Machine Learning in Forensic DNA Profiling

Machine Learning in Forensic DNA Profiling: A Critical Review

• Machine learning (ML) involves computational algorithms generating predictive models by intelligently analyzing large, unstructured data sets.
• ML is being used in forensics, streamlining complex data analysis while maintaining accuracy and reproducibility.
• Forensic scientists may not be aware of ML capabilities, while computer science professionals might lack knowledge of forensic science specifics.
• This study introduces ML methods for forensic DNA analysis and critically reviews current research.

Machine Learning Approach

• Classical scientific methods explore relationships between system elements to build mechanistic models.
• Engineering sciences use empirical formulas with coefficients to account for unknown environmental factors.
• Forensic science employs probabilistic genotyping algorithms using empirical formulas, like STRmix, to estimate likelihoods.
• ML helps identify dependencies in large data volumes when relationships between variables are unknown.
• ML algorithms transform input variables (X) to predict output variables (Y), expressed as Y = T(X).
• Forensic implementation requires transparency, standardization, and validation procedures like SWGDAM guidelines.
• Methods for ML were initiated in the 1950s.
• ML methods include linear regression, discriminant analysis, k-NN algorithms, naive Bayes, decision trees, random forests, and neural networks.
• ML strategies depend on the problem and data presentation.

Types of Machine Learning

• Machine learning has 4 categories; supervised, unsupervised, semi-supervised, and reinforcement learning.

Supervised learning

• Supervised learning trains models with structured datasets of input and output variables.
• Training uses labeled samples, like DNA fragments labeled with STR loci and flanking regions.
• The algorithm learns the mapping, assigning labels to new examples based on established rules.
• Supervised learning requires high-quality, normalized data to reduce bias.
• Approaches: classification and regression analyses.

Unsupervised learning

• Unsupervised learning develops functions based on input data (X) without corresponding output labels (Y).
• It requires large datasets to accommodate diverse scenarios in X-Y connections.
• Organizes the data, but this classification depends on the presented and extracted features.
• Beneficial for auto-organizing terabytes of unlabeled data into similar clusters.

Semi-supervised learning

• Combines supervised and unsupervised approaches.
• Use when there is a large amount of input data, while only a small portion the data is labeled.
• Semi-supervised learning uses this information to improve model performance, particularly when labeled data is limited or costly to obtain.
• It provides additional information in the unlabelled data to improve the models performance.
• Example is raw electropherograms that can be used to distinguish alleles from background noise.

Reinforcement learning

• Involves an agent learning via interaction with the environment, receiving rewards or penalties to improve its policy through trial and error.
• Used in tasks requiring informed decisions through trial and error, such as game playing, robotics, and autonomous systems.

Types of problems solved with ML approach

• Classification: assigns input data to predefined categories or classes
• Regression: predicts continuous numerical values based on input data
• Clustering: groups similar data points together based on shared characteristics
• Dimensionality reduction: retains meaningful features while eliminating redundant ones
• Image and Video Recognition: analyzes visual data such as object/facial detection for things like bloodstains, sperm cells, and video captioning.
• Anomaly Detection: identifies unusual patterns/outliers for fault, fraud detection, or anomalies in sensor data
• Natural Language Processing: understands and processes human language
• Generative Models: creates new data like images, music, or text

Classification

• Supervised learning method requiring an annotated dataset
• Model determines which predefined group new data is assigned to
• Can be binary, dividing data in two groups or able to classifying into multiple categories by assessing their best fit to one of the several groups
• Classification tasks use different ML methods like liner discriminant analysis, logistic regression, Bayes, and neural networks

Clustering

• Similar to classification except the training approach uses unlabelled data
• Used to find common patterns in a datasets, distinguishing between groups
• Done by the presence of the most similar characteristics within each group
• Can be solved with k-Means, DBSCAN, and hierarchical clustering
• Model-based likelihood estimation is a type of clustering algorithm used for the population assignment and is represented by Structure

Regression Analysis

• Used to understand the behaviour of an object by studying the effects of each parameter under different conditions
• Mathematical regression might be better for the number of parameters that does not lend itself to analytical description
• Used with supervised learning and can be used for DNA phenotyping to predict hair pigmentation from a DNA sample
• Problems requiring regression analyses can be approached with several ML methods like linear and polynomial regressions, neural networks, etc...

