Technical Summary: ProToxNet

Problem Statement

Predicting which drugs cause specific adverse events (AEs) and in which tissues remains a critical unsolved challenge in computational drug safety. Current approaches generally fall into two categories: structure-based methods (using molecular fingerprints or graph neural networks) and pharmacovigilance mining (extracting signals from databases like FAERS). Both share a fundamental limitation: they treat adverse event risk as an intrinsic property of the drug in isolation, ignoring the tissue-specific expression context. The paper argues that whether an off-target binding event manifests as clinical harm depends not only on binding affinity but on the relative expression of the target protein across different tissues (e.g., skin, liver, lung).

Methodology

ProToxNet is a computational framework that integrates proteome-wide predicted drug–protein binding with tissue-specific gene expression to generate 68-dimensional tissue exposure profiles. The pipeline consists of three stages:

1. Proteome-Wide Binding Scoring (Stage 1):

   • Data: 4,310 drugs and 1,507 core human proteins (the intersection of DrugCentral, GTEx, and STRING).
   • Model: A variant of ConPLex using the ESM-1b protein language model.
   • Process: Drug Morgan fingerprints (2048-bit) and ESM-1b mean-pool protein embeddings (1280-dimensional) are projected into a shared 1024-dimensional cosine similarity space.
   • Calibration: Binding probabilities are calibrated via Platt scaling using known DrugCentral binding pairs. This yields a binding score matrix $S \in \mathbb{R}^{4310 \times 1507}$.

2. Tissue Exposure Scoring (Stage 2):

   • Data: Gene expression data from the GTEx v10 project across 68 human tissues.
   • Computation: Tissue exposure vectors are generated by weighting the binding scores by tissue-specific expression levels.
   • Formula: For a drug $d$ and tissue $t$, the exposure score is:
     $$e_{d,t} = \sum_{p \in P} s_{d,p} \cdot x_{p,t}$$
     where $s_{d,p}$ is the binding score and $x_{p,t}$ is the normalized expression of protein $p$ in tissue $t$.
   • Preprocessing: The resulting matrix $E$ is z-scored per tissue to correct for bulk expression magnitude differences, yielding the final feature matrix $\tilde{E} \in \mathbb{R}^{4310 \times 68}$.

3. Adverse Event Prediction (Stage 3):

   • Data: 302,307 positive drug-AE signal pairs from the FDA Adverse Event Reporting System (FAERS) with log-likelihood ratio (LLR) > 0.
   • Model: A diagonal bilinear interaction model.
     • Drug Encoder: A two-layer neural network (68 → 256 → 128) with LayerNorm, ReLU, and dropout.
     • AE Embeddings: 13,200 MedDRA terms mapped to 64-dimensional embeddings, projected to 128 dimensions.
     • Scoring: The interaction score is computed as $score(d, a) = \sum_{k=1}^{128} h_d^{(k)} \cdot W_k \cdot h_a^{(k)} + b_d + b_a$, where $W$ is a trainable diagonal weight matrix and $b$ represents bias terms for reporting frequency.
   • Training: Optimized with AdamW, label smoothing, and cosine annealing over 100 epochs.

Key Contributions

1. Expression-Weighted Representation: Introduction of a 68-dimensional tissue engagement potential vector that combines proteome-wide binding with tissue expression, serving as a drug representation for AE prediction.
2. Performance without Graph Propagation: Demonstration that this vector representation, without explicit graph message passing over the protein-tissue network, achieves state-of-the-art performance.
3. Biological Validation: Direct validation that known hepatotoxic drugs show significantly higher liver tissue engagement scores than non-hepatotoxic drugs.
4. Case Study: Successful recovery of known adverse event profiles for approved kinase inhibitors, with known AEs ranking in the top 0.67% of all MedDRA terms.

Results

The model was evaluated across multiple protocols:

• FAERS Pair-Level Prediction: Achieved an AUC of 0.9576 (95% CI: 0.9565–0.9588) on held-out pairs.
• Cold-Start (Leave-Drug-Out): In a scenario where 244 drugs were held out entirely (zero training pairs), the model achieved an AUC of 0.8544. This confirms the model generalizes to novel drugs based on tissue engagement features rather than memorizing drug identities.
• Leave-AE-Out: When 1,980 MedDRA terms were held out, the model achieved an AUC of 0.8472, demonstrating generalization to unseen adverse events.
• External Validation (CT-ADE): On the independent CT-ADE clinical trial benchmark, ProToxNet achieved an AUC of 0.9157.
• Biological Enrichment:
  • Hepatotoxicity: Known hepatotoxic drugs had significantly higher liver tissue exposure scores than non-hepatotoxic drugs (Mann-Whitney U, $p = 0.017$, Cliff's $\delta = 0.41$).
  • DILIrank Validation: "Most-concern" hepatotoxic drugs showed significantly higher liver exposure than "no-concern" drugs ($p = 0.0002$, Cliff's $\delta = 0.59$).
• Kinase Inhibitor Case Study: Known adverse events for eight approved kinase inhibitors ranked at a median of 88 out of 13,200 terms (top 0.67%). For example, hepatotoxicity was ranked in the top 1% for imatinib, erlotinib, and lapatinib.

Significance and Claims

The paper claims that ProToxNet provides a mechanistically grounded, tissue-aware framework for adverse event prediction. By operationalizing the principle that adverse event risk is a joint property of a drug's binding landscape and the tissue-specific expression context, the framework bridges the gap between molecular interaction and organ-level toxicity.

The authors emphasize that the model's success in cold-start scenarios (novel drugs and novel AEs) suggests that tissue engagement potential captures transferable pharmacological signals beyond simple entity frequency or chemical structure memorization. The framework is presented as a tool applicable to drug discovery (early screening for tissue-specific toxicity) and pharmacovigilance (prioritizing safety signals), offering a complement to existing structure-based and mining-based methods.

Limitations Acknowledged by the Authors:

• The binding scores rely on a reinitialized target projector due to environment compatibility issues, meaning they reflect structural features more than fully pretrained affinity estimates.
• FAERS data contains reporting biases (e.g., larger patient populations generate more reports).
• The model does not account for drug dose, administration route, patient demographics, or drug-drug interactions.
• Tissue enrichment was statistically significant for hepatotoxicity but not for cardiotoxicity or nephrotoxicity in the tested sets, likely due to mechanism specificity (e.g., hERG blockade) or small sample sizes.