Multi-omics refers to the simultaneous study of multiple layers of biological information inside a cell. Cells are governed by DNA (the genetic blueprint), RNA (which reads that blueprint), proteins (which carry out cellular tasks), and metabolites (small molecules that fuel cell activity). The scientific fields that study each of these layers are called genomics, transcriptomics, proteomics, and metabolomics, respectively - and together they form the 'omics' family.
Traditional biological research usually examined only one layer at a time, such as looking at gene mutations alone. Multi-omics breaks that limitation by combining data from all these layers simultaneously, revealing how mutations affect gene expression, which then changes protein behavior and cell metabolism. This integrated view is far more powerful than any single measurement on its own.
Leukemia is a cancer of the blood and bone marrow in which abnormal white blood cells multiply uncontrollably. Because it arises from multiple interacting molecular changes, single-layer analysis often misses the full picture. Multi-omics approaches are therefore especially valuable for understanding and predicting leukemia, potentially enabling earlier diagnosis and more personalized treatment choices.
The enormous amount of data generated by multi-omics experiments presents a significant challenge for human researchers to analyze manually. This is where Artificial Intelligence (AI) becomes essential - machine learning and deep learning algorithms can process tens of thousands of molecular measurements simultaneously and identify patterns invisible to the naked eye.
This 2024 study from researchers at Beijing University of Posts and Telecommunications and Najran University set out to build an AI-powered leukemia prediction system using multi-omics data. The team collected three separate genomic datasets from the International Cancer Genome Consortium (ICGC) data portal, covering patient donor records, somatic mutation data, and tissue specimen information - a combined pool of over 10,000 samples.
The researchers compared two broad families of AI: Machine Learning (ML), which uses hand-crafted statistical rules, and Deep Learning (DL), which uses layered artificial neural networks that learn features directly from data. Five ML algorithms were tested - Random Forest, Naive Bayes, Decision Tree, Logistic Regression, and Gradient Boosting - alongside two deep learning architectures: Feedforward Neural Networks and Recurrent Neural Networks.
The key finding was that the Recurrent Neural Network (RNN) achieved the highest accuracy of 98%, narrowly outperforming the best ML algorithm (Gradient Boosting) at 97%. Naive Bayes performed significantly worse than all others, reaching only 73% accuracy on the balanced dataset, highlighting that algorithm selection matters greatly in medical AI.
The model used 17 patient and genomic features as inputs, including age, sex, mutation type, chromosome involved, and treatment method. This combination of clinical and molecular variables made the prediction more realistic and clinically relevant than studies relying on molecular data alone.
Machine Learning (ML) is a branch of AI that learns from examples. You feed it labeled data (for example, patient records tagged as 'leukemia' or 'no leukemia') and it builds a statistical model that can predict labels for new, unseen cases. ML algorithms work by finding mathematical patterns in a set of input features - the algorithm never truly 'understands' biology, but it can still achieve remarkable predictive accuracy.
Deep Learning (DL) is a specialized subset of ML that uses artificial neural networks inspired by the structure of the human brain. These networks contain many interconnected layers of mathematical units called neurons. Data flows through each layer, and each layer learns to recognize increasingly abstract features - from simple patterns in early layers to complex disease signatures in later layers. This layered architecture is what makes DL especially powerful for complex, high-dimensional data like multi-omics.
The paper tested two DL architectures. A Feedforward Neural Network (FNN) is the simplest type: information flows in one direction only, from input to output. A Recurrent Neural Network (RNN) adds feedback loops, allowing the network to retain information from earlier steps - a feature originally designed for sequential data like speech or text, but also useful for capturing dependencies among genomic features.
Both DL models used two types of mathematical tools: activation functions (ReLU and Softmax) that determine how each neuron fires, and loss functions (Binary Cross-Entropy and Mean Squared Error) that measure how wrong the model's predictions are during training. By minimizing the loss over 200 training cycles called epochs, the networks gradually tuned their internal weights to maximize accuracy.
The dataset came from the publicly accessible ICGC (International Cancer Genome Consortium) Data Portal, one of the world's largest repositories of cancer genomic data. Three separate data files were retrieved: a Donor file with 1,267 patient records covering age, sex, survival time, and relapse type; a Simple Somatic Mutation file with 5,983 records of DNA mutations verified by Illumina HiSeq sequencing technology; and a Specimen file with 2,844 tissue sample records. Together these formed a rich multi-omics dataset.
Raw genomic data is rarely clean enough to use directly. The team applied a careful preprocessing pipeline to remove unreliable data: any feature (column) with more than 50% missing values was dropped entirely, while features with 20-50% missing values were kept and their gaps filled using a technique called KNN Imputer - a method that estimates missing values based on the values of the most similar data points (K-nearest neighbors).
