User login
MILAN – , German researchers reported.
The team fed the medication plans of almost 630 patients into a deep neural network, which identified drug-drug interactions in more than 80% of cases, in particular when switching from one medication to another, alongside potential food interactions.
The tool was able to identify specific interactions that could be avoided if a drug was replaced with one with a similar pharmacologic profile, but a lower risk of adverse effects.
“Potential drug-drug interactions are a major safety concern in patients with MS,” said study presenter Michael Hecker, PhD, department of neurology, Rostock (Germany) University Medical Center.
Such deep learning–based methods are “useful” in screening for potential interactions both between drugs and with foods, they concluded.
The findings were presented at the 9th Joint ECTRIMS-ACTRIMS Meeting.
Unknown interactions
During his presentation, Dr. Hecker noted that most patients with MS take two or more drugs “to treat their disease and to mitigate their symptoms and comorbidities.” He pointed out, however, that patients who take multiple medications are at an increased risk for side effects, as one drug may affect the pharmacokinetic or pharmacodynamic properties of another.
“For instance, it may change its metabolism,” Dr. Hecker said, and therefore affect its mechanism of action and the response to the drug, with medications potentially having synergistic, antagonistic, or additive effects.
He explained that the online DrugBank database “provides a huge collection” of known drug-drug interactions for compounds that have a track record. “However, for other drugs, and especially those that are tested only in clinical trials, there’s no information about drug-drug and drug-food interactions,” Dr. Hecker said.
“Moreover, it is quite time-consuming to search a database for individual drug-drug interactions,” he added.
34 million parameters
Consequently, there is increasing interest in the use of deep neural networks to study drug-drug interactions, Dr. Hecker said. DeepDDI is the “state-of-the-art deep learning framework” for predicting interactions. It takes drug-drug or drug-food pairs and compares their structures to determine their similarity. This information is fed into a deep neural network with almost 34 million trainable parameters.
The framework then provides a prediction of any interactions in the same terms as the DrugBank, suggesting, for example, that Drug A may decrease the antihypertensive activities of Drug B.
For the current study, the researchers trained the deep neural network on the most recent release of the DrugBank database, finding it was able to replicate the drug-drug interactions in the database at an accuracy of 92.2% in the validation set and 92.1% in the testing set. They then put the medication plans of 627 patients with MS into the deep neural network.
The patients had a mean age of 48.6 years, 70.3% were women, and the median disease duration was 10 years. They were taking an average of 5.3 medications, and 62% were using disease-modifying therapies (DMT).
The team compared the structures of the drugs they were taking with those of 367 drugs used for the treatment of MS, as well as with structural data for 1,673 food compounds from the FooDB database.
Swapping drugs could reduce interactions
The overall prevalence of potential drug-drug interactions among the patients included in the study was 81.2%.
The researchers then determined the proportion of patients who would be at risk of additional drug-drug interaction if they switched from one DMT to another, or to a Bruton tyrosine kinase inhibitor, given all their other medications.
They found, for example, that more than 40% of patients who switched to the immunomodulator fingolimod (Gilenya) would be at increased risk for bradycardia.
Just under 40% of patients who changed their DMT to the purine analogue cladribine (Mavenclad) would have an increased risk, or worsening, of bleeding, as would approximately 25% of those who switched to the anthracenedione antineoplastic agent mitoxantrone (Novantrone).
Dr. Hecker also showed the deep neural network could make suggestions as to how critical drug-drug interactions could be avoided by replacing interacting drugs with alternatives that have similar pharmacological effects.
For example, carbamazepine (Tegretol, Equetro) could be replaced with topiramate (several brand names) to avoid hepatotoxicity in patients also taking acetaminophen, while liothyronine (Cytomel, Triostat) could replace levothyroxine in patients also taking teriflunomide (Aubagio).
Finally, Dr. Hecker reported there was a subset of 6,860 potential drug-food interactions with the patients’ medications, resulting in reduced or increased concentrations of the drugs, particularly with fish or mushroom consumption.
