🧬 biology

CausalDiffChem: Causal Network Restoration for Therapeutic Prioritization and Repurposing Across Diseases

CausalDiffChem is a novel framework that leverages transcriptomic data and structural causal models to prioritize drug repurposing candidates by simulating interventions that restore disease-specific causal network structures, thereby distinguishing causal drivers from downstream effects across diverse conditions.

Original authors: Francis Armah, Evans Abor, Pius Sam, Emmanuel Sarpong, George Ghartey-Kwansah

Published 2026-10-26

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Francis Armah, Evans Abor, Pius Sam, Emmanuel Sarpong, George Ghartey-Kwansah

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Finding a new use for an old drug is often a matter of looking for a match. For decades, scientists have tried to repurpose medications by comparing the molecular fingerprints of a disease against the fingerprints of drugs. If a drug seems to flip the disease's genetic switches in the opposite direction, researchers assume it might work as a treatment. This approach has found some real successes, but it has a blind spot. It treats the disease as a static list of symptoms, ignoring the fact that genes talk to one another in complex chains of cause and effect. A gene might be turned on not because it is the root of the problem, but because it is reacting to something else happening upstream. Treating the reaction without fixing the cause can be like trying to stop a fire by blowing on the smoke.

To move beyond this limitation, a team of researchers has developed a new way to map the hidden machinery of disease. Instead of just listing which genes are active, they built a model that shows how those genes influence each other, creating a causal network. They then tested how different drugs would ripple through this specific network to see if they could restore the system to a healthy state. Their work, which they call CausalDiffChem, suggests that by understanding the specific cause-and-effect structure of a disease, we can identify better candidates for treatment, even when those candidates are not obvious from a simple list of gene activity.

The researchers applied this method to four very different human conditions: multiple sclerosis, a disease of the immune system and nerves; sarcoidosis, a condition causing inflammation in the lungs and other organs; leishmaniasis, a parasitic infection; and Alzheimer's disease, a form of dementia. For each disease, they gathered data from hundreds of patients and healthy people. They did not use any pre-existing maps of how genes interact. Instead, they let the data speak for itself, using a computer algorithm to reconstruct a unique, cause-and-effect map for each disease based on the twenty most active genes in the samples. These maps revealed distinct architectures. In Alzheimer's, the network was dominated by a strong connection between two measurements of a protein called apolipoprotein D. In sarcoidosis, the map highlighted a regulatory axis involving immune cells known as macrophages. Each disease had its own unique shape and flow of information, proving that a single, generic approach could not capture the specific logic of every illness.

Once these disease maps were built, the team simulated the effect of thousands of drugs. They did not test the drugs in a petri dish or a living body; instead, they introduced a mathematical representation of each drug into the computer model of the disease network. They watched to see if the drug's simulated presence could push the network back toward the state seen in healthy people. They measured this "restoration" in three ways: how many parts of the network moved in the right direction, how strong that movement was, and whether the overall shift matched the path from sickness to health.

The results offered a fresh perspective on which drugs might work. For sarcoidosis, a chemotherapy drug called cisplatin emerged as the top candidate, ranking first among approved medicines. In leishmaniasis, memantine hydrochloride, a drug already used for Alzheimer's, ranked first across all stages of drug development, from early research to clinical trials. In Alzheimer's itself, a different drug, mecamylamine hydrochloride, took the top spot. What made these findings notable was that the researchers had built the disease maps and tested the drugs independently for each condition, with no shared settings or pre-set rules. The fact that the same drugs kept appearing as top candidates across different libraries and different diseases suggests the method is finding a genuine signal rather than a random coincidence.

However, the study also revealed that a high score does not tell the whole story. A drug might fix a large number of genes but fail to correct the most important connections, or it might fix the right genes but push the rest of the network in the wrong direction. In Alzheimer's, the top-ranked drug corrected many genes but left the strongest connection in the disease network—the link between the two apolipoprotein D measurements—largely untouched. In sarcoidosis, cisplatin strongly corrected a key group of immune genes but pushed other genes further away from a healthy state. This shows that a drug can be a strong candidate even if it does not fix everything perfectly, and that different drugs achieve their effects through different mechanisms.

The researchers also checked if their method could find drugs that are already known to work. In Alzheimer's, two established treatments, donepezil and memantine, appeared in the top ten percent of their rankings, even though they were not the number one pick. In other diseases, known treatments also ranked well, though some, like certain steroids, ranked lower. This suggests the computer model captures real biological truths about how these drugs work, but it also highlights that the model is not a perfect replacement for clinical knowledge. It focuses on a specific set of genes and may miss broader effects that happen elsewhere in the body or depend on how a drug is delivered.

The team acknowledges that their work is a simulation. The drug effects were estimated based on chemical structures rather than measured in a lab, and the disease maps were built from relatively small groups of patients. They did not prove that these drugs will cure these diseases in people. Instead, they provided a new, interpretable framework for generating hypotheses. By showing how a drug might travel through a specific disease's causal network, they offer a way to prioritize which candidates deserve to be tested in the lab. The repeated appearance of the same drugs across different diseases and different stages of development suggests that this approach can identify promising leads that might be missed by looking at gene lists alone, offering a more nuanced path toward finding new uses for old medicines.

