# Sparsity is All You Need: Rethinking Biological Pathway-Informed Approaches in Deep Learning

**Authors:** Isabella Caranzano, Corrado Pancotti, Cesare Rollo, Flavio Sartori, Pietro Liò, Piero Fariselli, Tiziana Sanavia  
**arXiv:** [2505.04300](https://arxiv.org/abs/2505.04300)  
**Date:** 2025-05-07  
**Summary page:** https://www.summarizepaper.com/en/paper/2505.04300v1/

## Key points

- Compositional sparsity in deep learning models involves breaking down efficiently computable functions into simpler functions that depend on a small subset of inputs.
- Biologically-informed neural networks leverage pathway annotations to enhance performance in biomedical applications.
- A recent study led by Isabella Caranzano challenges the benefits of pathway integration in neural network models for predictive tasks in biomedical settings.
- The study found that models based on randomized information performed as well as biologically informed ones, with some randomized versions even outperforming the latter.
- Pathway-informed models did not show a clear advantage in interpretability compared to their randomized counterparts.
- Current methods may not effectively filter out noise present in pathway annotations, leading to suboptimal performance gains.
- Caranzano and her team propose a novel methodology for systematically comparing pathway-informed models against randomized counterparts across different domains to evaluate the true value of incorporating biological knowledge into deep learning frameworks.

## Summary

In the realm of deep learning models, there is a concept known as "compositional sparsity" where efficiently computable functions can be broken down into simpler functions that depend on only a small subset of inputs. This idea is particularly relevant in the context of biologically-informed neural networks that leverage pathway annotations to enhance performance in biomedical applications. However, a recent study led by Isabella Caranzano and her team challenges the conventional wisdom surrounding the benefits of pathway integration. The research conducted by Caranzano et al. delved into a comprehensive analysis of various pathway-based neural network models used for predictive tasks in biomedical settings. The team critically evaluated each study's contributions and curated a subset of methods with publicly available source code for further investigation. Surprisingly, their findings revealed that models based on randomized information performed just as well as biologically informed ones across different metrics and datasets. Even more intriguing was the discovery that in 3 out of the 15 analyzed models, the randomized versions actually outperformed their biologically informed counterparts. This unexpected outcome raises questions about the true value of incorporating biological pathways into deep learning models. Moreover, pathway-informed models did not exhibit a clear advantage in interpretability compared to their randomized counterparts, as both were able to identify relevant disease biomarkers effectively. These results suggest that current methods may not adequately harness or filter out the noise present in pathway annotations, leading to suboptimal performance gains. As a response to these findings, Caranzano and her team propose a novel methodology that can serve as a robust benchmark for systematically comparing pathway-informed models against their randomized counterparts across different domains. By rigorously evaluating whether observed performance improvements can be attributed to biological insights or simply sparsity introduced by pathways, researchers can gain deeper insights into the efficacy of incorporating biological knowledge into deep learning frameworks. This study challenges existing paradigms and opens up new avenues for refining and optimizing pathway-informed neural network models for future biomedical applications.