Artificial intelligence represents a network of algorithms, mimicking the human brain’s neural network, created to reason, learn, and solve problems. Such a system can produce an output normally requiring human intelligence. AI became wildly popular, from targeted online marketing, chatbots, online search engines, and recommender systems, to hardware systems like robots and autonomous vehicles.

So it comes as no surprise that artificial intelligence has its place in genomic and medical diagnostics too.

After the completion of the Human Genome Project in 2003, the innovation in sequencing technology progressed rapidly. With the arrival of so-called next-generation sequencing, the cost of whole genome sequencing has decreased significantly, making it possible to sequence the entire human genome in less than two weeks for under $1000 (Behjati & Tarpey, 2013). Such rapid progress in technology resulted in a massive increase in data that could be shared with scientists all over the world. This new era called for more powerful algorithms as well as databases to tackle the abundance of data.

Traditionally, statistical models were used to help us understand the biological systems and to determine the relationship between, for example, gene dysregulation and disease. But let’s say we want to predict, whether someone is likely to contract a certain disease. There are, of course, statistical models that can be used for prediction too, but this is not their strong suit. Machine learning (ML) models, on the other hand, are designed to deliver predictions with high accuracy, however, they can be difficult to interpret. ML models are first trained on a subset of data and then validated on a test set. These models can handle more complex data with a larger number of variables and sample sizes, by finding patterns and nonlinear interactions within the data.

 

 

Tools like DeepVariant from Google Health were developed to identify individual genetic variants from next-generation sequencing data. This convolutional neural network (CNN) based algorithm is trained directly on read alignments and was shown to outperform some of the existing tools (Poplin et al., 2018).

PrimateAI-3D is a deep-learning network trained on common genetic variants from primate species to identify benign missense variants in humans – which then helps to highlight pathogenic mutations (Gao et al., 2023).

Similarly, tools that specialize in identifying variants in noncoding parts of the genome were developed. SpliceAI is a deep neural network trained on pre-mRNA sequences identifying cryptic splice mutations (Jaganathan et al., 2019).

Epigenetics is another field that can benefit from AI algorithms. Tools like DeepSEA can accurately predict transcription factor binding, DNase hypersensitive sites, and histone marks (Zhou et al., 2015).

Another deep learning-based framework, ExPecto, can predict gene expression from DNA-sequence and the tissue-specific transcriptional effects (Zhou et al., 2018).

Many tools have become popular in the clinical setting as well, used to assist physicians in diagnostics.

For example, tools like DeepGestalt are using face recognition to help identify genetic syndromes of certain genetic conditions, like Noonan syndrome (Gurovich et al., 2019). Image recognition is widely used in analyzing X-rays, MRIs, ultrasounds or CT scans more accurately and quickly in various cancer diagnostics.

Time series analysis can be used with a computerized electrocardiogram to help automate the analysis and help identify arrhythmia (Hannun et al., 2019). Speech recognition has been successfully used to diagnose diseases with an effect on speech like chronic pharyngitis (Li et al., 2019), but even less obvious like Alzhaimer’s (Fraser et al., 2016) and Parkinson’s disease (Zhan et al., 2018), major depressive disorder (Ringeval et al., 2019), posttraumatic stress disorder (Marmar et al., 2019), and even coronary artery disease (Maor et al., 2018).

Using computer vision can help diagnose Parkinson’s disease by quantification of bradykinesia from a video taken at the physician’s office (Gareth et al., 2023). Survival convolutional neural networks (SCNNs) use histological features associated with survival together with somatic mutations to predict outcomes in cancer patients (Mobadersany et al., 2018). And the list of applications goes on.

 

 

Artificial intelligence has undoubtedly many advantages in the field of precision medicine – by eliminating human errors, decreasing the costs of diagnosis, and producing faster, more accurate results. But as with every new technology we have to be careful when interpreting results from these models and ensure their explainability and transparency. We have to understand the biological, chemical, and physical mechanisms that influence the decision-making process and be able to confidently interpret the results produced by these models. Especially if the ML algorithms make decisions that can have a profound influence on the quality of human life. This poses a problem, as for many AI is a “black box” and it is difficult to explain how the algorithm arrived at its conclusion. This is a challenge for clinicians as, for example, under the current GDPR laws the user has the “right to explain” (European Parliament and Council of the European Union, 2016).

Another caveat with AI models being used in genomic and clinical applications is that the majority of the existing tools are supervised AI models that require already verified datasets for training and validation. But there are known biases presented in current research and training AI models on these datasets can further gap the health benefits for underrepresented racial and gender populations.

Early detection and prevention are crucial in diagnostics. For example, we use screenings at certain age thresholds to monitor types of cancer. The challenge is to be able to identify individuals at risk fast and offer more precise diagnosis and treatment.

Polygenic risk scores (PRSs) are a “weighted sum of the number of risk alleles and individual carries” (Lewic and Vassos, 2020). In other words, PRSs estimate an individual’s risk of a specific condition or disease based on their genetic profile. PRSs have the potential to be great biomarkers, however, they have a poor generalisability in populations with different ancestries because of different allele frequencies, linkage disequilibrium, and different genetic effects. (Wang et al., 2022).

