MARS: A neurosymbolic approach for interpretable drug discovery

Technical Summary: MARS – A Neurosymbolic Approach for Interpretable Drug Discovery

Problem Statement
Drug discovery (DD) traditionally relies on screening thousands of chemical compounds, a process often accelerated by Knowledge Graphs (KGs). While standard KG-based methods excel at link prediction (e.g., matching drugs to indications), they often lack interpretability regarding a drug's Mechanism of Action (MoA). MoA deconvolution—the computational prediction of the molecular paths by which a drug achieves its effect—is critical for de-risking side effects and predicting clinical success. However, existing interpretable methods often rely on "associative patterns" (e.g., shared side effects or indications) rather than "mechanistic patterns" (physical molecular interactions). Furthermore, there are no clear guidelines to assess the biological plausibility of model interpretations. A significant challenge identified in Neurosymbolic (NeSy) AI is the "reasoning shortcut," where models achieve high accuracy by exploiting unintended semantics, such as node degree bias, rather than adhering to domain-specific logical rules.

Methodology
The authors propose MARS (MoA Retrieval System), a NeSy approach designed to perform MoA deconvolution. The methodology consists of three core components:

1. MoA-net (Dataset): A tailored KG constructed specifically for MoA deconvolution. Unlike general biomedical KGs, MoA-net focuses on entities and relations discoverable in pre-clinical laboratory settings. It comprises triples linking Drugs, Proteins, and Biological Processes (BPs) via causal relations (e.g., upregulates, downregulates) and interactions. The dataset includes 300 drugs, 9,301 proteins, and 86 BPs.
2. NeSy Framework: MARS utilizes a Deep Reinforcement Learning (RL) agent to traverse MoA-net. The agent's goal is to find paths connecting a Drug node to a BP node (the induces relation).
   • Metapath-based Rules: The system employs logical rules where the body consists of metapaths (sequences of relations) and the head is the target induces relation.
   • Dynamic Weight Learning: Unlike previous approaches (e.g., PoLo) that use static, pre-computed rule weights, MARS learns rule weights dynamically during training.
3. Weight Update Mechanisms: The paper introduces two methods for updating rule weights to make the system "shortcut-aware":
   • Naive Updates: Adjusts weights based on the frequency of rule satisfaction relative to expected frequency.
   • P2H (Two-Hop Joint Probability): A novel algorithm that approximates rule usefulness by analyzing the joint probability of two-hop fragments within successful agent trajectories. This allows the system to update weights based on partial path matches, providing a more robust measure of rule importance.
4. Evaluation Strategy: The system is evaluated on link prediction tasks using standard metrics (Hits@k, MRR) and "pruned metrics" (calculated only on predictions that satisfy a logical rule). The study specifically tests for "reasoning shortcuts" by comparing performance on the original MoA-net against a permuted version (MoA-net-permuted) and a sparse version (MoA-net-10k) to isolate the effects of node degree bias.

Key Results

• Reasoning Shortcuts: Initial experiments revealed that MARS, along with baseline NeSy models like PoLo, frequently relied on "associative patterns" (specifically, inverse edges suggesting that drugs inducing a common BP likely induce another BP) rather than the intended mechanistic rules. This behavior was driven by degree-bias, where the model exploited the high connectivity of certain nodes to maximize rewards without learning the underlying biological logic.
• Mitigation: By applying the P2H weight update strategy and training on a trimmed version of the graph (MoA-net-10k), the authors demonstrated that the model could be made "shortcut-aware." This mitigation reduced the reliance on degree-bias and forced the agent to utilize the domain-based logical rules.
• Performance: The optimized MARS (specifically the P2H variant) achieved performance on par with state-of-the-art KG embedding models and NeSy baselines. Crucially, unlike the baselines, MARS produced interpretations aligned with known MoAs when the reasoning shortcuts were mitigated.
• Interpretability: The system successfully generated paths between drugs and BPs that corresponded to known mechanistic pathways (validated against DrugMechDB), demonstrating that the learned weights and trajectories reflect biologically meaningful patterns rather than statistical artifacts.

Significance and Claims
The paper claims that MARS showcases a novel computational task: MoA deconvolution. Its primary significance lies in demonstrating that while NeSy models offer enhanced interpretability, they remain susceptible to reasoning shortcuts that can render their "explanations" biologically implausible.

The authors emphasize that:

1. Evaluation is Contextual: Interpretability must be assessed against specific domain knowledge (mechanistic rules) rather than just statistical performance metrics.
2. Shortcuts are Mitigatable: By identifying degree-bias and employing strategies like P2H weight updates and graph trimming, NeSy models can be guided to produce reliable, biologically meaningful predictions.
3. Reliability for Downstream Research: By ensuring that predictions are driven by domain rules rather than statistical biases, MARS provides more reliable MoA predictions for downstream drug discovery research, offering a unique AI-based alternative to traditional, non-scalable laboratory assays.

The study concludes that for NeSy approaches to be truly trustworthy in biomedical applications, they must be designed and evaluated with specific mechanisms to detect and mitigate reasoning shortcuts.