A machine learning framework for autism risk gene discovery using large-scale mouse phenotyping data

Background

Neuropsychiatric disorders such as autism spectrum disorder (ASD) are characterised by complex genetics, diverse pathomechanisms and a lack of objective biomarkers, posing major challenges for diagnosis and treatment. In human studies, behavioural variability and environmental influences limit systematic investigation of genotype–phenotype relationships. In contrast, large-scale mouse phenotyping enables controlled, multidimensional interrogation of gene function in behavioural and physiological domains. The International Mouse Phenotyping Consortium (IMPC) provides standardised datasets for thousands of gene knockouts, representing an extensive phenomic resource to interrogate ASD risk genes and bridge preclinical models with human neuropsychiatric research. The multisystemic nature of IMPC data further enables integrated assessment of brain and peripheral phenotypes, offering a unique opportunity to explore brain – body interactions relevant to neuropsychiatric disease.

Aims

This study aimed to determine whether ASD risk genes exhibit structured phenotypic signatures in large-scale mouse phenotyping data and to evaluate whether machine learning approaches can capture these patterns within a translational framework for ASD gene discovery.

Methods

Machine learning models, including neural networks and random forest, were applied to large-scale IMPC phenotyping datasets. A curated set of ASD risk genes, informed by the SFARI Gene database and including high-confidence genes such asSHANK2, CHD8and NRXN1, was analysed alongside wild-type controls and non-ASD mutant lines. Features spanned neuropsychiatric-relevant domains, including locomotor activity and acoustic startle response (ASR), together with broader phenotypic measures. Model performance was evaluated using accuracy, precision, recall and F1 score. To address class imbalance and biological heterogeneity, threshold adjustment was applied to improve sensitivity to ASD mutant detection, alongside functional stratification of ASD risk genes into biologically defined categories.

Results

Neural network and random forest models achieved comparable accuracy (~0.78 – 0.80), with stronger classification of controls than for ASD mutants. Recall for ASD mutants remained low (~0.10 – 0.15), with precision ~0.60 – 0.62 and F1 scores ~0.18–0.24, reflecting phenotypic heterogeneity and class imbalance. Feature importance analysis identified ASR and locomotor activity as key drivers. However, prediction was not driven by single-feature effects and animals with pronounced behavioural phenotypes were not consistently detected, indicating classification reflects multivariate phenotypic structure. Lowering the classification threshold (0.5 to 0.3) increased sensitivity but also false positives (recall=0.34 at accuracy=0.74), highlighting a trade-off between sensitivity and specificity. Class weighting and functional stratification are being explored to improve performance. Comparison with non-ASD mutant lines enables discrimination of ASD-relevant signatures from general behavioural perturbations.

Conclusions

These findings support the use of large-scale phenotyping to identify structured signatures associated with ASD risk genes. The IMPC provides a uniquely standardised framework, enabling systematic dissection of genotype–phenotype relationships across behavioural and peripheral domains at scale. Ongoing work focuses on refining model performance through functional stratification and integration of multisystem data to better capture ASD-related biology. This framework provides a foundation for systems-level stratification strategies and highlights the translational potential of large-scale phenomics for ASD gene discovery.

References

No references