Creating interpretable deep learning models to identify species using environmental DNA sequences.

Area of Science:

• Ecology and Conservation Biology
• Bioinformatics and Computational Biology
• Machine Learning and Artificial Intelligence

Background:

• Monitoring species presence is vital for ecosystem conservation and habitat assessment.
• Environmental DNA (eDNA) analysis offers a cost-effective alternative to traditional methods for species detection.
• Current deep learning methods like Convolutional Neural Networks (CNNs) are fast but lack interpretability.

Purpose of the Study:

• To develop an interpretable deep learning framework for eDNA species identification.
• To improve the accuracy and transparency of CNN-based eDNA analysis.
• To visualize distinctive DNA sequences associated with specific species.

Main Methods:

• Utilized the ProtoPNet framework to create a prototype-based, interpretable CNN.
• Introduced a novel skip connection to enhance the interpretability of the original ProtoPNet.
• Evaluated the model on a challenging eDNA dataset, comparing its performance to existing methods.

Main Results:

• The interpretable CNN achieved higher accuracy than previous methods on the eDNA dataset.
• The model successfully visualized species-specific DNA base sequences, aiding in 'fact-checking' predictions.
• Reducing reliance on convolutional output improved both model interpretability and predictive accuracy.

Conclusions:

• An interpretable, prototype-based CNN (ProtoPNet) offers a significant advancement in eDNA analysis.
• Visualizing distinctive DNA sequences enhances the transparency and trustworthiness of deep learning models in ecology.
• This approach holds promise for more accurate and understandable biodiversity monitoring.