Cancer cell: Scientists have developed an artificial intelligence system that allows AI to match the best combination of drugs for cancer patients

For cancer therapies, dozens of new treatments enter clinical trials each year, but less than 4% of the drugs are eventually approved by the FDA.
although there are many factors that influence this outcome, the main problem is that we don't fully understand how or why a particular cancer responds to treatment.
, it is not currently possible to optimally combine the right drugs and match them to the right patients.
, however, the ad emerging of artificial intelligence may help us.
most machine learning models are "black boxes" that can be optimized for predictive accuracy without having to understand or pay attention to the biological mechanisms of predictions.
recently, researchers at the University of California, San Diego School of Medicine said they had created a new artificial intelligence (AI) system called DrugCell that makes it possible for tumors to match the best combination of drugs.
DrugCell, the system returns the best-known drugs, the biological pathways that control the response to the drug, and the best combination of drugs after entering data about the tumor.
study was published in Cancer Cell.
https://doi.org/10.1016/j.ccell.2020.09.014 Previously, researchers developed a visible neural network (VNN) that simulates a simple nuclear cell, brewing yeast.
the system can accurately predict the effects of genetic mutations on cell growth responses, while identifying the most relevant molecular pathways driving these predictions.
, they created a VNN called DrugCell that simulates the response of human cancer cells to therapeutic compounds.
DrugCell combines the internal workings of the model with the hierarchy of human cell biology to predict the response of any drug in any cancer and to design effective combination therapies.
drug reaction is a complex phenomenon, depending on biological and chemical factors.
In order to capture the two determinants of drug reactions in explanatory models, the researchers designed DrugCell as a neural network with two branches, the first being VNN, modeled from 2,086 biological processes recorded in the Human Gene (GO) database to simulate the hierarchy of molecular subsyscies in human cells.
Each of these subsyscies, from subsyscies involving small protein complexes (e.g. b-serial protein destruction complexes) to larger signaling pathways (e.g. MAPK signaling pathways) to overall cellular functions (e.g. glycolysis), is assigned a set of artificial neurons to represent the state of the subsyscies.
VNN used a total of 12,516 neurons, which were layered over six different layers.
a second branch of DrugCell's design is the traditional artificial neural network (ANN), which embeds the drug's Morgan fingerprint, the standard vector representation of the chemical structure.
the output of the two branches of the model (VNN embedded in the cell genotype and ANN embedded in the drug structure) is merged into a single layer of neurons and then integrated to produce a given genotype response to a particular treatment.
, the model trained 684 drugs by responding to 1,235 tumor cell line.
results show that DrugCell is able to accurately predict the response of cell line to treatment (the total accuracy of all cell line-drug pairs is: spearman correlation coefficient: 0.80).
, the predicted combination improved the non-progressive survival of patient-sourced heterogeneity transplant tumor models and layered clinical outcomes in ER-positive breast cancer patients. Of all the (cell line, drug) pairs studied in
, DrugCell, who predicted and actually drug responses originated from patients with heterogeneity transplant tumors, received training in the response of more than 1,200 tumor cell line to nearly 700 FDA-approved drugs and experimental therapeutic drugs, with a total of more than 500,000 cell line/drug pairs.
first author Kuenzi said: "We were surprised by DrugCell's ability to transform from laboratory cell line to tumors in mice and patients, as well as clinical trial data.
but our ultimate goal is to get DrugCell into the clinic for the benefit of the patient.
, much remains to be done.
also stressed that while the 1,200 cell line is a good start, it does not represent the complete heterogeneity of the cancer.
team is now adding more single-cell data and experimenting with different drug structures.
they also want to work with existing clinical studies to embed DrugCell in diagnostic tools and to proactively test them in real life.