24P - Assessment of an AI algorithm to classify germline variants in the ATM cancer predisposition gene

Background

The ATM gene is a widely-known tumour suppressor gene, and pathogenic mutations in ATM are associated with increased cancer susceptibility. Advances in gene sequencing have accelerated genetic variant identification, posing challenges for interpretation. Variants of Uncertain Significance (VUS) constitute a significant proportion of this challenge, requiring comprehensive assessment and more evidence to determine their clinical significance. The use of AI in various clinical fields, is gaining attention, but its lack of transparency complicates trust and validation. This study assesses the performance of a transparent, explainable AI algorithm (https://doi.org/10.3390/cancers15041118) in predicting ATM variant pathogenicity.

Methods

The AI algorithm, based on a graph deep learning model was trained on ClinVar 2020 data. To evaluate its accuracy, 3 distinct test datasets of ATM variants were used. A) Novel variants, recently submitted to ClinVar 2024. B) Reclassified VUS, comprising variants with final classification as pathogenic or benign in ClinVar 2024 that in ClinVar 2020 due to lack of evidence were classified as VUS. C) Dataset of 50 pilot variants, recently described (https://doi.org/10.1016/clinchem/hvaa250), some of which with conflicting interpretation that were resolved by expert consensus.

Results

682 out of 710 variants from dataset A were accurately classified (accuracy = 0.96) highlighting the algorithm's proficiency in classifying the novel variants. Regarding dataset B, 124 out of 128 cases (accuracy = 0.97) were correctly classified remarking the algorithm's capacity to reclassify the VUS. Out of the 50 pilot variants from dataset C, 100% of the classifications aligned with consensus classification with clear significance (29 out of 50) demonstrating algorithm's agreement with experts decisions. For the remaining VUS (21 out of 50), AI could reclassify 95% of them as pathogenic or benign.

Conclusions

The AI effectively classified the novel pathogenic and benign variants, and reclassified previous VUS accurately. Its decision align with those of consensus experts. Additionally, it holds potential to reclassify newly described VUS, that their significance may become apparent in the future.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

Fujitsu Research.

Funding

Fujitsu Research.

Disclosure

N. Bayat, V. Raúl, L. Martínez Gomez: Financial Interests, Institutional, Full or part-time Employment: Fujitsu Research of Europe Limited Sucursal en España. C. Fernández Rozadilla, A. Fernandez Montes, J. Garcia Mata: Financial Interests, Personal, Advisory Role: Fujitsu Research of Europe Limited Sucursal en España. S. Abe, S. Tago, M. Fuji: Financial Interests, Institutional, Full or part-time Employment: Fujitsu Research. S. Georgescu: Financial Interests, Institutional, Coordinating PI: Fujitsu Research of Europe Ltd.