A systematic review of biologically-informed deep learning models for cancer: fundamental trends for encoding and interpreting oncology data

Abstract

Background There is an increasing interest in the use of Deep Learning (DL) based methods as a supporting analytical framework in oncology. However, most direct applications of DL will deliver models with limited transparency and explainability, which constrain their deployment in biomedical settings. Methods This systematic review discusses DL models used to support inference in cancer biology with a particular emphasis on multi-omics analysis. It focuses on how existing models address the need for better dialogue with prior knowledge, biological plausibility and interpretability, fundamental properties in the biomedical domain. For this, we retrieved and analyzed 42 studies focusing on emerging architectural and methodological advances, the encoding of biological domain knowledge and the integration of explainability methods. Results We discuss the recent evolutionary arch of DL models in the direction of integrating prior biological relational and network knowledge to support better generalisation (e.g. pathways or Protein-Protein-Interaction networks) and interpretability. This represents a fundamental functional shift towards models which can integrate mechanistic and statistical inference aspects. We introduce a concept of bio-centric interpretability and according to its taxonomy, we discuss representational methodologies for the integration of domain prior knowledge in such models. Conclusions The paper provides a critical outlook into contemporary methods for explainability and interpretability used in DL for cancer. The analysis points in the direction of a convergence between encoding prior knowledge and improved interpretability. We introduce bio-centric interpretability which is an important step towards formalisation of biological interpretability of DL models and developing methods that are less problem- or application-specific.

Inference of drug off-target effects on cellular signaling using interactome-based deep learning

“…The advent of ML and big data in biology holds promise for a more data-driven life science. However, ML models have been criticized for their lack of interpretability 54 , 55 and thus many times fail to explain the underlying MoA in a biological phenomenon or were never designed to do so. Embedding prior knowledge into the structure of ML models can improve their interpretability.…”

AutoTransOP: translating omics signatures without orthologue requirements using deep learning

“…Interpretability of deep learning models in biology remains a challenge. These models have been criticized for providing a poor understanding of which biological relationships they capture 51 , 52 . On this front, we demonstrate in our framework how integrated gradient approaches 35 can be used to estimate the importance of features used by different parts of the framework for various tasks, enabling some biological interpretation of the model.…”

Reliable interpretability of biology-inspired deep neural networks

“…Biology-inspired neural networks enable a unique type of interpretability by incorporating domain knowledge of biological relationships into the architecture of deep learning models 11 , 12 . However, critical steps must be taken to ensure reliability of interpretations.…”

“…Interpretation approaches of deep learning models are primarily focused on the feature level 8 – 10 . In computational biology 11 , 12 , a recent alternative approach utilizes prior knowledge on biological networks to influence the structure of the neural network, in so-called “visible”, “biologically-inspired”, or “knowledge-primed” neural networks 13 – 18 . In such biology-inspired deep learning models, hidden layers consist of nodes that correspond to biological entities, for example, Gene Ontology (GO) terms 15 , Reactome pathways 17 , or signaling proteins 16 .…”

“…Layers are then sparsely connected based on existing knowledge of relationships of the encoded biological entities. The architecture of biology-inspired models is thus informed by domain knowledge, which enables a unique interpretability 11 , 12 : After training, a measure of importance is calculated for each node in the hidden layers (“node importance score”), which quantifies the importance of each biological entity in the network for the prediction task. For example, the P-NET model 17 used mutation data to predict cancer state (primary versus metastatic cancer) in a network, where hidden nodes correspond to genes and Reactome pathways, thus identifying genes and pathways relevant for metastasis.…”