When I was younger, I thought drugs were problems in the other countreis. In South Korea, it's not allowed... Nowadays, it seems to be a serious problem. I haven't seen any incidents around me, but I often see it on the news or in ads, to prevent use of drugs particularly in school or among students. I feel that it's getting serious than before. While I was watching news of people addicted to drug, it suddenly came to me that I may predict the toxicity of drugs. I'm not expert on drug, so I did little bit of search. I decided to compare the toxicity of tetrahydrocannabinol (THC) and cannabinol using VEGA. These two copmounds are ingredients of marijuana. THC is the one that actually causes hallucinations. There are actually a lot of different uses for canabbis: Sleep aid, pain reliever, appetite stimulant. Some people even cook and eat it as food. I've never tried it, so I don't say about the taste. If you search for cannabis on Coursera, there are various courses. Cannabis courses are about medical uses of cannabis.
I first checked smiles code from PubChem. First prediction was cannabinol.
I ran calculations with all the models in VEGA. There are 63 models for human toxicity and ADME (absorption, distribution, metabolism, and excretion: these are all steps from the compound getting into and removed from the body). Since I don't want to see all 63 values in detail, I decided to view the Summary, which saves all the results in a single csv file.
The summary file doesn't have a detailed report on applicability domain (AD) analysis. However, it divides the AD into four categories
1. EXPERIMENT: There are experimental values in the data. (Experimental values are much more reliable than predicted values)
2. GOOD: the predicted value is very likely to be correct.
3. MODERATE: Moderate confidence in the prediction. Not sure if the value is accurate or not.
4. LOW: Prediction confidence is low. Don't trust the prediction.
Of these, only EXPERIMENT and GOOD are reliable. I only listed up predictions with EXPERIMENT and GOOD.
There are two cases where there are experimental results. (Experiment)
1. Carcinogenicity (female rat) TD50: 243.22 mg/kg
2. Genotoxicity: Toxic O
The following 5 cases are predictions with high confidence. (Good)
1. reproductive toxicity: Toxic O
2. Carcinogenicity: Toxicity X
3. Acute toxicity: 2306 mg/kg
4. Thyroid: Toxicity X
5. Steatosis: Toxicity X
Animal tests are conducted on multiple animals. The 50 in TD50 means 50%. The concentration at which 50% of the animals tested develop cancer is the TD50. According to my research, retail cannabis contains about 10 mg of THC. The amount of cannabinol is much lower. So it's probably not carcinogenic.
Genotoxicity results are based on in vivo micronucleus tests. When the nucleus of a cell is fragmented, it is called a micronucleus. When the nucleus is fragmented, it means that the genetic information has somehow been adversely affected. Genes are contained in DNA. And DNA is packaged in the cell nucleus in the form of chromosomes. When the nucleus is fragmented, it means that something is wrong with the chromosomes. So, according to the predictions, cannabinol causes damage to genes. There are many causes of cancer, but genetic damage is one of them.
The second toxicity identified is developmental toxicity. This refers to the potential for birth defects. As the fetus develops limbs and organs in the mother"s womb, substances that adversely affect this process are called developmental toxicants. As predicted, cannabinol has the potential to cause developmental toxicity.
The QMRF documentation for the developmental toxicity model shows that the model was developed using human and animal data, which indicates that the prediction value is highly relevant to the human toxicity.
Acute toxicity is the toxicity that occurs when a large amount is ingested in a short period of time. The acute toxicity values are quite high, so it's probably not acutely toxic in humans.
What about THC?
The structure is almost identical to cannabinol. The differences are circled in red. I thought that the predicted values would be roughly similar if they have this much similar structure.
Experimental values
1. Genotoxicity: Toxicity O
2. carcinogenicity: toxicity X
3. NOAEL: Unit is unclear... (Predictions are 5mg/kg or 1572mg/kg. Why is there such a difference? The prediction report mentioned the unit is mg/kg, but the QMRF said mmol/kg. It's unclear which one is correct.)
4. Plasma protein binding (PPB): 1.279
5. Aromatase: active antagonist
Good
1. genotoxicity: Toxicity X
2. developmental toxicity: toxicity O
3. Estrogen: Active
4. thyroid: Inactive
5. half-life: 9.86 hours
6. fatty liver: Cautious
Two endpoints had same prediction values between THC and cannabinol.
The values from the micronucleus test were predicted as toxic for both THC and cannabinol.
In terms of developmental toxicity, both were found to be toxic.
For THC, a much larger number of models provided predictions that were in good agreement with experimental values.
Plasma protein binding is a ratio of the compound binding to proteins in the blood. Even if a substance is absorbed, the substance that binds to proteins in the blood, will be excreted doing nothing. Therefore, a substance with high PPB is unlikely to show much effect in the body. A PPB of 0 means that the ratio of protein-bound to unbound substance is 1:1. A value greater than zero indicates more binding to protein, and in the case of THC, it seems to bind to proteins a lot. Among the predicted values, a half-life of 9.86 hours was predicted. This means that about 10 hours after absorption, only half of the THC remains in the blood. It doesn't seem to stay in the body for very long.
As I mentioned before, toxicity is all about concentration. Some predictive models simply categorize something as toxic or not toxic. In this case, I don't have sufficient information to know at what concentrations toxicity was actually observed. This criteria is essential to interpret the prediction value. The source of the data is mentioned in QMRF, but there is no indication of the criteria used to categorize substances in the data as toxic or not toxic. So I can't say anything about the predictions. Determining toxicity is not a simple matter, but a combination of information. I think it would be more effective if someone with expertise in drugs used this program. I may predict few more compounds and report their result like this post.
'AI & Chemistry' 카테고리의 다른 글
| Toxicity prediction services making money without deeplearning (0) | 2024.11.27 |
|---|---|
| AlphaFold explained in the simplest way (1) | 2024.11.25 |
| AI won Nobel prize in Chemistry? well.. I don't use it much. (0) | 2024.11.23 |
| Malicious AI use cases? (0) | 2024.11.22 |
| AI doesn't know what it's doing (0) | 2024.11.21 |