Explainable artificial intelligence toward usable and trustworthy computer-aided diagnosis of multiple sclerosis from Optical Coherence Tomography.

BACKGROUND: Several studies indicate that the anterior visual pathway provides information about the dynamics of axonal degeneration in Multiple Sclerosis (MS). Current research in the field is focused on the quest for the most discriminative features among patients and controls and the development of machine learning models that yield computer-aided solutions widely usable in clinical practice. However, most studies are conducted with small samples and the models are used as black boxes. Clinicians should not trust machine learning decisions unless they come with comprehensive and easily understandable explanations. MATERIALS AND METHODS: A total of 216 eyes from 111 healthy controls and 100 eyes from 59 patients with relapsing-remitting MS were enrolled. The feature set was obtained from the thickness of the ganglion cell layer (GCL) and the retinal nerve fiber layer (RNFL). Measurements were acquired by the novel Posterior Pole protocol from Spectralis Optical Coherence Tomography (OCT) device. We compared two black-box methods (gradient boosting and random forests) with a glass-box method (explainable boosting machine). Explainability was studied using SHAP for the black-box methods and the scores of the glass-box method. RESULTS: The best-performing models were obtained for the GCL layer. Explainability pointed out to the temporal location of the GCL layer that is usually broken or thinning in MS and the relationship between low thickness values and high probability of MS, which is coherent with clinical knowledge. CONCLUSIONS: The insights on how to use explainability shown in this work represent a first important step toward a trustworthy computer-aided solution for the diagnosis of MS with OCT.

Machine learning classifier for identification of damaging missense mutations exclusive to human mitochondrial DNA-encoded polypeptides.

BACKGROUND: Several methods have been developed to predict the pathogenicity of missense mutations but none has been specifically designed for classification of variants in mtDNA-encoded polypeptides. Moreover, there is not available curated dataset of neutral and damaging mtDNA missense variants to test the accuracy of predictors. Because mtDNA sequencing of patients suffering mitochondrial diseases is revealing many missense mutations, it is needed to prioritize candidate substitutions for further confirmation. Predictors can be useful as screening tools but their performance must be improved. RESULTS: We have developed a SVM classifier (Mitoclass.1) specific for mtDNA missense variants. Training and validation of the model was executed with 2,835 mtDNA damaging and neutral amino acid substitutions, previously curated by a set of rigorous pathogenicity criteria with high specificity. Each instance is described by a set of three attributes based on evolutionary conservation in Eukaryota of wildtype and mutant amino acids as well as coevolution and a novel evolutionary analysis of specific substitutions belonging to the same domain of mitochondrial polypeptides. Our classifier has performed better than other web-available tested predictors. We checked performance of three broadly used predictors with the total mutations of our curated dataset. PolyPhen-2 showed the best results for a screening proposal with a good sensitivity. Nevertheless, the number of false positive predictions was too high. Our method has an improved sensitivity and better specificity in relation to PolyPhen-2. We also publish predictions for the complete set of 24,201 possible missense variants in the 13 human mtDNA-encoded polypeptides. CONCLUSIONS: Mitoclass.1 allows a better selection of candidate damaging missense variants from mtDNA. A careful search of discriminatory attributes and a training step based on a curated dataset of amino acid substitutions belonging exclusively to human mtDNA genes allows an improved performance. Mitoclass.1 accuracy could be improved in the future when more mtDNA missense substitutions will be available for updating the attributes and retraining the model.

Rebooting the human mitochondrial phylogeny: an automated and scalable methodology with expert knowledge.

BACKGROUND: Mitochondrial DNA is an ideal source of information to conduct evolutionary and phylogenetic studies due to its extraordinary properties and abundance. Many insights can be gained from these, including but not limited to screening genetic variation to identify potentially deleterious mutations. However, such advances require efficient solutions to very difficult computational problems, a need that is hampered by the very plenty of data that confers strength to the analysis. RESULTS: We develop a systematic, automated methodology to overcome these difficulties, building from readily available, public sequence databases to high-quality alignments and phylogenetic trees. Within each stage in an autonomous workflow, outputs are carefully evaluated and outlier detection rules defined to integrate expert knowledge and automated curation, hence avoiding the manual bottleneck found in past approaches to the problem. Using these techniques, we have performed exhaustive updates to the human mitochondrial phylogeny, illustrating the power and computational scalability of our approach, and we have conducted some initial analyses on the resulting phylogenies. CONCLUSIONS: The problem at hand demands careful definition of inputs and adequate algorithmic treatment for its solutions to be realistic and useful. It is possible to define formal rules to address the former requirement by refining inputs directly and through their combination as outputs, and the latter are also of help to ascertain the performance of chosen algorithms. Rules can exploit known or inferred properties of datasets to simplify inputs through partitioning, therefore cutting computational costs and affording work on rapidly growing, otherwise intractable datasets. Although expert guidance may be necessary to assist the learning process, low-risk results can be fully automated and have proved themselves convenient and valuable.