Savir Lab
Harnessing AI for biomedical applications
Multimodal data integration for decision support systems in medicine
Advances in computational pathology and radiology together with developments in omics technologies enable high-throughput generation of complex personalized health data that can provide better treatment prediction. However, practical and reliable methods that integrate multi-omics and image-based data are still lacking.
We develop an artificial intelligence approach that will learn the relation between omics and image-based data and utilize it to provide better, personalized, diagnosis, prognosis, and treatment.
AI representing chemical compounds for drug design
One of the promising future trajectories is harnessing the power of machine learning to screen for billions of compounds in silico. The main advantage of this approach is that it requires a reasonable experimental effort to produce a training set based only on the order of thousands of compounds.
One of the key challenges in applying machine learning for drug discovery is the representation of the molecule. Such a representation must account for the chemical features of the molecule and its 3D structure. We develop graph neural network approaches to represent compounds and utilize them for drug discovery for various conditions.
Generating synthetics images of tissues
One of the current challenges in taking AI to the next level is not just the amount of data but rather the distribution of scenarios with the data – it is not about the numbers it’s about the frequencies. Many medical images (e.g. radiology, pathology, MRI) are biased. These biased could be biological (many healthy patients and many very sick but not enough around the decision threshold) or Technical (images from a particular device or a specific campus).
There is a dual challenge in producing proper synthetic images. The first is the need to control the scene of images. The second is the images should be photorealistic. Many simulators can provide the first but not the second and vice-versa for typical GAN-based solutions.
We develop tools to generate synthetic images of tissues that are photorealistic and allow control over their spatial distribution. This allows us to generate balanced datasets that can be used to facilitate AI development.