A critical challenge in medical datasets is class imbalance - where far more patients are labeled 'healthy' than 'diseased', causing AI models to simply predict the majority class. To fix this, the ML experiments used SMOTE (Synthetic Minority Oversampling Technique), which generates artificial examples of the minority class to balance the dataset. The final balanced dataset contained 1,690 samples, split 85% for training and 15% for testing.
To identify which of the many available variables were most predictive, the team used Pearson correlation coefficient analysis - a standard statistical measure that scores how strongly two variables move together on a scale from -1 to 1. Features with a positive correlation to the leukemia label were selected for model training, ultimately yielding 17 meaningful features including patient age, mutation type, chromosome location, and treatment method.
The five ML algorithms were each trained on the same balanced dataset of 1,690 samples (1,436 for training, 254 for testing). Every algorithm received the same 17 selected features as input and was asked to predict whether a sample represented a leukemia case. Performance was measured using accuracy (percentage of correct predictions), precision (how often a positive prediction was truly positive), recall (how many actual positives were caught), and the F1 score (a combined measure balancing precision and recall).
For the deep learning models, the team used a sequential model architecture in which layers are stacked one after another. Both FNN and RNN were trained with a batch size of 32 (meaning 32 samples were processed together in each update step) over 200 epochs. To prevent overfitting - a situation where a model memorizes training data but fails on new data - early stopping criteria were applied, halting training when validation performance stopped improving.
The DL experiments did not use SMOTE, instead training directly on 1,267 samples split into training (633), validation (317), and testing (317) sets. This three-way split allowed the team to tune model settings on the validation set while keeping the test set completely untouched for final evaluation - a best-practice design that prevents inflated performance estimates.
Model quality was visualized using two standard diagnostic tools. The ROC curve (Receiver Operating Characteristic) plots the true positive rate against the false positive rate at different decision thresholds - a curve that hugs the upper-left corner indicates near-perfect discrimination. Confusion matrices provided a breakdown of exactly how many samples were correctly classified versus misclassified in each category.
An important finding emerged during preprocessing: when the dataset was left imbalanced, nearly all algorithms reported 100% accuracy - a clear warning sign. This spurious perfection disappeared once SMOTE balanced the classes, revealing true algorithmic performance. This underscores why proper data balancing is essential before drawing conclusions from medical AI models.
On the balanced dataset, four of the five ML algorithms performed well. Gradient Boosting (GB) was the standout ML algorithm, achieving 97% accuracy with 96% precision, 98% recall, and a 97% F1 score. It correctly predicted 246 out of 254 test samples, misclassifying only 8. Decision Tree (DT) matched this 97% accuracy. Random Forest (RF) and Logistic Regression (LR) both reached 96% accuracy.
Naive Bayes (NB) was the clear outlier, achieving only 73% accuracy on balanced data. It correctly predicted just 230 of 254 samples, with 24 misclassifications. Naive Bayes assumes all input features are statistically independent of each other - an assumption that rarely holds true in biological data where genes and clinical variables interact closely. This explains its relatively poor performance on multi-omics data.
The ROC curve analysis confirmed the rankings visually. The DT, GB, RF, and LR curves all bent sharply toward the upper-left corner of the plot, indicating strong discriminative power. The NB curve sat noticeably further from that ideal corner, consistent with its lower numerical accuracy. These visual and numerical results together make a clear case that ensemble and tree-based methods outperform probability-based methods like Naive Bayes on this type of high-dimensional genomic data.
The deep learning experiments tested FNN and RNN with two activation functions (ReLU and Softmax) and two loss functions (Binary Cross-Entropy and Mean Squared Error), creating several model configurations. The best-performing configuration was the RNN using Binary Cross-Entropy (BCE) loss with the Softmax activation function, which reached 98% accuracy - the highest result in the entire study.
The ReLU (Rectified Linear Unit) activation function produced better results than Softmax in most configurations. ReLU outputs zero for any negative input and the value itself for positive inputs - this simple rule prevents a problem called 'vanishing gradients' that slows down learning in deep networks. Softmax, by contrast, converts outputs into a probability distribution over classes and is more commonly used in the final output layer rather than hidden layers, which may explain its inferior performance here.
When the RNN used MSE loss with the Softmax function, it still correctly predicted 306 of 317 test samples (about 96%). With BCE loss and Softmax, it accurately predicted 307 of 317 samples (98%). The FNN showed more variability: with ReLU and MSE it correctly predicted 300 of 317 samples (97%), but with Softmax and MSE it correctly predicted only 177 of 317 (about 56%), illustrating how sensitive neural networks can be to the combination of architectural choices made.
