FLAN: feature-wise latent additive neural models for biological applications

Area of Science:

• Machine Learning
• Computational Biology
• Bioinformatics

Background:

• Deep learning models offer high performance but often lack interpretability, posing challenges in critical applications like healthcare and legal systems.
• Understanding algorithmic decisions is crucial due to their potential long-lasting effects on individuals.
• Existing interpretable models may not match the performance of complex deep learning architectures.

Purpose of the Study:

• To propose a novel class of deep neural networks, Feature-wise Latent Additive Networks (FLAN), designed for enhanced interpretability.
• To develop a model that processes each input feature independently, allowing for the estimation of individual feature effects.
• To demonstrate the utility of FLAN in biological domains and beyond, balancing performance with interpretability.

Main Methods:

• FLAN processes each input feature separately, generating a unique latent representation for each.
• These feature-wise latent representations are aggregated through summation before the final prediction.
• The structural constraint of FLAN facilitates the interpretation of its decision-making process.

Main Results:

• FLAN achieves competitive performance on benchmark datasets, including complex biological tasks like TCR-epitope binding prediction.
• The inherent interpretability of FLAN allows for deciphering decision processes and extracting biological insights, such as identifying marker genes.
• FLAN demonstrates comparable performance on non-biological datasets, indicating broader applicability.

Conclusions:

• FLAN offers a promising solution for interpretable deep learning, particularly in domains requiring transparency and accountability.
• The model's ability to isolate feature effects provides a mechanism for understanding complex AI decisions.
• FLAN successfully integrates high performance with interpretability, paving the way for more trustworthy AI systems.