Skip to content

Biologically Informed Neural Networks

Summary

Biologically informed neural networks replace some or all of a standard dense neural network's fully-connected layers with sparse connectivity constrained by known biological relationships — gene-to-pathway membership (Reactome, curated pathway databases) or gene-to-program annotations (gene sets, gene ontology) — so that each hidden node corresponds to an interpretable biological entity rather than an arbitrary learned combination of inputs. This design appears across three distinct application domains reviewed here: cancer genomics (P-NET), proteomic biomarker/pathway discovery (BINN), and single-cell reference mapping (expiMap) — and in each case the biological constraint is reported to improve either predictive performance (especially at small sample sizes), interpretability, or both, relative to a comparably-sized dense network or classical ML baseline.

Shared Design Pattern

  1. Prior knowledge as network topology: a database of biological relationships (pathway membership, gene programs, ontology hierarchies) is converted into a sparse connectivity mask between consecutive layers, so a node in layer n+1 only receives input from the layer-n nodes representing its known biological constituents.
  2. Hierarchical abstraction: successive layers represent increasingly abstract biological entities — e.g., genes → fine-grained pathways → higher-level biological processes (P-NET, BINN) — mirroring the natural hierarchy of biological knowledge bases.
  3. Post-hoc interpretability by design, not by approximation: because each node has a fixed biological identity, standard feature-attribution methods (DeepLIFT in P-NET, SHAP in BINN) directly yield per-gene, per-pathway, and per-process importance scores, rather than requiring a separate surrogate-model explanation step.
  4. Reported small-sample advantage: P-NET's sparse architecture significantly outperformed a dense network with the same parameter count specifically at smaller training-sample sizes, consistent with sparsity acting as a strong, biologically-motivated regularizer.

Applications Compared

Tool Domain Biological prior Task
P-NET Cancer genomics 3,007 curated biological pathways Classify primary vs. metastatic castration-resistant prostate cancer from mutation/copy-number data
BINN Clinical proteomics Reactome pathway database Classify clinical subphenotypes (septic AKI, COVID-19 severity, ARDS aetiology) from plasma proteomics
expiMap Single-cell genomics Curated + de novo "gene programs" Map query single-cell datasets onto a reference atlas while keeping the resulting latent space biologically interpretable

Despite spanning three different data modalities (bulk mutation/CNV, bulk plasma proteomics, single-cell transcriptomics) and three different downstream tasks (classification, classification, representation learning/integration), all three papers converge on the same core claim: constraining network structure with a curated biological knowledge base yields models that are simultaneously more accurate (or comparably accurate) and substantially easier to interpret than unconstrained dense architectures of similar capacity.

See Also

Citations

[1] Elmarakeby, H.A., Hwang, J., Arafeh, R., Crowdis, J., Gang, S., Liu, D., AlDubayan, S.H., Salari, K., Kregel, S., Richter, C., Arnoff, T.E., Park, J., Hahn, W.C., & Van Allen, E.M. (2021). Biologically informed deep neural network for prostate cancer discovery. Nature, 598, 348–352. [2] Hartman, E., Scott, A.M., Karlsson, C., Mohanty, T., Vaara, S.T., Linder, A., Malmström, L., & Malmström, J. (2023). Interpreting biologically informed neural networks for enhanced proteomic biomarker discovery and pathway analysis. Nature Communications, 14, 5359. [3] Lotfollahi, M., Rybakov, S., Hrovatin, K., Hediyeh-zadeh, S., Talavera-López, C., Misharin, A.V., & Theis, F.J. (2023). Biologically informed deep learning to query gene programs in single-cell atlases. Nature Cell Biology, 25, 337–350.