Deep learning models in genomics and systems biology face a notorious curse of dimensionality: they must learn to make predictions from tens of thousands of genes using only dozens or hundreds of patient samples. To navigate this sparse landscape, a powerful paradigm has emerged: biologically informed machine learning. By embedding structured prior knowledge—such as gene ontologies, metabolic pathways, and protein-protein interaction (PPI) networks—directly into model architectures, researchers can radically constrain the search space, reduce overfitting, and build highly interpretable "glass-box" systems.
However, this elegant solution introduces a fundamental vulnerability: prior knowledge dependency. When a model's architecture or regularization scheme is hardwired to existing biological databases, it inherits all the systemic flaws, gaps, and historical research biases of those databases. Instead of discovering new biology, these models risk merely echoing what is already documented in scientific literature.
How Knowledge Constraints Shape Model Architectures
To understand why prior knowledge dependency occurs, we must look at how these priors are integrated into modern neural networks. In a traditional black-box network, every neuron in one layer can connect to every neuron in the next, allowing the model to learn any arbitrary interaction given enough data. In contrast, "visible" or biologically informed neural networks (VNNs) restrict these connections based on curated databases like Reactome or the Gene Ontology (GO).
For instance, in the multimodal survival prediction framework proposed in Deep Biological Pathway Informed Pathology-Genomic Multimodal Survival Prediction, the network topology is strictly mapped to known biological pathways. As reviewed by David Antony Selby et al. in Visible neural networks for multi-omics integration: a critical review, this structural mapping forces the model to process information through the lens of established biology.
While this makes the internal states human-readable, it creates a rigid mathematical boundary. If a critical biological mechanism or cross-talk between two pathways is missing from the underlying database, a sparse network restricted to known pathways is structurally incapable of routing that signal. The model cannot learn an interaction that its architecture explicitly forbids.
The Blind Spots of Historical Consensus
The dependency on prior knowledge creates a significant blind spot for novel biological discovery. Curated databases are not objective ground truths; they are historical records of scientific publication, which is heavily biased toward well-studied genes and popular disease pathways.
When models rely on these databases to guide feature selection or regularize loss functions, they inadvertently penalize emerging phenomena. As demonstrated by Jusang Lee et al. in Prioritizing biological pathways by recognizing context in time-series gene expression data, integrating literature-based priors can suppress novel discoveries simply because they lack historical documentation.
This creates a self-reinforcing feedback loop. Supervised models designed to identify disease drivers are often trained on "gold standard" reference sets that suffer from the same curation biases (Evaluation and integration of cancer gene classifiers: identification and ranking of plausible drivers). Consequently, the model becomes excellent at identifying well-documented interactions but blind to rare, niche, or context-specific mechanisms.
Error Propagation and Performance Degradation
Beyond missing novel insights, incomplete or inaccurate priors can actively degrade a model's predictive performance. In graph neural networks (GNNs) used for multi-omics integration, external networks act as the foundational graph structure over which messages are passed between nodes.
In AGMI: Attention-Guided Multi-omics Integration for Drug Response Prediction with Graph Neural Networks, the authors note that the performance of models incorporating biological priors is strictly bounded by the completeness of external PPI and pathway networks. If the reference graph contains false positives or lacks tissue-specific edges, these errors propagate through the message-passing layers. Instead of regularizing the model to prevent overfitting, a flawed prior misguides the attention mechanism, leading to suboptimal predictions.
Similarly, in dimensionality reduction, methods like the one explored in Pathway Activity Score Learning for Dimensionality Reduction of Gene Expression Data by Ioulia Karagiannaki show that restricting latent spaces to predefined gene sets forces the model to discard significant variance driven by undocumented biological processes.
Breaking the Dependency: Soft Priors and Knowledge Distillation
To overcome these constraints, recent research has shifted away from treating prior knowledge as an immutable blueprint, moving instead toward frameworks that treat priors as flexible guides.
One approach is to allow the model to dynamically refine the prior during training. For example, the factor analysis model developed in f-scLVM: scalable and versatile factor analysis for single-cell RNA-seq by Ploy N. Pratanwanich et al. can actively add or remove genes from predefined pathway annotations based on the experimental data, capturing unannotated variation.
A more radical departure is found in KINDLE: Knowledge-Guided Distillation for Prior-Free Gene Regulatory Network Inference. Here, Rui Peng et al. introduce a teacher-student distillation framework. A "teacher" model is trained using existing biological priors to capture regulatory intuition, but this knowledge is then distilled into a "prior-free" student model. By decoupling the final inference engine from hard database constraints, the student model retains the predictive power of the prior knowledge while regaining the freedom to discover entirely novel regulatory interactions.
Key Foundational Work
Detection of gene pathways with predictive power for breast cancer prognosis (2010) — contributed 1 concept elements to this direction.
Causal reasoning identifies mechanisms of sensitivity for a novel AKT kinase inhibitor, GSK690693 (2010) — contributed 1 concept elements to this direction.
Systematic survey reveals general applicability of "guilt-by-association" within gene coexpression networks (2005) — contributed 1 concept elements to this direction.