He conceded, however, there were several limitations to their study, including that it included only small-molecule drugs, and that they did not ask patients about their diet or if they had observed any undesired drug effects.
Furthermore, “only a small number” of the potential drug-drug or drug-food interactions they identified would be “clinically relevant.”
Dr. Hecker also pointed out that each drug has one record, but it is used for different indications, with different dosages, and has different side effects, depending how it is used. “The model does not distinguish this,” he said, and so some of the interactions it highlights could be related to other doses than the one used in MS, for example.
Promise for the future
Pavan Bhargava, MBBS, MD, associate professor of neurology, Johns Hopkins Precision Medicine Center of Excellence for Multiple Sclerosis, Baltimore, commented that, as with all AI tools, “it’s only as good as what we’re putting into it.”
Dr. Bhargava, who cochaired the session, said that “there’s limitations on the information in the databases” that are being fed into the deep neural network.
He also highlighted that, “at this point, it didn’t seem like it was coming up with much clinically useful information,” but noted that, “we may get to that point.”
“Right now, there’s promise,” Dr. Bhargava said, but “it’s still not quite there.”
No funding was declared. Dr. Hecker declares relationships with Bayer HealthCare, Biogen, Merck Healthcare, Novartis, and Teva. Several other coauthors also declared financial relationships with industry.
A version of this article first appeared on Medscape.com.
MILAN – , German researchers reported.
The team fed the medication plans of almost 630 patients into a deep neural network, which identified drug-drug interactions in more than 80% of cases, in particular when switching from one medication to another, alongside potential food interactions.
The tool was able to identify specific interactions that could be avoided if a drug was replaced with one with a similar pharmacologic profile, but a lower risk of adverse effects.
“Potential drug-drug interactions are a major safety concern in patients with MS,” said study presenter Michael Hecker, PhD, department of neurology, Rostock (Germany) University Medical Center.
Such deep learning–based methods are “useful” in screening for potential interactions both between drugs and with foods, they concluded.
The findings were presented at the 9th Joint ECTRIMS-ACTRIMS Meeting.
Unknown interactions
During his presentation, Dr. Hecker noted that most patients with MS take two or more drugs “to treat their disease and to mitigate their symptoms and comorbidities.” He pointed out, however, that patients who take multiple medications are at an increased risk for side effects, as one drug may affect the pharmacokinetic or pharmacodynamic properties of another.
“For instance, it may change its metabolism,” Dr. Hecker said, and therefore affect its mechanism of action and the response to the drug, with medications potentially having synergistic, antagonistic, or additive effects.
He explained that the online DrugBank database “provides a huge collection” of known drug-drug interactions for compounds that have a track record. “However, for other drugs, and especially those that are tested only in clinical trials, there’s no information about drug-drug and drug-food interactions,” Dr. Hecker said.
“Moreover, it is quite time-consuming to search a database for individual drug-drug interactions,” he added.
34 million parameters
Consequently, there is increasing interest in the use of deep neural networks to study drug-drug interactions, Dr. Hecker said. DeepDDI is the “state-of-the-art deep learning framework” for predicting interactions. It takes drug-drug or drug-food pairs and compares their structures to determine their similarity. This information is fed into a deep neural network with almost 34 million trainable parameters.
The framework then provides a prediction of any interactions in the same terms as the DrugBank, suggesting, for example, that Drug A may decrease the antihypertensive activities of Drug B.
For the current study, the researchers trained the deep neural network on the most recent release of the DrugBank database, finding it was able to replicate the drug-drug interactions in the database at an accuracy of 92.2% in the validation set and 92.1% in the testing set. They then put the medication plans of 627 patients with MS into the deep neural network.
The patients had a mean age of 48.6 years, 70.3% were women, and the median disease duration was 10 years. They were taking an average of 5.3 medications, and 62% were using disease-modifying therapies (DMT).
The team compared the structures of the drugs they were taking with those of 367 drugs used for the treatment of MS, as well as with structural data for 1,673 food compounds from the FooDB database.
