Institute of AI and Robotics

Decision making for medical diagnosis is a major challenge even for well highly qualified doctors with long experience. Further such decisions are subjective and highly dependent on the health condition, perception and eyesight of the persons who make decisions. Often the medical cases in the borderline can be at the risk of being misdiagnosed which is a major threat for the patients for being operated on instead of cured using chemotherapy or vice versa. So, there is a long going discussion on introducing AI for medical diagnosis to support the decision making in medical diagnosis. The machine learning models are trained with various types of medical data (can be X-ray images, CT scan, MRI, PET images, ECG data) and a well trained AI model can classify the medical condition in a superior way when compared to its human counterpart. Further with the increase in the size of the datasets the reliability of the machine learning model improves significantly. Currently, the machine learning models to classify a medical condition as normal, infected with tuberculosis, Covid-19 or other types of lung infections based on X-ray or CT images have been developed and shown success.

Drug discovery involves design of small molecular compounds which can inhibit certain biomolecules aberrantly expressed in the disease state. The biomolecules expressed aberrantly are referred to as biomarkers as they can be used as indicators to provide first hand information about disease conditions. A potential therapeutic route is to identify small molecules that can inhibit these biomarkers. However, the major challenge here is to identify the compounds from a chemical space having about trillions (ZINC15 database) to 10^60 (GDP17) compounds. Even a screening using a fast scoring function as implemented in Autodock4.0 or Autodock Vina can be time taking and so machine learning approaches can be successfully employed. In the ligand based approach, the datasets of descriptors along with inhibition constants (Ki, Kd, IC50, pIC50) are used for developing a model. Using these models, any chemical spaces can be screened to identify potential lead compounds. One can also employ structure based machine learning models such as OnionNET or AtomDeep however they are comparably expensive as they make use of descriptors based on the intermolecular distances of protein-ligand complexes along with other features.

Materials design suitable for energy production in a cheaper, faster and sustainable way is another major research area where AI can see potential applications. There are many research works ongoing to develop machine learning models to develop suitable dye molecules for dye-sensitized solar cell applications. In these cases, the database of various dye molecules along with various descriptors (1D,2D and 3D) and energy efficiency is used to develop machine learning model. In addition to using descriptors based on empirical or semi-empirical approaches, one can also use QM based descriptors such as energy gap, energies of frontier molecular orbitals and chemical potential to develop more reliable machine learning models. Once the model has been developed they can be used to get insight into the energy efficiency of selected dye molecules and in addition can be used to design novel dye molecules with superior energy efficiency. Such AI based design can be carried out with minimal computational resources and within a limited time frame and so is a major boost for the research sectors working in this area. However, the lead compounds proposed from the machine learning models need to be synthesized. However, the synthesizability can also be optimized using suitable machine learning models.

We will study and analyse your current business model and report on the possible ways to improve it with the use of machine learning approaches. The preliminary level assessment is completely free and your data that will be used for this purpose is completely secure with us. Usually, many companies adopt a static business model that often ignores the dynamical nature of various parameters that potentially affect the business development, costs and profits. Due to this reason, many times there is a wastage of products developed or underproduction of consumables that are in high demand which eventually lead to a not-so profitable business. However with the use of machine learning models we can optimize the various processes in business developments and the loss can be reduced and profits can be increased. Further the use of machine learning models in the business also makes the business sustainable as the wastage is largely reduced.