The loss curves for the RNN showed smooth, gradual convergence over 200 epochs, indicating stable and reliable learning. A notable spike in the BCE loss curve appeared between epochs 75 and 100, suggesting a critical transition point in training where the network rapidly adjusted its internal parameters before stabilizing again. The FNN's training curves showed more irregular behavior, reflecting its simpler architecture's relative difficulty handling the complexity of multi-omics data.
The study's comparative analysis confirms a pattern seen across computational biology: deep learning handles multi-omics data better than classical machine learning because biological datasets contain many nonlinear relationships - interactions between genes, proteins, and clinical variables that simple statistical rules cannot capture. Neural networks can learn arbitrarily complex patterns given enough data and training time.
The RNN's advantage over the FNN likely stems from its ability to model dependencies between features. While an FNN treats all 17 input features as independent quantities, an RNN's internal memory allows it to capture sequential or conditional relationships among them. In the context of genomic data, this might correspond to recognizing that a particular mutation type matters more or less depending on the patient's age or which chromosome is affected.
Compared to previous studies in the literature, this work achieved notably higher accuracy. Earlier deep learning approaches applied to leukemia bone marrow images reached only 89% accuracy. Studies on neuroblastoma using SVM and random forests reached 60-65%. Even sophisticated multi-omics integration models on other cancer types topped out around 76%. The 98% achieved here represents a meaningful improvement, though the authors acknowledge this was tested on a single dataset.
A key limitation the authors openly discuss is that the model was evaluated in a simulation environment using a single database. Before clinical deployment, the model would need validation on independent patient cohorts from different hospitals, ideally in real-time diagnostic settings. The authors also note that future work should incorporate additional omics layers beyond genomics - such as proteomics and methylation data - to further boost predictive power and applicability to other cancers.
Leukemia is a family of blood cancers that primarily affects white blood cells - the cells responsible for fighting infections. The disease is classified by how quickly it progresses (acute vs. chronic) and by which type of blood cell is affected (myeloid vs. lymphoid). The four main types are Acute Myeloid Leukemia (AML), Acute Lymphoblastic Leukemia (ALL), Chronic Myeloid Leukemia (CML), and Chronic Lymphocytic Leukemia (CLL).
Acute Myeloid Leukemia (AML) is one of the most aggressive forms, affecting both adults and children. In patients aged 18-60, the 2-year survival rate remains a critical benchmark. Encouragingly, outcomes for children with AML have improved substantially over the past three decades: overall survival rates have risen to 70-75%, attributed to better supportive care and optimized chemotherapy protocols.
Acute Lymphoblastic Leukemia (ALL), the most common childhood cancer, has also been studied with multi-omics methods. A landmark multi-omics analysis of 49 ALL cell lines quantified over 12,000 proteins and transcripts and identified sensitivity patterns to 528 drugs - precisely the kind of molecular profiling that AI tools like those developed in this study are designed to exploit for personalized treatment decisions.
The clinical promise of this AI approach lies in its use of features readily available in hospital records: patient age, sex, mutation type, chromosome location, and treatment history. A model that can predict leukemia risk or subtype from these variables - without requiring expensive new tests - would be genuinely practical and accessible, especially in resource-limited healthcare settings.
This study demonstrates that integrating multi-omics data with AI - specifically deep learning - can achieve high-accuracy leukemia prediction that surpasses both classical machine learning and most prior AI approaches for cancer diagnosis. The RNN with Binary Cross-Entropy loss achieved 98% accuracy, while Gradient Boosting led the ML group at 97%, confirming that both families of AI have genuine value but deep learning holds an edge on complex biological data.
The study's methodological contributions extend beyond the accuracy numbers. The authors demonstrate a replicable eight-step pipeline - from multi-source data collection through preprocessing, class balancing, feature selection, model training, and performance evaluation - that could serve as a template for applying AI to other cancer types and other multi-omics datasets. All code and data are publicly available on GitHub, promoting transparency and reproducibility.
The authors identify several priorities for future research. First, expanding the model to additional omics layers such as methylation patterns, protein expression, and metabolomics could capture even richer biological information. Second, the model should be tested on larger, more diverse patient cohorts from multiple institutions to confirm that its accuracy generalizes beyond the ICGC dataset used here.
Perhaps most importantly, the study calls for eventual testing in real-time clinical scenarios with live patient data. Transitioning from a research simulation to a clinical decision-support tool requires regulatory oversight, physician collaboration, and prospective trials - but this work represents a strong scientific foundation for that journey. The authors envision AI-generated genomic biomarker insights eventually improving care quality across all cancer types.