Convergent transcriptomic and neuroimaging signature of Autism Spectrum Disorder

Elisa Ferrari, Alessandra Retico, Alessandro Cellerino, Davide BacciuarXiv 2025 · DOI · arXiv

Abstract

Autism Spectrum Disorder (ASD) is a multi-factorial neurodevelopmental disorder with causes that remain poorly understood. While early behavioral interventions can significantly improve patient outcomes, diagnosis often occurs late due to the subjective nature of clinical assessments. This research presents an imaging transcriptomics approach to identify a reliable, objective biomarker for ASD. A bias-resilient deep learning classifier was developed to analyze resting-state functional MRI (rsfMRI) data, achieving an AUC of 0.89 on an independent test set. Concurrently, five gene network modules involved in ASD were identified through post-mortem brain transcriptomics. By mapping these findings, the study demonstrates that the brain regions most relevant to the MRI classifier are the same regions where ASD-linked gene modules show significantly high expression. These results provide a neuroimaging-based biomarker for ASD that is mutually validated by transcriptomic data.

Introduction

The search for a quantitative biomarker for Autism Spectrum Disorder (ASD) is driven by the need for early, objective diagnosis to facilitate timely intervention. Magnetic Resonance Imaging (MRI) is a promising candidate, yet machine learning models in this domain are frequently compromised by "confounders"—biases related to age, sex, or scanner site—that lead to poor generalizability. While transcriptomics has identified reproducible biological signatures in post-mortem tissue, these findings cannot be directly applied to diagnosis in living patients. This study bridges this gap by developing a bias-resilient deep learning model and validating its learned spatial patterns against brain-wide gene expression profiles.

Methods

The research utilizes a three-stage "imaging transcriptomics" pipeline:

1. Debiased Deep Learning for Neuroimaging

A Two-Stream Convolutional Neural Network (2S-CNN) was designed to process 4D rsfMRI data from the ABIDE I & II datasets (n=1,423). The model captures temporal dynamics by analyzing both raw video frames and optical flow. To mitigate the impact of confounders, a Bias-Resilient Neural Network (BR-NN) was employed. This adversarial setup forces the primary classifier to ignore non-diagnostic features (such as scanner site or "eyes open/closed" status) by penalizing it if a secondary network can predict these confounders from the extracted features. Explainability was achieved using SmoothGrad to isolate the 19 brain regions most critical for classification.

2. Transcriptomic Analysis

Post-mortem brain transcriptomic data (48 ASD, 49 controls) were analyzed using Linear Mixed Models and Weighted Gene Co-Expression Network Analysis (WGCNA). This identified five gene co-expression modules significantly dysregulated in ASD, associated with ribosomal activity, immune response, and synaptic transmission.

3. Spatial Convergence

The spatial expression of the identified gene modules was mapped using the Allen Human Brain Atlas. The study then tested the overlap between these genetic expression patterns and the 19 brain regions prioritized by the MRI classifier.

Key Results

The integrated pipeline yielded several breakthrough findings:

  • High Diagnostic Accuracy: The debiased classifier achieved an AUC of 0.89, with 86% accuracy, 88% sensitivity, and 85% specificity on an independent test set, outperforming the literature average of ~70%.
  • Biological Validation of AI: The 19 brain regions identified by the AI (including the amygdala and insula) perfectly aligned with the spatial expression patterns of 4 out of 5 ASD-linked gene modules. This spatial alignment validates the AI's "black box" decisions as biologically grounded.
  • Confounder Resilience: The adversarial training successfully removed biases, including the significant "eyes open/closed" confounder, which was found to have a stronger influence on raw data than patient sex.
  • Regional Sensitivity: Genetic analysis revealed that comparing expression differences between brain regions (e.g., Frontal vs. Cerebellum) is more sensitive for identifying ASD-linked genes than single-region analysis.

Discussion

This research demonstrates that a convergent biomarker for ASD exists across neuroimaging and transcriptomic scales. By successfully decoupling diagnostic signatures from dataset biases, the study provides a more robust framework for objective diagnosis.

Future work should focus on the specificity of the classifier against other neurodevelopmental disorders and its performance in infant populations to support early intervention. Additionally, while the spatial mapping relies on healthy adult brain atlases, the strong convergence observed suggests that these baseline patterns are relevant to the ASD phenotype. This multi-modal approach paves the way for a deeper understanding of the biological underpinnings of ASD and more reliable diagnostic tools.

Summary updated

Abstract

Autism Spectrum Disorder (ASD) is a multi-factorial neurodevelopmental disorder, whose causes are still poorly understood. Effective therapies to reduce all the heterogeneous symptoms of the disorder do not exists yet, but behavioural programs started at a very young age may improve the quality of life of the patients. For this reason, many efforts have been dedicated to the research of a reliable biomarker for early diagnosis. Machine learning approaches to distinguish ASDs from healthy controls based on their brain Magnetic Resonance Images (MRIs) have been plagued by the problem of confounders, showing poor classification performance and inconsistency in the biomarker definition. Brain transcriptomics studies, instead, showed some converging results, but being based on data that can be acquired only post-mortem they are not useful for diagnosis. In this work, using an imaging transcriptomics approach, the following results have been obtained. 1) A deep learning based classifier resilient to confounders and able to exploit the temporal dimension of resting state functional MRIs has been developed, reaching an AUC of 0.89 on an independent test set. 2) Five gene network modules involved in ASD have been identified, by analyzing brain transcriptomics data of subjects with ASD and healthy controls. 3) By comparing the brain regions relevant for the classifier obtained in the first step and the brain-wide gene expression profiles of the modules of interest obtained in the second step, it has been proved that the regions that characterize ASD brain at the neuroimaging level are those in which four out of the five gene modules take a significantly high absolute value of expression. These results prove that, despite the heterogeneity of the disorder, it is possible to identify a neuroimaging-based biomarker of ASD, confirmed by transcriptomics.
BibTeX
@misc{ferrari2025convergent,
  author = {Elisa Ferrari and Alessandra Retico and Alessandro Cellerino and Davide Bacciu},
  title = {{Convergent transcriptomic and neuroimaging signature of Autism Spectrum Disorder}},
  year = {2025},
  eprint = {2505.23776},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2505.23776},
  howpublished = {arXiv},
}
Figure