Accounts & Settings
Could you please tell me if this service is completely free?
Researchers are completely free of charge, only companies that choose to sponsor the platform or use the tool for internal R&D will have to pay, in that case, contact us.
I lost my password. What should I do?
If you lost your credentials, please go to the log in page, click on “Forgot password?” and follow the procedure.
Can I test it without an academic email?
Yes, you can but you will have only 3 recommendations at your disposal.
Does the number of recommendations have a limit?
This is mainly a research platform, in function of your account, there is a limit:
You are limited to 3 recommendations per demo account (no academic email required).
In the basic account (registration with academic email), you have 10 recommendations per day and 10 recommendations per month.
If you have a full access account (registration plus consent to the processing of personal data), you have 30 recommendations per day and 100 recommendations per month.
If you need more please contact email@example.com
I need more reccomendation for a specific project, or I need to process many input sets, what can I do?
If you are planning to use Generecommender intensively and you may need more reccomendations, please, contact us, we can find the best way to support you.
Would it be possible to save recommendations for later inspection?
The answer is yes. You can find them under the Main tab in “Your saved searches”. Additionally, previous searches can be edited.
How to interpret the recommendations?
In the n-dimensional representation of human genes performed by the neural network, distance is a measure of similarity. The recommendations are the genes nearest to the point in space that summarize all the characteristics of the inputs.
Is there a limit to the number of genes in input?
There is no limit to the number of genes that can be input. As a general rule of thumb, insert more than 5 inputs to avoid too general recommendation.
Is it possible to export the output?
It is possible to export a detailed report in pptx with all the details, or a xlsx with the predicted gene list.
The platform contains links to commercial products. Why is that?
Generecommender was created by TheProphetAI, who does not produce or sell biological products. Sponsors of your country provide the link you receive.
Obviously, you don’t have to buy the eventual products from them… just be aware that they paid for the service you are receiving for free.
What is the purpose of having multiple algorithms?
For new searches, you can choose between three different algorithms, DeepProphet2, a deep neural network for biomedical, is the most advanced and performs better, but we’ve kept the old ones available for continuity.
What makes the recommendation process so quick?
During training, all the parameters are tuned to reach optimal results through a trial and error process, which can take hours or days in some cases. Having trained the algorithm, the recommendation process can be extremely fast, like a series of simple mathematical operations. Genomic research is a slow process, the speed of AI can be of great value.
GeneRecommender accepts sponsors. How does that work?
It’s possible to have a sponsor per country. Therefore, if you are interested in an area, please let us know, and we could plan a meeting to explain in detail the benefits.
Enrichment of pathways and diseases: how does it work?
The enrichment is performed considering, for every pathway/disease, the probability (p-value) that genes in a given set are annotated to that particular pathway/disease. The p-value is then adjusted with the correction for multiple testing.
Artificial Intelligence: What is it?
A branch of computer science, artificial intelligence is concerned with the development of software capable of simulating human intelligence. As with human brains, intelligence refers to their ability to learn from mistakes and correct themselves.
Deep Learning: What is it?
Deep learning is a subset of Artificial Intelligence that develops machine learning algorithms that have multiple layers, hence the name. As each layer learns a specific characteristic of the data, the software can perform harder tasks.
Does the network have the capability of discovering new gene relationships?
There is no doubt that it can. By scanning all scientific papers, the platform could uncover hidden knowledge already present in literature but ignored by humans, and with this broader perspective, new connections could be made. Artificial intelligence in the life science can boost knowledge to new levels and help the research. Do you want to study the association of new genes to a particular pathology? Give it a try!
In order to train a machine to perform this task, what steps should you take?
In machine learning, algorithms are mathematical operations carried out by using a variety of parameters. In the training process, we run an algorithm on a portion of data that has already been labeled with the characteristic we want the machine to learn, and then compare the output with the true labels. The parameters of the algorithm are tweaked in order to reduce a metric that represents the global error of the machine.
Which data is used for training the algorithm?
In order to train the algorithm, all the freely available scientific papers in Pubmed (about 7 million) are processed and used.
Is it possible to identify why two genes appear similar to the neural network?
Unfortunately, the answer is no. The neural network “thinks” in a 64-dimensional space. Each one of them conveys a part of the information that makes it possible for the algorithm to represent and distinguish genes. Unfortunately this is not human readable. What can be done is to trust the suggestion and start finding information about the recommended genes. Most probably, you will find interesting insights regarding the genes you are not expecting to appear.
In the neural network, which specie's genes are considered?
Currently, only human genes are considered by the neural network, but other species may be added in the future. Please let us know if you have any requests at firstname.lastname@example.org.
Did you ever test the algorithm's performance?
They were tested using gene sets for pathways and diseases. Each time a gene was left out and the remaining input was given, we tested if the algorithm could predict, in its first recommendations, which gene was missing. During training, the algorithm does not know the gene sets, so if it can complete them, then it has learned an accurate representation of genes. The performance in this kind of analysis was astonishingly good, reaching an AUC of 0.9, read our paper for more information. [DOI]