Incorporating biological structure into machine learning models in biomedicine

Abstract

In biomedical applications of machine learning, relevant information often has a rich structure that is not easily encoded as real-valued predictors. Examples of such data include DNA or RNA sequences, gene sets or pathways, gene interaction or coexpression networks, ontologies, and phylogenetic trees. We highlight recent examples of machine learning models that use structure to constrain model architecture or incorporate structured data into model training. For machine learning in biomedicine, where sample size is limited and model interpretability is critical, incorporating prior knowledge in the form of structured data can be particularly useful. The area of research would bene t from performant open source implementations and independent benchmarking e orts.Structure of ontology, phylogeny, etc. Labels or cluster assignments Tabular dataSchematic showing the main categories of models incorporating structured biological data covered in this review. The rst panel shows an example of a model operating on raw sequence data, the second panel shows a model in which dimension reduction is constrained by the connections in a gene network, and the third panel shows a neural network with structure constrained by a phylogeny or ontology.

Network-based machine learning in colorectal and bladder organoid models predicts anti-cancer drug efficacy in patients

“…Our research further supports the utility of network analysis to reduce the biological complexity of input datasets by capturing robust signals, which can improve the predictive performances of ML tasks 10 , 11 . Recently, the PsycheENCODE Consortium embedded a regulatory network into a deep-learning model and demonstrated that functional genomic data improved the detection of gene-phenotype associations in neurological disorders 42 .…”

“…Interpretability of ML models is becoming increasingly important, especially when it comes to the high-stakes decision-making of healthcare 40 . To generate interpretable ML models, incorporation of biological structures, such as PPI networks, into the ML models assists in making comprehensive predictions 10 , 11 . Recently, Ma et al showed that the infusion of hierarchical structures of biological functions into neural networks accurately predicted cellular growth from genotypes 41 .…”

“…However, preclinical models for drug biomarker identification or ML model development frequently fail to predict drug sensitivity in human tumors 8 , 9 . Differences in the complexity of the biological systems are one challenge of these models 10 , 11 . Also, limited training data can hinder the performance of ML techniques, in contrast to data-rich input features, such as gene expression profiles.…”

Bridging “Big Data” and Mechanistic Insight To Enable Precision Medicine

“…Finally, several groups are developing statistical learning methods that are augmented with experimental knowledge which benefits both computational and experimental communities [15] . From the computational side, including prior knowledge helps reduce overhead by biasing methods and parameters towards physical relevance [15b] . From the experimental side, these methods can provide a “blueprint” output and reduce false positives that can consume experimental resources to disprove.…”

MethylSPWNet and MethylCapsNet: Biologically Motivated Organization of DNAm Neural Network, Inspired by Capsule Networks

“…[51][52][53] Incorporation of prior biological knowledge not only improves the transparency and interpretability of the modeling approach but also reduces noise while increasing signal by meaningfully pruning redundant relationships between predictors. [54] Capsule Networks are a promising approach for grouping CpGs that harness their statistical interactions and relate the groupings of predictors to clinical and biological outcomes. [27] Capsule networks explicitly model the relationship between images or genomic information by its constituent parts, or capsules, by parameterizing pose matrices (unitary transformations), and then hierarchically associate each of these parts independently to higher-order targets of interest.…”