P-NET
Summary¶
Elmarakeby, Hwang, Arafeh, et al. (2021) present P-NET, a biologically informed deep neural network in which layer connectivity is constrained by 3,007 curated biological pathways (features → genes → pathways of increasing abstraction → outcome), trained to classify prostate cancer patients as having primary or metastatic castration-resistant disease from somatic mutation and copy-number data. P-NET outperformed standard ML baselines (AUC=0.93, AUPRC=0.88, accuracy=0.83) and its full-model interpretability implicated MDM4 as a treatment-resistance driver, a finding the authors validated experimentally with CRISPR knockdown and a selective inhibitor.
Architecture and Training¶
- Layer structure: patient molecular features (mutations, copy-number alterations) feed into a gene layer, which feeds into successive pathway layers of increasing biological abstraction, ending in the disease-state outcome — connections between layers are constrained to known child-parent (gene-pathway, pathway-pathway) relationships rather than being fully connected, making the network sparse and interpretable by construction.
- Training data: 1,013 prostate cancer samples (333 castration-resistant, 680 primary), 80/10/10 train/validation/test split, using somatic mutation and copy-number data.
- Interpretation: DeepLIFT attribution scores rank the relative importance of every gene and pathway node; node activation patterns can additionally be traced across layers to observe how information differs between primary and metastatic input samples.
Performance¶
- Held-out test set: AUC=0.93, AUPRC=0.88, accuracy=0.83, outperforming linear/RBF SVM, logistic regression, and decision trees on the same task.
- Sparse vs. dense comparison: a dense network with the same number of parameters as P-NET was trained on logarithmically increasing sample sizes (100–811); P-NET's mean AUC was higher at every sample size, with the difference statistically significant up to 500 samples (e.g., P=0.004 at n=155) — evidence that the biological sparsity constraint acts as an effective regularizer, particularly valuable given the modest sample sizes typical of curated cancer cohorts.
- External validation: on two independent cohorts (one primary, one metastatic), P-NET correctly classified 73% of primary tumors (true-negative rate) and 80% of metastatic tumors (true-positive rate).
- Clinical stratification signal: among primary-tumor patients, those misclassified by P-NET as having resistant disease ("high P-NET score") had significantly higher rates of biochemical recurrence than correctly-classified low-score patients (P=8×10⁻⁵, log-rank test) — suggesting the P-NET score itself carries prognostic information beyond the binary classification task it was trained on.
Biological Discovery: MDM4¶
- Interpreting the trained network's gene layer (via DeepLIFT) ranked known prostate cancer drivers highly (AR, PTEN, RB1, TP53) alongside less-expected genes including MDM4, FGFR1, NOTCH1, and PDGFA.
- MDM4 amplification was significantly more prevalent in resistant than primary samples (χ²=40.8, P<0.00001) and scored as the strongest hit for enzalutamide resistance in an independent genome-scale ORF overexpression screen (17,255 ORFs) in LNCaP cells.
- CRISPR-Cas9 knockdown of MDM4 significantly reduced proliferation across multiple prostate cancer cell lines (P<0.0001), and TP53-wild-type cells were more sensitive than TP53-mutant cells to the MDM4-selective inhibitor RO-5963 — together supporting MDM4 as a therapeutically targetable vulnerability in a genomically-defined (TP53-wild-type) patient subgroup.
See Also¶
- Biologically Informed Neural Networks — the shared design philosophy across P-NET, BINN, and expiMap.
- BINN — applies the same pathway-constrained sparse architecture to clinical proteomics rather than cancer genomics.