Predicting Therapeutic Outcome via Aligning Patient-Specific Knowledge Graph and Gene-Level Perturbation Representations

Abstract: Accurate prediction of patient-specific therapeutic response from pre-treatment transcriptomes is hindered by the scarcity of matched clinical response labels and post-treatment molecular profiles. Preclinical transfer-learning models can simulate drug-induced expression changes but are often hard to interpret and unstable, whereas knowledge-graph methods provide mechanistic context yet remain static and fail to capture drug-induced transcriptomic perturbation dynamics. We propose PREDIKTOR, a patient-centered multi-view framework that aligns a personalized network view with a transferable transcriptomic perturbation view to predict clinical drug response. For each patient, we construct an individualized gene regulatory network from tumor expression using DysRegNet and augment it with drug-target links from DrugBank; a graph neural encoder yields a drug-centric, mechanistically grounded embedding. In parallel, a frozen condition-specific gene-gene attention model pretrained on LINCS L1000 generates a simulated post-perturbation transcriptomic profile for the same patient-drug pair. We align the two views in a shared latent space via a CLIP-style contrastive objective with drug-context hard negatives, then concatenate the representations for end-to-end response classification. On TCGA, PREDIKTOR consistently outperforms state-of-the-art baselines under patient-, drug-, and tissue-split evaluations, and transfers zero-shot to the I-SPY2 trial, improving AUROC by 5.6% over competing methods. The aligned embeddings yield stable gene and pathway attributions that recover known mechanisms, supporting actionable and interpretable precision oncology.

• The paper introduces PREDIKTOR, which contrastively aligns personalized patient gene-regulatory networks with frozen, LINCS-trained perturbation representations to predict binary treatment response without paired post-treatment profiles.
• PREDIKTOR achieves AUROC scores of 0.837 for patient splits, 0.594 for drug splits, and 0.603 for tissue splits on TCGA, while zero-shot testing on I-SPY2 reaches 0.661 AUROC for paclitaxel response.
• The paper shows that personalized network structure, perturbation modeling, and patient-specific hard negatives drive performance and produce biologically meaningful attributions enriched for paclitaxel-related pathways, although cold-start drugs and small cohorts remain major limitations.

PREDIKTOR addresses a persistent bottleneck in transcriptome-based precision oncology: clinical cohorts such as TCGA provide pre-treatment tumor expression but almost never paired post-treatment molecular profiles, so models must either rely on static preclinical associations or on perturbation priors learned from cell-line data. The proposed framework couples a patient-specific mechanistic view with a transferable perturbation view, aligns them contrastively, and predicts binary clinical response end-to-end.

Motivation and positioning

Prior work falls into two families. Preclinical-centered models such as DeepCDR and DeepTTA learn static mappings from cell-line multi-omics to IC50 readouts and transfer poorly to patients because they do not represent drug-induced dynamics or patient-specific regulatory context. Knowledge-guided and perturbation-informed models (DRPreter, Precily, CSG²A) improve biological plausibility and transfer, but remain driven by preclinical representations and do not exploit patient-level heterogeneity. PREDIKTOR's premise is that these two information sources are complementary and can be fused per patient-drug pair, even without matched post-treatment measurements.

Architecture

The framework comprises two encoders whose outputs are concatenated for a response MLP.

Network view. For each patient, DysRegNet infers dysregulated TF–target interactions against matched healthy tissue baselines (HTRIdb-constrained pairs, BH FDR < 0.05), yielding a personalized GRN over a 7,800-gene universe. A drug node is attached via DrugBank target links, node embeddings are initialized with DREAMwalk, and a two-layer GCN (hidden dimension 128) produces a drug-centric embedding, concatenated with layer-normalized patient expression.

Perturbation view. A condition-specific gene–gene attention module (CSG²A), pretrained on LINCS L1000 with PPI-constrained attention and MAT-encoded drug structure, is frozen during PREDIKTOR training. It generates a simulated post-perturbation 978-gene profile for each patient-drug pair by attention-weighted aggregation of baseline expression.

