Behera at IIT-Mandi through the Indian Knowledge System and Mental Health Applications (IKSMHA) Centre and the Centre for Artificial Intelligence and Robotics (CAIR). FTIR spectroscopy is widely used in healthcare, pathology, microbiology, and pharmaceutical research for molecular characterisation. However, conventional FTIR analysis often requires multiple preprocessing steps, specialised software, and domain expertise, limiting its use in real-time clinical and point-of-care settings.
How BioFASTNet Works
BioFASTNet addresses these challenges through a lightweight, end-to-end deep learning architecture that directly analyses raw infrared spectra. The model incorporates a Multiresolution Convolutional Feature Extractor (MCFE) to capture both fine absorption peaks and broad spectral patterns, along with a Fragment-wise Attention Module (FAM) that focuses on chemically meaningful regions of the spectrum. This chemistry-informed design significantly improves both prediction accuracy and model interpretability.
Performance Benchmarks
The AI framework demonstrated state-of-the-art performance on two internationally recognised benchmark datasets. On the Functional Group Prediction Dataset comprising 8,272 molecular samples across 21 functional groups, BioFASTNet achieved an impressive 97.81% accuracy and an F1-score of 91.97%, outperforming existing models such as CIRSNet and AggMapNet. It also recorded the best reported F1-score of 36.68% on the highly imbalanced Odor Prediction Dataset containing 3,018 molecules and 109 odor categories.
Potential Applications
According to the researchers, BioFASTNet has significant potential for future applications in cancer tissue characterisation, histopathology, microbial identification, biofluid analysis, pharmaceutical quality assessment, biomedical research, and point-of-care diagnostics. While these applications require further clinical validation, the technology could substantially reduce dependence on expert interpretation and accelerate laboratory workflows.
Future Plans
A spokesperson of IIT-Mandi stated that the research team is now working towards integrating BioFASTNet into smart and handheld FTIR spectrometers for real-time spectral analysis. Currently at the proof-of-concept stage, the technology is designed as a software-based solution that can be integrated with existing FTIR systems without requiring modifications to optical hardware. Future efforts will focus on large-scale clinical validation, cross-platform testing, regulatory compliance, and collaborations with healthcare and industry partners to translate the innovation into practical diagnostic tools.