Swapping drugs could reduce interactions
The overall prevalence of potential drug-drug interactions among the patients included in the study was 81.2%.
The researchers then determined the proportion of patients who would be at risk of additional drug-drug interaction if they switched from one DMT to another, or to a Bruton tyrosine kinase inhibitor, given all their other medications.
They found, for example, that more than 40% of patients who switched to the immunomodulator fingolimod (Gilenya) would be at increased risk for bradycardia.
Just under 40% of patients who changed their DMT to the purine analogue cladribine (Mavenclad) would have an increased risk, or worsening, of bleeding, as would approximately 25% of those who switched to the anthracenedione antineoplastic agent mitoxantrone (Novantrone).
Dr. Hecker also showed the deep neural network could make suggestions as to how critical drug-drug interactions could be avoided by replacing interacting drugs with alternatives that have similar pharmacological effects.
For example, carbamazepine (Tegretol, Equetro) could be replaced with topiramate (several brand names) to avoid hepatotoxicity in patients also taking acetaminophen, while liothyronine (Cytomel, Triostat) could replace levothyroxine in patients also taking teriflunomide (Aubagio).
Finally, Dr. Hecker reported there was a subset of 6,860 potential drug-food interactions with the patients’ medications, resulting in reduced or increased concentrations of the drugs, particularly with fish or mushroom consumption.
He conceded, however, there were several limitations to their study, including that it included only small-molecule drugs, and that they did not ask patients about their diet or if they had observed any undesired drug effects.
Furthermore, “only a small number” of the potential drug-drug or drug-food interactions they identified would be “clinically relevant.”
Dr. Hecker also pointed out that each drug has one record, but it is used for different indications, with different dosages, and has different side effects, depending how it is used. “The model does not distinguish this,” he said, and so some of the interactions it highlights could be related to other doses than the one used in MS, for example.
Promise for the future
Pavan Bhargava, MBBS, MD, associate professor of neurology, Johns Hopkins Precision Medicine Center of Excellence for Multiple Sclerosis, Baltimore, commented that, as with all AI tools, “it’s only as good as what we’re putting into it.”
Dr. Bhargava, who cochaired the session, said that “there’s limitations on the information in the databases” that are being fed into the deep neural network.
He also highlighted that, “at this point, it didn’t seem like it was coming up with much clinically useful information,” but noted that, “we may get to that point.”
“Right now, there’s promise,” Dr. Bhargava said, but “it’s still not quite there.”
No funding was declared. Dr. Hecker declares relationships with Bayer HealthCare, Biogen, Merck Healthcare, Novartis, and Teva. Several other coauthors also declared financial relationships with industry.
A version of this article first appeared on Medscape.com.
MILAN – , German researchers reported.
The team fed the medication plans of almost 630 patients into a deep neural network, which identified drug-drug interactions in more than 80% of cases, in particular when switching from one medication to another, alongside potential food interactions.
The tool was able to identify specific interactions that could be avoided if a drug was replaced with one with a similar pharmacologic profile, but a lower risk of adverse effects.
“Potential drug-drug interactions are a major safety concern in patients with MS,” said study presenter Michael Hecker, PhD, department of neurology, Rostock (Germany) University Medical Center.
Such deep learning–based methods are “useful” in screening for potential interactions both between drugs and with foods, they concluded.
The findings were presented at the 9th Joint ECTRIMS-ACTRIMS Meeting.
Unknown interactions
During his presentation, Dr. Hecker noted that most patients with MS take two or more drugs “to treat their disease and to mitigate their symptoms and comorbidities.” He pointed out, however, that patients who take multiple medications are at an increased risk for side effects, as one drug may affect the pharmacokinetic or pharmacodynamic properties of another.
“For instance, it may change its metabolism,” Dr. Hecker said, and therefore affect its mechanism of action and the response to the drug, with medications potentially having synergistic, antagonistic, or additive effects.
He explained that the online DrugBank database “provides a huge collection” of known drug-drug interactions for compounds that have a track record. “However, for other drugs, and especially those that are tested only in clinical trials, there’s no information about drug-drug and drug-food interactions,” Dr. Hecker said.