PRSs created by AI in combination with Electronic health records (EHRs) or gene expression data, on top of currently used SNPs, can provide a more complex picture. Moreover, access to population EHR and information about patients’ race, gender, age, or socioeconomic status can help to create more representative models (Fritzsche et al., 2023). Yet another area where AI has a great potential to improve the quality of human life, but comes with a set of ethical challenges.

The usage of patients’ data requires better legal and health policies, focused on data protection, patient privacy, and what informed consent should look like. It is important to understand to what degree can these high-dimensional genetic data be anonymized to ensure privacy and autonomy and avoid a risk of genomic identifiability (Kaya et al., 2023 ). Even though AI has a lot of promising applications, its use should augment rather than replace the work of physicians to ensure a transparent and understandable explanation of the diagnoses to patients.

Without a doubt, AI will have a significant effect on healthcare, improving patient outcomes by helping physicians make faster and more accurate decisions. However, as this brief review outlines, there are challenges that need to be addressed to ensure proper regulation and transparency of these methods to avoid any malpractice, misuse, or misinterpretation.

 

Reference list

Behjati, S. and Tarpey, P.S. (2013). What Is next Generation sequencing? Archives of Disease in Childhood – Education & Practice Edition, [online] 98(6), pp.236–238. doi:https://doi.org/10.1136/archdischild-2013-304340.

Border, R., Athanasiadis, G., Buil, A., Schork, A.J., Cai, N., Young, A.I., Werge, T., Flint, J., Kendler, K.S., Sankararaman, S., Dahl, A.W. and Zaitlen, N.A. (2022). Cross-trait assortative mating is widespread and inflates genetic correlation estimates. Science, 378(6621), pp.754–761. doi:https://doi.org/10.1126/science.abo2059.

Choi, S.W., Mak, T.S.-H. and O’Reilly, P.F. (2020). Tutorial: a guide to performing polygenic risk score analyses. Nature Protocols, 15(9), pp.2759–2772. doi:https://doi.org/10.1038/s41596-020-0353-1.

European Parliament, Council of the European Union (2016). Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data, and repealing Directive 95/46/EC (General Data Protection Regulation). Official J. L 119, 1–88.

Fraser, K.C., Meltzer, J.A. and Rudzicz, F. (2015). Linguistic Features Identify Alzheimer’s Disease in Narrative Speech. Journal of Alzheimer’s Disease, [online] 49(2), pp.407–422. doi:https://doi.org/10.3233/jad-150520.

Fritzsche, M.-C., Akyüz, K., Cano Abadía, M., McLennan, S., Marttinen, P., Mayrhofer, M.Th. and Buyx, A.M. (2023). Ethical layering in AI-driven polygenic risk scores—New complexities, new challenges. Frontiers in Genetics, 14. doi:https://doi.org/10.3389/fgene.2023.1098439.

Gao, H., Hamp, T., Ede, J.M., Schraiber, J.G., McRae, J.F., Moriel Singer-Berk, Yang, Y., Anastasia, Petko Fiziev, Lukas, Sundaram, L., Wu, Y., Adhikari, A.N., Field, Y., Chen, C., Serafim Batzoglou, François Aguet, Lemire, G., Reimers, R. and Balick, D.J. (2023). The landscape of tolerated genetic variation in humans and primates. Science, 380(6648). doi:https://doi.org/10.1126/science.abn8197.

Gareth Morinan, Yuriy Dushin, Grzegorz Sarapata, Rupprechter, S., Peng, Y., Girges, C., Salazar, M., Milabo, C., Sibley, K., Foltynie, T., Ioana Cociasu, Ricciardi, L., Baig, F., Morgante, F., Leyland, L.-A., Weil, R.S., Ro’ee Gilron and O’Keeffe, J. (2023). Computer vision quantification of whole-body Parkinsonian bradykinesia using a large multi-site population. npj Parkinson’s disease, 9(1). doi:https://doi.org/10.1038/s41531-023-00454-8.

Gurovich, Y., Hanani, Y., Bar, O., Nadav, G., Fleischer, N., Gelbman, D., Basel-Salmon, L., Krawitz, P.M., Kamphausen, S.B., Zenker, M., Bird, L.M. and Gripp, K.W. (2019). Identifying facial phenotypes of genetic disorders using deep learning. Nature Medicine, [online] 25(1), pp.60–64. doi:https://doi.org/10.1038/s41591-018-0279-0.

Hannun, A.Y., Rajpurkar, P., Haghpanahi, M., Tison, G.H., Bourn, C., Turakhia, M.P. and Ng, A.Y. (2019). Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network. Nature Medicine, [online] 25(1), pp.65–69. doi:https://doi.org/10.1038/s41591-018-0268-3.

