Interpretable tissue-resolved drug toxicity prediction with pathway-level mechanism attribution

Imagine you are a chef trying to create the perfect new dish. You have a recipe (the drug) and a kitchen (the human body). The problem is that the human body isn't just one big kitchen; it's a city of different neighborhoods, each with its own unique rules, ingredients, and sensitivities. A spice that makes the "Liver District" sizzle might make the "Heart District" collapse, while the "Brain District" doesn't even notice. For decades, scientists have struggled to predict which neighborhoods a new drug will ruin before they actually serve it to patients. Usually, they just guess based on what the drug looks like chemically, missing the fact that the location matters just as much as the ingredients. This is the heart of drug toxicity: figuring out not just if a medicine is dangerous, but where in the body it might cause trouble and why.

Enter ToxiPred, a new digital detective developed by researchers at the Norwegian University of Science and Technology and the University of Helsinki. Think of ToxiPred as a super-smart, two-tower robot that doesn't just look at a drug's chemical shape; it also studies the "personality" of every tissue in the body. The robot has a special trick: it forces the drug and the tissue to meet in a narrow hallway of 70 specific "pathways" (like biological assembly lines inside our cells). This hallway acts as a translator, revealing exactly how a specific drug messes with a specific tissue's machinery. If the robot predicts a drug will hurt the heart, it doesn't just say "Danger!"; it points a finger and says, "It's messing up the cholesterol factory in the heart cells."

The team tested this robot on thousands of drugs, including some it had never seen before. They found that ToxiPred is surprisingly good at spotting trouble. When they checked it against known toxic drugs, it correctly identified the dangerous ones about 85% of the time for liver issues and 73% of the time for general clinical trial failures. But the real magic happened when they tested it on drugs used for mental health, like quetiapine and sertraline. The robot predicted these drugs would be tough on the heart, specifically by disrupting lipid (fat) and cholesterol pathways. To prove it wasn't just a computer hallucination, the scientists grew tiny, living "mini-organs" (organoids) of human hearts, brains, and retinas in a lab. When they added the drugs, the mini-hearts stopped beating and showed the exact same fat-related stress signals the robot had predicted.

This isn't a magic wand that replaces all safety testing, but it's a powerful new flashlight. It suggests that by understanding the specific biological pathways a drug targets in different tissues, we can prioritize which medicines need extra safety checks and which ones are likely safe to use. It turns a blind guess into a targeted investigation, potentially saving time, money, and most importantly, patient safety.

The Story of the Two-Tower Robot

The Problem: The "One-Size-Fits-All" Mistake
Imagine you have a key (a drug) and a bunch of different locks (tissues like the liver, heart, and brain). For a long time, scientists tried to predict if the key would break the lock just by looking at the shape of the key's teeth. They often missed the fact that some locks are made of brittle glass (sensitive tissues) while others are made of steel (tough tissues). If a key fits a steel lock perfectly, it might shatter the glass lock. The old computer models were like that: they gave a single score for "is this drug toxic?" without telling you which part of the body would get hurt or how it would happen.

The Solution: ToxiPred's Two Towers
The researchers built ToxiPred, a neural network (a type of computer brain) that works like a high-tech matchmaking service. It has two main towers:

1. The Drug Tower: This tower looks at the drug's chemical structure, its known targets, and how it changes gene activity in cells.
2. The Tissue Tower: This tower studies the specific "vibe" of different tissues, looking at which genes are active and which are essential for survival in the liver, heart, brain, etc.

These two towers meet in the middle at a "bottleneck"—a narrow hallway with exactly 70 pathways. Think of these pathways as 70 different assembly lines in a factory. The robot forces the drug and the tissue to interact only through these 70 lines. If the drug jams the "Cholesterol Assembly Line" in the "Heart Factory," the robot sees it immediately. This design is crucial because it doesn't just predict a number; it explains the mechanism. It tells you, "This drug is toxic to the heart because it clogs the cholesterol line."

