Protein kinases control important cell functions, including phosphorylation which plays a role in the activation of proteins. If cancer cells are phosphorylated, there is a risk of the cancer spreading. Some molecules are kinase inhibitors, which can slow the spread of cancer by blocking kinases from phosphorylating proteins. Many cancer treatments rely on the use of kinase inhibitors to slow the spread of cancer. In particular, cytostatic chemotherapy is a targeted cancer treatment which relies on kinase inhibitors. Scientists are constantly testing new drugs and molecules to determine which may be most effective for such treatments.

Mihir Putcha is a high school student with an interest in biology and machine learning. He conducted a research project with the Jinso research advising program to develop machine learning models to classify molecules as either kinase inhibitors or non-inhibitors.

Mihir used a dataset involving thousands of molecules whose effects were analyzed on 8 protein kinases. He applied several machine learning models to the dataset to determine which model is most effective. The models included fully connected neural networks and convolutional neural networks.

To train his models, Mihir divided his data into a testing and training dataset using a standard 80/20 train-test split. He compared the results of his test data across several metrics. Mihir recorded an accuracy for each model, which is the percentage of all predictions it makes about whether molecules are kinase inhibitors that are correct. He also calculated recall, which is the percentage of actual positives that were correctly predicted.

While one of his fully connected neural networks performed best in terms of recall (.9603), Mihir could that one of the convolutional neural networks had the best results for all other metrics, including accuracy (.863). Although the accuracy of the model increased with more training data, the accuracy seemed to plateau around .9, suggesting a limitation on the machine learning model to analyze such a dataset. However, models such of these can still be extremely useful for scientific research and drug discovery with human oversight.

By the application of such machine learning classifiers, molecules can be marked as having potential for cancer treatments before significant experimental research. The models are able to detect characteristics of the molecules which may make them more or less effective kinase inhibitors than would be detectable to a human.

Scientists are very optimistic about cytostatic chemotherapy for treating cancer. Its focus on preventing the spread of cancer cells through inhibition is generally more effective than methods which attempt to kill tumor cells, such as cytotoxic chemotherapy.

The project built on prior work by researchers at Stanford University who trained deep learning models to analyze the same dataset. Mihir's models outperformed more sophisticated algorithms for some metrics.

Mihir is currently a junior at Adlai E. Stevenson High School in Lincolnshire, Illinois. He shared his preprint online via OSF Preprints, which is titled "Machine Learning Models to Classify Molecules as Cancer Kinase Inhibitors or Non Inhibitors". He published his paper in February 2022.

GitHub is a popular platform used by computer scientists to manage their collaborative projects, but a similar program does not exist for academic work. There is no standard platform to create work, connect with others, and share work in one place. Most platforms only fall into one or two of these categories.The Jinso collaboration tool is a better way for groups to work on projects. By bringing the entire academic collaboration process onto one tool, it simplifies workflows and communication.The first steps for using the Jinso platform are:

Create an account
Create your first group

Once a user builds a network, they can create new Groups that consist of their network members. By default, the creator of a group is the admin. The most common Group is a research group, but the platform can manage several other types of academic projects. Platform users can create study groups for sharing course materials or groups of club members for extracurricular work.The admin of the Group has the ability to add new members at any time.
Admins are also responsible for creating Projects within Groups.

A Project for a research group is usually a research paper, but Projects can also be other forms of documents that could benefit from discussion and revisions. Examples include study guides, business plans, articles, and essays. Each Group can have an unlimited number of Projects within it, and all Projects within a Group are shared among the same members. 

Once a user builds a network, they can create new Groups that consist of their network members. By default, the creator of a group is the admin. The most common Group is a research group, but the platform can manage several other types of academic projects.

Platform users can create study groups for sharing course materials or groups of club members for extracurricular work.The admin of the Group has the ability to add new members at any time. Admins are also responsible for creating Projects within Groups.

A Project for a research group is usually a research paper, but Projects can also be other forms of documents that could benefit from discussion and revisions. Examples include study guides, business plans, articles, and essays. Each Group can have an unlimited number of Projects within it, and all Projects within a Group are shared among the same members. 

Example of Research group
Revisions of the paper

When a new Project is created, an initial revision must be shared. This can either be plain text or a PDF.
The Project will be immediately visible to all Group members with the first revision shown. Group members can comment on the revision with questions or feedback, and others can reply to comments.When another revision of the paper has been completed, the Group admin can add a new revision to the same Project.
The revision will become visible above the prior revision, and it will have a new comment box associated with it. Projects make it simple to keep track of a paper’s entire revision history and discussions at each stage. 

For each revision, Group admins can also create subtasks. Arrows allow Group members to view all of the different subtasks and comment on them individually. Subtasks allow a paper to be analyzed in unique components. For example, a research paper can have a unique subtask for each of its sections, and collaborators can discuss them all separately in the comment boxes. Jinso is a quicker way to collaborate on long-term projects. It makes it easier to connect, share, and manage the development of ideas and papers. You can create a Jinso account and start using the platform today for your research and academic needs at jinso.io.

Protein kinases control important cell functions, including phosphorylation which plays a role in the activation of proteins. If cancer cells are phosphorylated, there is a risk of the cancer spreading. Some molecules are kinase inhibitors, which can slow the spread of cancer by blocking kinases from phosphorylating proteins. Many cancer treatments rely on the use of kinase inhibitors to slow the spread of cancer. In particular, cytostatic chemotherapy is a targeted cancer treatment which relies on kinase inhibitors. Scientists are constantly testing new drugs and molecules to determine which may be most effective for such treatments.

Mihir Putcha is a high school student with an interest in biology and machine learning. He conducted a research project with the Jinso research advising program to develop machine learning models to classify molecules as either kinase inhibitors or non-inhibitors.

Mihir used a dataset involving thousands of molecules whose effects were analyzed on 8 protein kinases. He applied several machine learning models to the dataset to determine which model is most effective. The models included fully connected neural networks and convolutional neural networks.

To train his models, Mihir divided his data into a testing and training dataset using a standard 80/20 train-test split. He compared the results of his test data across several metrics. Mihir recorded an accuracy for each model, which is the percentage of all predictions it makes about whether molecules are kinase inhibitors that are correct. He also calculated recall, which is the percentage of actual positives that were correctly predicted.

While one of his fully connected neural networks performed best in terms of recall (.9603), Mihir could that one of the convolutional neural networks had the best results for all other metrics, including accuracy (.863). Although the accuracy of the model increased with more training data, the accuracy seemed to plateau around .9, suggesting a limitation on the machine learning model to analyze such a dataset. However, models such of these can still be extremely useful for scientific research and drug discovery with human oversight.

By the application of such machine learning classifiers, molecules can be marked as having potential for cancer treatments before significant experimental research. The models are able to detect characteristics of the molecules which may make them more or less effective kinase inhibitors than would be detectable to a human.

Scientists are very optimistic about cytostatic chemotherapy for treating cancer. Its focus on preventing the spread of cancer cells through inhibition is generally more effective than methods which attempt to kill tumor cells, such as cytotoxic chemotherapy.

The project built on prior work by researchers at Stanford University who trained deep learning models to analyze the same dataset. Mihir's models outperformed more sophisticated algorithms for some metrics.

Mihir is currently a junior at Adlai E. Stevenson High School in Lincolnshire, Illinois. He shared his preprint online via OSF Preprints, which is titled "Machine Learning Models to Classify Molecules as Cancer Kinase Inhibitors or Non Inhibitors". He published his paper in February 2022.

All posts