Technical Summary: CausalDiffChem

Problem Statement
Current computational drug repurposing strategies predominantly rely on transcriptomic signature matching, identifying compounds that reverse disease-associated gene expression profiles. While effective in some contexts, this paradigm is fundamentally correlational; it fails to distinguish between genes that act as causal drivers of pathology and those whose altered expression is a downstream consequence. Furthermore, signature-based predictions often suffer from reproducibility issues across independent datasets. There is a need for a framework that moves beyond static expression similarity to model the directed, causal relationships within disease-specific regulatory networks, thereby enabling more robust therapeutic prioritization without relying on curated pathway annotations.

Methodology
The authors introduce CausalDiffChem, a causal-network-guided framework that reconstructs disease-specific Structural Causal Models (SCMs) directly from transcriptomic data. The methodology proceeds through the following steps:

Data Preparation: Transcriptomic datasets for four distinct diseases (Multiple Sclerosis, Sarcoidosis, Leishmaniasis, and Alzheimer's Disease) comprising 408 disease samples and 374 healthy controls were utilized. For each disease, the 20 most variable genes were selected to construct the network.
SCM Reconstruction: Using the NOTEARS (Notears) continuous optimization algorithm, the framework learns a weighted directed acyclic graph (DAG) representing the causal relationships among the selected genes. This process minimizes a least-squares reconstruction loss subject to a smooth acyclicity constraint, utilizing an L1 sparsity penalty. No curated pathway information or manually specified regulatory relationships are shared between disease modules.
Perturbation Generation: Candidate compounds (from FDA-approved, clinical-stage, and preclinical libraries) are represented by molecular fingerprints (ECFP4 and ECFP6) and a drug-likeness scale (derived from QED, logP, and TPSA). These are combined to generate a perturbation vector ( $\delta_T$ ).
Simulated Intervention: The perturbation vector is introduced into the reconstructed SCM via a simulated intervention, $do(T)$. The post-intervention state is propagated through the inferred linear structural causal model using a closed-form solution.
Restoration Metrics: The therapeutic effect is quantified using three complementary metrics:
- net_correction: The mean direction-aware fractional restoration of network nodes toward the healthy reference state.
- %H (Breadth): The percentage of nodes corrected toward the healthy state.
- R (Global Alignment): The cosine similarity between the intervention-induced shift and the disease-to-healthy expression shift.
Prioritization: Compounds are ranked based on net_correction. The framework also includes a positive-control analysis to check if established therapeutics receive favorable scores within the full library rankings.

Key Results

Biologically Coherent Networks: Independently reconstructed SCMs for the four diseases revealed distinct regulatory architectures. Notable findings included a dominant APOD interaction in Alzheimer's Disease, a macrophage-associated regulatory axis in Sarcoidosis, and neutrophil-associated interactions in Multiple Sclerosis.
Disease-Specific Prioritization: The framework identified distinct leading candidates for each disease. For example, Cyclizine hydrochloride ranked first for Multiple Sclerosis, Cisplatin for Sarcoidosis, Memantine hydrochloride for Leishmaniasis, and Mecamylamine hydrochloride for Alzheimer's Disease.
Cross-Disease and Cross-Library Convergence: Despite independent network reconstruction, certain compounds recurred across different disease modules and development stages. Cisplatin appeared in the top ten for three of the four diseases, and Memantine hydrochloride ranked first across all three compound libraries (FDA-approved, clinical, preclinical) for Leishmaniasis with identical restoration scores.
Complementary Dimensions of Response: The study demonstrated that net_correction, breadth (%H), and global alignment (R) capture distinct aspects of therapeutic response. High aggregate restoration could coexist with limited correction of dominant network interactions or discordant global alignment. For instance, Cisplatin in Sarcoidosis showed strong node-level restoration but negative global alignment, while Mecamylamine hydrochloride in Alzheimer's Disease restored specific nodes without fully correcting the dominant APOD/APOD_1 axis.
Positive Control Validation: Established therapeutics (e.g., Donepezil and Memantine in Alzheimer's; Siponimod in Multiple Sclerosis) received favorable net_correction scores, often ranking in the top 10–20% of the library, though they were not always the top-ranked candidates. This suggests the framework captures aspects of therapeutic biology but does not perfectly replicate systemic clinical efficacy.

Significance and Claims
The paper claims that CausalDiffChem provides an interpretable framework for prioritizing therapeutics based on their predicted effects on disease-specific regulatory structures, moving beyond simple correlation-based signature reversal. By reconstructing SCMs without curated pathway annotations, the method generates experimentally testable hypotheses for drug repurposing across heterogeneous diseases.

The authors emphasize that the recurrence of candidates across independently reconstructed networks suggests that causal-network restoration can identify therapeutic signals extending beyond single disease signatures. However, the paper remains modest regarding clinical translation, acknowledging that the framework evaluates a compact transcriptional network and may not capture systemic, tissue-specific, or pharmacokinetic mechanisms. The gene-level analyses further clarify that high net_correction does not guarantee global alignment with the healthy state, necessitating a multi-metric evaluation of therapeutic potential. The work establishes causal-network restoration as a complementary strategy for drug repurposing, generating hypotheses that require experimental validation.

More like this