The Training: Learning from the Past
To teach the robot, the scientists fed it data from 9,275 different drug-and-tissue combinations found in public databases. They taught it to predict a "Drug Sensitivity Score" (called DSStox) on a scale of 0 to 50.

• 0–3.0: The drug is likely safe (low liability).
• 3.0–5.0: The drug is in the "maybe" zone (borderline).
• 5.0+: The drug is likely dangerous (high liability).

They tested the robot on drugs it had never seen before, using a "drug-disjoint" method. This means they hid entire families of chemical structures during training and asked the robot to guess on completely new shapes. It got it right with a correlation of 0.73, which is a very strong score for such a complex task.

The Real-World Test: The Anti-Infective Challenge
First, they tested the robot on 93 anti-infective drugs (antibiotics and antivirals). These drugs are designed to kill bacteria or viruses, not human cells. The robot correctly predicted that these drugs would have low toxicity to human tissues (scores below 3.0), while flagging a few reference drugs known to be toxic to human cells. When they tested these drugs on six different human tissue systems (lung, liver, intestine, kidney, pancreas, and breast) in the lab, the robot's predictions matched the real-world results with a correlation of 0.78 to 0.84. This proved the robot wasn't just guessing; it understood the difference between killing a bug and hurting a human.

The Deep Dive: The Psychotropic Mystery
The most exciting part of the story involved 91 psychotropic drugs (medicines for mental health like antidepressants and antipsychotics). About 98% of these drugs were completely new to the robot; it had never seen their chemical structures before. The robot scanned them and flagged several, including quetiapine and sertraline, as having high toxicity risks, specifically pointing to lipid and cholesterol pathways.

To see if the robot was right, the scientists grew human organoids (tiny, 3D versions of human organs) in a dish. They used:

• Forebrain organoids
• Cardiac (heart) organoids
• Retinal (eye) organoids

They treated these mini-organs with the drugs and watched what happened.

• The Heart Test: When they added quetiapine and sertraline, the tiny heart organoids stopped beating at a concentration of 20 µM. This was a huge red flag that matched the robot's prediction.
• The Molecular Test: They took a snapshot of the genes inside the heart cells. They found that sertraline and quetiapine turned on genes related to cholesterol biosynthesis and lipid metabolism. This perfectly matched the robot's "pathway attribution."
• The Contrast: Another drug, aripiprazole, showed a different pattern, affecting muscle development genes instead. This showed the robot could distinguish between different types of toxicity, not just a generic "bad" score.

The Multi-Omics Follow-Up
To get even deeper, the researchers studied quetiapine in detail. They used a clever trick: they treated the organoids with retinoic acid (a known stressor) and then added quetiapine. This helped isolate exactly what quetiapine was doing. They found that quetiapine suppressed genes related to lipoprotein transport and cholesterol homeostasis. They also looked at the metabolites (the tiny chemical building blocks) and found changes in fumarate, malate, and acylcarnitines, suggesting the drug was messing with how the cells handle energy and fat.

What This Means
The paper concludes that ToxiPred is a powerful tool for prioritization. It's not a crystal ball that guarantees safety, but it's a highly effective filter. It can take a list of thousands of new drugs and say, "Check these ones out first; they look like they might clog the cholesterol factory in the heart."

The researchers are careful to note that this is a mechanism-guided framework, not a replacement for clinical trials. The organoids are simplified models, and the data comes from cell lines that might not perfectly mimic a whole human body. However, the fact that the robot's predictions about pathways (like cholesterol) matched the actual biological changes in living human cells is a massive step forward. It suggests that we can now move from asking "Is this drug toxic?" to "How, where, and why is this drug toxic?" before we ever give it to a patient.

In the end, ToxiPred is like a translator that speaks both "Chemical" and "Biological." By forcing these two languages to meet in a narrow hallway of 70 pathways, it reveals the hidden stories of how drugs interact with our bodies, offering a clearer, more hopeful path to safer medicines.