Contrastive alignment. A symmetric CLIP-style loss with a learnable temperature aligns the two projected views. A key empirical finding is the negative-selection strategy: hard negatives drawn from pairs sharing the patient but differing in drug ($\mathcal{L}_{\text{CLIP-context}}$) outperform negatives based on tissue or supervised labels, which introduce semantic noise by conflating biological variability with treatment effects. The total objective is BCE plus a weighted CLIP term ($\lambda_{\text{CLIP}}$ initialized at 5), trained end-to-end with Adam; CSG²A and MAT remain frozen to preserve transferable priors, consistent with earlier observations that fine-tuning CSG²A degrades downstream performance.

Predictive performance

On 383 TCGA patient-drug pairs (358 patients, 21 drugs), evaluated with stratified 5-fold cross-validation under three mutually exclusive splits, PREDIKTOR achieves the best results on most metrics:

Setting	AUROC	AUPRC
Patient split	0.837 (±0.034)	0.829 (±0.049)
Drug split	0.594 (±0.090)	0.565 (±0.252)
Tissue split	0.603 (±0.085)	0.595 (±0.284)

Two observations temper these results. In the patient-split setting, most baselines already exceed 0.80 AUROC, so the margin over competitors is modest. In the drug-split setting—arguably the most clinically relevant scenario for novel compounds—absolute performance is near 0.59 AUROC, and knowledge-guided and transformer baselines collapse below chance while simpler models remain comparatively stable; PREDIKTOR retains the best average but the gap is defined against weak baselines. The high AUPRC standard deviations across all methods reflect the small cohort size.

Ablations attribute the gains to both views and to personalization: removing the perturbation module costs 0.151/0.146 AUROC (drug/tissue splits), removing the network view costs 0.109/0.118, and replacing patient-specific GRNs with a static shared topology significantly degrades performance (paired t-test, p < 0.05). Removing hard negatives or the contrastive objective altogether reduces drug-split AUROC to 0.574 and 0.559 respectively, indicating that the hard-negative strategy, not contrastive learning per se, drives much of the drug-split robustness.

External validation

In a zero-shot transfer to 178 I-SPY2 breast cancer patients treated with paclitaxel (pCR labels), PREDIKTOR reaches AUROC 0.661 (±0.043), AUPRC 0.296, and F1 0.306, a 5.6% AUROC improvement over the best competitor (GeneFormer+ECFP, 0.626). The stability claim is notable: PREDIKTOR's AUROC standard deviation is roughly half that of the transfer-learning baselines (~0.08). Still, 0.661 AUROC on a single-drug, single-tissue cohort is moderate in absolute terms, and the evaluation covers one compound, limiting the strength of the generalization claim.

Interpretability

Gene-level attributions via expected gradients distribute mass more broadly than TreeExplainer on a Random Forest baseline, and GSEA on the top-ranked genes recovers paclitaxel-relevant pathways (cell cycle, proteoglycans in cancer, chemokine signaling, p53, PI3K–Akt), whereas the baseline's eight enriched KEGG pathways have no mechanistic connection to paclitaxel response. Cross-referencing with LINCS L1000 paclitaxel consensus signatures yields significant enrichment for both upregulated (64/241 genes, p = 1.4E-7) and downregulated (62/249, p = 3.8E-6) signatures, with 16 consensus genes in the top 100 attributions; the baseline recovers only one (UCP2). Overlap counts are stable across random seeds (12–17 of top 100). A GO-term prediction probe further shows the network module's embeddings carry roughly 1.5–3× the functional information ($F_{max}$, AUPRC) of the perturbation module across all three GO domains, supporting the claim that mechanistic priors improve semantic coherence.

Limitations

The authors are candid that drug-split generalization remains the weak point: the network view requires known DrugBank targets, so cold-start or investigational drugs without annotated targets cannot be attached to the patient GRN, and the perturbation view relies on SMILES alone. The TCGA cohort is small and skewed in drug and tissue coverage, and the underlying knowledge graphs are incomplete. The zero-shot validation involves a single drug and tissue, so cross-domain transfer breadth is untested. Proposed remedies—inferred drug-target edges, richer drug representations, and adaptive (attention/gated) fusion—remain unimplemented.

Conclusion

PREDIKTOR demonstrates that contrastively aligning a patient-specific knowledge graph view with a frozen, transfer-learned perturbation view yields the best reported performance on TCGA drug response prediction across patient-, drug-, and tissue-split protocols and a modest but stable zero-shot gain on I-SPY2, while producing attributions that recover known pharmacology. The main open question left by the paper is whether the framework retains its advantage for compounds lacking curated targets and perturbation data, where both views currently degrade.