ALyzer3D.AI: a more generalizable deep learning predictor of light chain amyloidogenicity powered by structural and evolutionary Artificial Intelligence.

Abstract

Background: Predicting amyloidogenic risk of immunoglobulin light chains in amyloid light-chain (AL) amyloidosis is a major challenge, and existing computational models fail to generalize to new patient data.

Methods: We developed ALyzer3D.AI, a multi-modal deep learning architecture that integrates evolutionary features from the ESM-2 Protein Language Model, structural metrics from ColabFold and engineered biophysical features. Generalizability was evaluated on published and independent datasets, and model interpretability was assessed using IMGT-aligned SHapley Additive exPlanations (SHAP).

Results: While existing models dropped in performance on new data (accuracy 0.42-0.56), ALyzer3D.AI maintained an accuracy of 0.65 when trained and tested on the same datasets. A final model built on the full combined cohort of 5261 sequences achieved stable performance across repeated splits, with independent test AUC of 0.86 and accuracy of 0.84 at an optimized decision threshold. Performance arose from synergistic integration of PLM, structural and scalar features. PLM attributions were broadly distributed and slightly enriched in framework regions, whereas structural SHAP values were concentrated in CDRs and FR4. The most informative positions had intermediate sequence entropy.

Conclusions: ALyzer3D.AI provides a robust, interpretable tool for predicting light-chain amyloidogenicity that generalizes better to independent datasets than existing methods. It is publicly available via https://colab.research.google.com/github/pcmay/ALyzer3D.AI/blob/main/ALyzer3DAI.ipynb.