Kanishk Yadav

Machine learning (ML) is transforming cancer research and care by enabling analysis of complex, high-dimensional datasets spanning genomics, transcriptomics, proteomics, imaging, and clinical records. By improving risk stratification, accelerating detection and diagnosis, and supporting treatment selection, ML has the potential to enhance survival outcomes while increasing efficiency across oncology workflows. This review synthesizes key developments in ML for oncology, covering foundational algorithms alongside emerging approaches. We highlight major application areas including early cancer detection, tumor classification, molecular subtyping, biomarker discovery, prognosis estimation, multi-omics integration, computational pathology, pharmacogenomics, and clinical decision support. We also summarize commonly used datasets, discuss the importance of interpretability for clinical trust, and outline barriers to translation such as data heterogeneity, bias, and regulatory constraints. Finally, we describe future directions, including federated learning, graph neural networks, longitudinal modeling, and integration of real-world and wearable data to support precision oncology. This review is intended for cancer researchers, clinicians, and data scientists seeking a practical overview of ML methods, opportunities, and translational considerations in oncology.

Background: Inherited variants in cancer susceptibility genes play a critical role in breast cancer risk, yet comprehensive comparative analyses across clinically relevant genes remain limited. While BRCA1 and BRCA2 are well-established high-risk contributors, the broader landscape of germline pathogenicity across breast cancer associated genes warrants systematic evaluation. Methods: We curated a panel of 15 breast cancer associated genes ( BRCA1 , BRCA2 , TP53 , PTEN , CHEK2 , ATM , CDH1 , PALB2 , LSP1 , MAP3K1 , NF1 , RECQL , TOX3 , FANCD2 , RAD51C ) based on clinical and epidemiological relevance. Publicly available germline variant data from gnomAD v4.1 and the 1000 Genomes Project were integrated with ClinVar annotations. Variants were classified by clinical significance and filtered using allele frequency and phenotype annotation to stratify into high-risk and non-high-risk groups. We compared the distribution of variants and assessed penetrance patterns, and we evaluated the discriminatory power of computational pathogenicity predictors including CADD, REVEL, SIFT, and PolyPhen-2. Results: BRCA1 and BRCA2 harbored the highest burden of pathogenic variants, consistent with their high penetrance status. Moderate-risk genes ( PALB2 , RAD51C , CHEK2 ) showed intermediate levels of pathogenicity, while low-risk genes such as RECQL , TOX3 , and FANCD2 were predominantly enriched for variants of uncertain significance. High-risk variants exhibited significantly elevated CADD (>30), REVEL (>0.75), and SIFT (2 ), whereas PolyPhen-2 showed limited discriminatory power. Odds ratio analysis further supported CADD, REVEL, and ClinVar pathogenicity labels as strong discriminators between variant categories. Conclusions: Our analysis reaffirms the pathogenic burden concentrated in high penetrance genes and highlights the utility of combined in silico prediction tools and population frequency filters in variant prioritization. These findings support refined variant curation pipelines for hereditary breast cancer and provide a framework for interpreting gene-level pathogenicity across diverse susceptibility loci.

In veins, clotting initiation displays a threshold response to flow intensity and injury size. Mathematical models can provide insights into the conditions leading to clot growth initiation under flow for specific subjects. However, it is hard to determine the thrombin generation curves that favor coagulation initiation in a fast manner, especially when considering a wide range of conditions related to flow and injury size. In this work, we propose to address this challenge by using a neural network model trained with the numerical simulations of a validated 2D model for clot formation. Our surrogate model approximates the results of the 2D simulations, reaching an accuracy of 94% on the test dataset. We used the trained artificial neural network to determine the threshold for thrombin generation parameters that alter the coagulation initiation response under varying flow speed and injury size conditions. Our model predictions show that increased levels of the endogenous thrombin potential (ETP) and peak thrombin concentration increase the likelihood of coagulation initiation, while an elevated time to peak decreases coagulation. The lag time has a small effect on coagulation initiation, especially when the injury size is small. Our surrogate model can be considered as a proof-of-concept of a tool that can be deployed to estimate the risk of bleeding in specific patients based on their Thrombin Generation Assay results.