Dimensionality Reduction

• A study involves collection of a large amount of different types of data related to an object or phenomenon under study.
• Some parameters measured are only generally related to the object of study
• Should be more productive to discard such irrelevant parameters to to facilitate the construction of a model
• Can be approched via linear discriminant analysis, PCA, generalized discriminant analysis, and t-distributed stochastic neighbour embedding

Generative Models

• By combining various methods of ML, it is possible to build more complex models can predict not only the class of an object or a specific value of a parameter, but create a comprehensive model of a system
• Falls under the umbrella of machine learning and is inspired by the structure and functionality of the human brain
• It has a goal of creating intelligent machines capable of making independent decisions
• Utilitizes variations of a hierarchical organization of artificial neurons with connections to other neurons
• A major public demonstration of deep learning; 2016 when the AlphaGo beat Lee Sedol in four games of Go

Benefits of machine learning methods

• Can streamline processing of large amounts of "big data".
• Automation significantly reduces the burden of manual data analysis tasks.
• Helps scientists focus on higher-level problem-solving and creativity.
• Can be applied in cases where data is missing.

Weaknesses and pitfalls of machine learning methods

• Preprocessing data requires mostly manual intervention still
• Using separate train/test/validation sets are used to evaluate if algorithms or overfitting data
• "Curse of dimensionality" refers to degrading performance of algorithms if the datasets dimensionality is too high
• Can also over-relies on ML algorithms as a substitute for human judgement
• Potential lack of transparency can undermine validity and acceptability of algorithms

Machine learning applications in forensic DNA analysis

• It can be especially beneficial for the field of DNA analysis
• Advancement in genomic technologies has caused complex gathering of data
• DNA analysis is used for solving numerous problems in genetics and genomics
• Forensic data requires draws conclusion from different sources
• Requires extensive knowledge and experience, rigid standards, and zero biases
• The data can be extremely complex as it includes DNA markers used for identity and investigative purposes and more
• Each if these includes numerous hidden variables, patterns and artifacts. Powerful software and thoroughly trained specialists are required for interpretation

STR allele designation from CE- and MPS- generated data

• Processes of the genotyping and data consists of steps, such as separation of signals from different color channels, identification of all peaks, designate allele calls, removing artifacts
• The complexity of separating the allele from background nose and artifacts is the basic problem
• Currently, the process is carried out in semi-automatic manner using dedicated expert software
• With the help of built in algorithms and numbers of validated thresholds, resulting in designating a DNA profile
• In this niche, most studies rely on raw EPG data to generate a model, and eliminate the data analysis thresholds used in the ST profiling
• In order words all the available information is used to learn and make informed predictions

STR genotyping using CE - generated data

• Includes Common technical artefacts of the CE process
• A possibility to solve has been offered by Adelman if al, describing a method for automatic detection and removal of fluorescence pull-ups
• Three more quantitative filters were applied, resulting in removal of all all electric and stutter peaks
• It was tested that peak heights was the most significant variable
• Dynamic local-specific AT together with ML albums to demonstrate better performance
• Pull-ups for ever not not the only are technical signals

Additional applications of machine learning in forensic DNA analysis

• ML methods have been successfully applied to number of other aspects
• One successful aspect is the Y-SNP halogroup
• All models showed high prediction accuracy or over 95% in the load halogroups resolution
• The RF model demonstrated a superior outcome compared to other models

Conclusions

• Algorithms are considered field of artificial intelligence
• Designed to train a computer program from experience with some tasks
• Use of the L algorithms and forensic science can provide valuable science
• Improve throughout the reliability as well as reducing subjectivity to human interpretation