Jaganathan, K., Kyriazopoulou Panagiotopoulou, S., McRae, J.F., Darbandi, S.F., Knowles, D., Li, Y.I., Kosmicki, J.A., Arbelaez, J., Cui, W., Schwartz, G.B., Chow, E.D., Kanterakis, E., Gao, H., Kia, A., Batzoglou, S., Sanders, S.J. and Farh, K.K.-H. (2019). Predicting Splicing from Primary Sequence with Deep Learning. Cell, [online] 176(3), pp.535-548.e24. doi:https://doi.org/10.1016/j.cell.2018.12.015.

Kaya Akyüz, Goisauf, M., Gauthier Chassang, Kozera, Ł., Mezinska, S., Tzortzatou, O. and Michaela Th. Mayrhofer (2023). Post-identifiability in changing sociotechnological genomic data environments. Biosocieties. doi:https://doi.org/10.1057/s41292-023-00299-7.

Lewis, C.M. and Vassos, E. (2020). Polygenic risk scores: from research tools to clinical instruments. Genome Medicine, 12(1). doi:https://doi.org/10.1186/s13073-020-00742-5.

Li, Z., Huang, J. and Hu, Z. (2019). Screening and Diagnosis of Chronic Pharyngitis Based on Deep Learning. International Journal of Environmental Research and Public Health, [online] 16(10), p.1688. doi:https://doi.org/10.3390/ijerph16101688.

Maor, E., Sara, J.D., Orbelo, D.M., Lerman, L.O., Levanon, Y. and Lerman, A. (2018). Voice Signal Characteristics Are Independently Associated With Coronary Artery Disease. Mayo Clinic Proceedings, 93(7), pp.840–847. doi:https://doi.org/10.1016/j.mayocp.2017.12.025.

Marmar, C.R., Brown, A.D., Qian, M., Laska, E., Siegel, C., Li, M., Abu‐Amara, D., Tsiartas, A., Richey, C., Smith, J., Knoth, B. and Vergyri, D. (2019). Speech‐based markers for posttraumatic stress disorder in US veterans. Depression and Anxiety, 36(7), pp.607–616. doi:https://doi.org/10.1002/da.22890.

Mobadersany, P., Yousefi, S., Amgad, M., Gutman, D.A., Barnholtz-Sloan, J.S., Velázquez Vega, J.E., Brat, D.J. and Cooper, L.A.D. (2018). Predicting cancer outcomes from histology and genomics using convolutional networks. Proceedings of the National Academy of Sciences, 115(13), pp.E2970–E2979. doi:https://doi.org/10.1073/pnas.1717139115.

Poplin, R., Chang, P.-C., Alexander, D., Schwartz, S., Colthurst, T., Ku, A., Newburger, D., Dijamco, J., Nguyen, N., Afshar, P.T., Gross, S.S., Dorfman, L., McLean, C.Y. and DePristo, M.A. (2018). A universal SNP and small-indel variant caller using deep neural networks. Nature Biotechnology, 36(10), pp.983–987. doi:https://doi.org/10.1038/nbt.4235.

Ringeval, F., Messner, E.-M., Song, S., Liu, S., Zhao, Z., Mallol-Ragolta, A., Ren, Z., Soleymani, M., Pantic, M., Schuller, B., Valstar, M., Cummins, N., Cowie, R., Tavabi, L., Schmitt, M., Alisamir, S. and Amiriparian, S. (2019). AVEC 2019 Workshop and Challenge: State-of-Mind, Detecting Depression with AI, and Cross-Cultural Affect Recognition. Proceedings of the 9th International on Audio/Visual Emotion Challenge and Workshop – AVEC ’19. doi:https://doi.org/10.1145/3347320.3357688.

Wang, Y., Tsuo, K., Kanai, M., Neale, B.M. and Martin, A.R. (2022). Challenges and Opportunities for Developing More Generalizable Polygenic Risk Scores. Annual Review of Biomedical Data Science, 5(1), pp.293–320. doi:https://doi.org/10.1146/annurev-biodatasci-111721-074830.

Zhan, A., Mohan, S., Tarolli, C., Schneider, R.B., Adams, J.L., Sharma, S., Elson, M.J., Spear, K.L., Glidden, A.M., Little, M.A., Terzis, A., Dorsey, E.R. and Saria, S. (2018). Using Smartphones and Machine Learning to Quantify Parkinson Disease Severity. JAMA Neurology, 75(7), p.876. doi:https://doi.org/10.1001/jamaneurol.2018.0809.

Zhou, J., Theesfeld, C.L., Yao, K., Chen, K.M., Wong, A.K. and Troyanskaya, O.G. (2018). Deep learning sequence-based ab initio prediction of variant effects on expression and disease risk. Nature Genetics, [online] 50(8), pp.1171–1179. doi:https://doi.org/10.1038/s41588-018-0160-6.

Zhou, J. and Troyanskaya, O.G. (2015). Predicting effects of noncoding variants with deep learning–based sequence model. Nature Methods, 12(10), pp.931–934. doi:https://doi.org/10.1038/nmeth.3547.