“Moreover, it is quite time-consuming to search a database for individual drug-drug interactions,” he added.
34 million parameters
Consequently, there is increasing interest in the use of deep neural networks to study drug-drug interactions, Dr. Hecker said. DeepDDI is the “state-of-the-art deep learning framework” for predicting interactions. It takes drug-drug or drug-food pairs and compares their structures to determine their similarity. This information is fed into a deep neural network with almost 34 million trainable parameters.
The framework then provides a prediction of any interactions in the same terms as the DrugBank, suggesting, for example, that Drug A may decrease the antihypertensive activities of Drug B.
For the current study, the researchers trained the deep neural network on the most recent release of the DrugBank database, finding it was able to replicate the drug-drug interactions in the database at an accuracy of 92.2% in the validation set and 92.1% in the testing set. They then put the medication plans of 627 patients with MS into the deep neural network.
The patients had a mean age of 48.6 years, 70.3% were women, and the median disease duration was 10 years. They were taking an average of 5.3 medications, and 62% were using disease-modifying therapies (DMT).
The team compared the structures of the drugs they were taking with those of 367 drugs used for the treatment of MS, as well as with structural data for 1,673 food compounds from the FooDB database.
Swapping drugs could reduce interactions
The overall prevalence of potential drug-drug interactions among the patients included in the study was 81.2%.
The researchers then determined the proportion of patients who would be at risk of additional drug-drug interaction if they switched from one DMT to another, or to a Bruton tyrosine kinase inhibitor, given all their other medications.
They found, for example, that more than 40% of patients who switched to the immunomodulator fingolimod (Gilenya) would be at increased risk for bradycardia.
Just under 40% of patients who changed their DMT to the purine analogue cladribine (Mavenclad) would have an increased risk, or worsening, of bleeding, as would approximately 25% of those who switched to the anthracenedione antineoplastic agent mitoxantrone (Novantrone).
Dr. Hecker also showed the deep neural network could make suggestions as to how critical drug-drug interactions could be avoided by replacing interacting drugs with alternatives that have similar pharmacological effects.
For example, carbamazepine (Tegretol, Equetro) could be replaced with topiramate (several brand names) to avoid hepatotoxicity in patients also taking acetaminophen, while liothyronine (Cytomel, Triostat) could replace levothyroxine in patients also taking teriflunomide (Aubagio).
Finally, Dr. Hecker reported there was a subset of 6,860 potential drug-food interactions with the patients’ medications, resulting in reduced or increased concentrations of the drugs, particularly with fish or mushroom consumption.
He conceded, however, there were several limitations to their study, including that it included only small-molecule drugs, and that they did not ask patients about their diet or if they had observed any undesired drug effects.
Furthermore, “only a small number” of the potential drug-drug or drug-food interactions they identified would be “clinically relevant.”
Dr. Hecker also pointed out that each drug has one record, but it is used for different indications, with different dosages, and has different side effects, depending how it is used. “The model does not distinguish this,” he said, and so some of the interactions it highlights could be related to other doses than the one used in MS, for example.
Promise for the future
Pavan Bhargava, MBBS, MD, associate professor of neurology, Johns Hopkins Precision Medicine Center of Excellence for Multiple Sclerosis, Baltimore, commented that, as with all AI tools, “it’s only as good as what we’re putting into it.”
Dr. Bhargava, who cochaired the session, said that “there’s limitations on the information in the databases” that are being fed into the deep neural network.
He also highlighted that, “at this point, it didn’t seem like it was coming up with much clinically useful information,” but noted that, “we may get to that point.”
“Right now, there’s promise,” Dr. Bhargava said, but “it’s still not quite there.”
No funding was declared. Dr. Hecker declares relationships with Bayer HealthCare, Biogen, Merck Healthcare, Novartis, and Teva. Several other coauthors also declared financial relationships with industry.
A version of this article first appeared on Medscape.com.
FROM ECTRIMS 2023