Summary: For years, neuroscientists have prioritized the strongest 10% of brain connectivity signals and dismissed the remainder as “noise.” A new large-scale study shows that the other 90%—the connections typically discarded—can predict behavior just as accurately, and sometimes even better. The results indicate that predictive information is distributed widely across the brain rather than concentrated in a single, dominant network.
The findings reshape how we think about brain-based biomarkers and suggest new directions for diagnosis and treatment in psychiatry and precision medicine.
Key Findings
- Multiple pathways: There are numerous, non-overlapping networks that can predict the same behavior, demonstrating redundancy and functional flexibility across brain systems.
- Psychiatric implications: In conditions like depression, different individuals may use distinct neural pathways to produce similar symptoms or behaviors.
- Therapeutic targets: Focusing only on the strongest networks may miss viable treatment targets. Overlooked circuits could be effective for patients who do not respond to current interventions.
- Rethinking “accuracy”: Strong statistical association does not automatically equal greater biological relevance. Signals once labeled as “noise” may hold important clinical and mechanistic information.
Source: Yale
Why the overlooked signals matter
Neuroimaging aims to reveal the brain mechanisms that underlie cognition, behavior and mental health. But the sheer number of possible brain connections makes analysis difficult: researchers often apply feature selection to narrow down the data to the strongest signals—commonly the top 10% of connections—to build interpretable models.
The new study, published in Nature Human Behavior, questions that practice. Using more than 12,000 participants across four major U.S. datasets and examining both functional and structural connectomes, the team tested whether the connections normally discarded as “noise” might still carry meaningful information about behavior and clinical outcomes.
Study approach
For each participant the researchers measured the association strength between each brain connection (edge) and target behavioral outcomes. They ranked all connections from strongest to weakest and divided them into ten non-overlapping groups. Group 1 held the top 10% of connections—the ones typically selected—while groups 2 through 10 represented the remaining 90% of connections most analyses ignore. The team then built separate prediction models using each group independently.
Surprisingly, models based on lower-ranked groups (groups 2–9) consistently matched or even outperformed models built from the top 10% of connections. This pattern held across multiple cognitive, developmental and psychiatric outcomes and was robust to external validation.
Implications for research and treatment
These results imply that predictive information is distributed broadly across the brain. Narrowing analyses to only the strongest features can simplify modeling but risks oversimplifying the underlying neurobiology. For psychiatric disorders such as depression, this means different people could reach the same behavioral presentation via different neural routes. Consequently, therapeutic strategies that target only the canonical or “loudest” networks may be effective for many but not all patients.
Recognizing and targeting alternative circuits could explain why some individuals are resistant to treatments like certain medications or neuromodulation (for example, TMS) and could expand therapeutic options for those patients.
Clinical biomarkers and precision medicine
Expanding the set of features considered in brain-based models may produce more accurate, individualized biomarkers. Instead of searching for a single, universal “depression signal,” clinicians and researchers could identify which specific pathway or combination of pathways is most relevant for each person—moving toward more precise, personalized interventions.
Key questions answered
A: The brain contains billions of possible connections, so researchers simplify analyses by focusing on the strongest signals to reduce dimensionality and improve interpretability. That practice is practical, but it overlooks that weaker, distributed signals can carry similar predictive information in different formats.
A: Not wrong—more incomplete. Many existing interventions target well-replicated networks that benefit most people. But some patients may have pathology in alternative networks that are overlooked, which helps explain variable treatment response.
A: Potentially yes. Including broader, brain-wide signals in predictive models could improve biomarkers and help clinicians identify which specific neural pathways contribute to an individual’s symptoms, enabling better-tailored therapy.
Editorial notes
- Edited by a Neuroscience News editor.
- Journal paper reviewed in full.
- Additional context provided by editorial staff.
About this mental health and neuroscience research news
Author: Colleen Moriarty
Source: Yale
Contact: Colleen Moriarty – Yale
Image credit: Neuroscience News
Original research: Open access. “Feature selection leads to divergent neurobiological interpretations of brain-based machine learning biomarkers” by Brendan D. Adkinson et al., Nature Human Behavior. DOI: 10.1038/s41562-026-02447-y
Abstract (summary)
Machine learning models trained on neuroimaging data are increasingly used to predict behavioral phenotypes and develop biomarkers for precision medicine. However, the high dimensionality of brain connectivity complicates interpretation. Common practice uses univariate feature selection, implicitly treating the selected feature networks as the unique neurobiological representation of a phenotype while ignoring others.
Using four large-scale datasets with more than 12,000 participants and 13 behavioral outcomes, the study demonstrates that edges discarded by feature selection can still achieve significant prediction accuracy and produce different neurobiological interpretations. These effects extend across cognitive, developmental and psychiatric measures and apply to both functional MRI and diffusion imaging connectomes. The results suggest that focusing exclusively on top-ranked features may present an incomplete view of brain–behavior relationships and that subtle, distributed brain-wide signals deserve greater attention.