Identifying a cancer patient’s precise type of cancer and primary location is the first step in selecting an effective treatment. However, even with rigorous testing, the origin of cancer cannot be determined in rare instances. Although these tumours of unclear sources tend to be aggressive, oncologists are required to treat them with non-targeted medicines, which typically result in high toxicity and low survival rates.
With this, researchers at the Koch Institute for Integrative Cancer Research at Massachusetts Institute of Technology (MIT) and Massachusetts General Hospital (MGH) have created a new deep-learning approach that may assist categorise tumours of unknown origin by examining gene expression programmes related to early cell development and differentiation.
Developing a diagnostic tool from a machine learning model that exploits variations between healthy and normal cells and among various types of cancer requires a delicate balancing act.
If a model is very sophisticated and accounts for an excessive number of aspects of cancer gene expression, it may appear to learn the training data flawlessly yet fail when presented with new data. However, by reducing the number of characteristics to simplify the model, the model may fail to capture the types of information that would lead to correct classifications of cancer types.
To achieve a compromise between lowering the number of features and selecting the most pertinent data, the scientists centred the model on cancer cell markers of disrupted developmental pathways. As an embryo develops and undifferentiated cells specialise into diverse organs, a plethora of pathways governs cell division, growth, shape change, and migration.
As the tumour grows, cancer cells lose several specialised characteristics of mature cells. In addition, as they acquire the ability to multiply, change, and metastasise to other tissues, they begin to resemble embryonic cells in some respects. In cancer cells, many of the gene expression pathways that drive embryogenesis are reactivated or dysregulated.
The researchers took the gene expression of tumour samples from the Cancer Genome Atlas (TCGA) and broke it down into separate parts that each correspond to a certain point in a tumour’s development. They then gave each of these parts a mathematical value; and turned it into a machine learning model tag as Developmental Multilayer Perceptron (D-MLP), which gives a tumour a score based on how it grew and then predicts where it came from.
Meanwhile, when DALL-E came out, it made everyone on the internet feel good. DALL-E is an image generator based on artificial intelligence that was inspired by the artist Salvador Dali and the cute robot WALL-E.
It uses natural language to make any mysterious and beautiful image your heart desires. When people typed in things like “smiling gopher holding an ice cream cone,” they saw them come to life right away.
To make an image, DALL-E 2 uses something called a “diffusion model,” which tries to fit all the text into one description. But when there are a lot more details in the text, it’s hard for one description to cover everything.
Even though they are very adaptable, they sometimes have trouble understanding how certain ideas are put together. To make more complex images that are easier to understand, scientists at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) changed the way the typical model is set up.
The “magical” models that are used to make images work by suggesting a series of steps that can be taken over and over to get to the desired image. It starts with a “bad” picture and then makes it better and better until it is the one that is chosen.
By putting together several models, they can refine the look together at each step, making an image that has all the features of each model. By having several models work together, users can choose from a lot more creative image combinations.
