Innovation in life sciences is advancing at possibly the fastest pace among all branches of science. The same can be said for deep learning – with the relentless advancement in computational power coupled with the continuous improvement in learning algorithms. The confluence of these two disciplines has the potential to change the world in a dramatic way. Today we are living in a golden age of these technologies, a renaissance where the synergy between deep learning and biology will provide the catalyst for historic advancements in life sciences. The effects of this union have already had a substantial impact in biological research, medical diagnosis, and precision medicine – and it's only the beginning of the realization of the potential.
Deep learning has great potential to transform basic research in biology. Particularly in cases where research teams produce enormous volumes of data – so much that it is difficult to analyze all of it. One such domain is deep learning in drug discovery. Over the past decade, there has been a remarkable increase in the amount of available compound activity and biomedical data owing to the emergence of new experimental techniques such as high throughput screening, parallel synthesis, among others Deep learning algorithms are used to predict likely interactions between complex proteins and potential drug molecules (1,2). The introduction of a representation of complex protein molecules into grids of 3D pixels, called voxels, has allowed data scientists to train deep learning models directly from the 3D structure of proteins and small molecules with atomic precision. The latent representations learned from training the deep learning models from these features can be used to predict drug molecules are likely to interact with a given protein.
Another area of research that has been aided by recent developments in deep learning is in microscopy. For example, Steve Finkbeiner’s team at the Gladstone Institute of Neurological Disease. His team is leveraging robotic microscopy to study brain cells, and generating large image datasets that were a challenge to analyze with traditional machine learning techniques (3,4).
His efforts in collaborating with Google in applying deep learning to this data paid off in astonishing good predictions of labeled robotic imaging data, coupled with the ability to analyze all of the data his team was generating.
Another example is the research conducted by Alexandra Maslova et al in applying deep learning to immunology (5).. In that study she trained a convolutional neural network model to determine cell type-specific open chromatin regions. Their model was able to learn the sequence syntax that underlies the globality of immune cell differentiation.
There are several drivers that are fueling the advancement of medical diagnostics through deep learning technology including:
Geoffrey Hinton, whom many consider the Godfather of Deep Learning, famously predicted that radiologists would soon lose their jobs (6). “People should stop training radiologists right now,” he announced, “It’s obvious that within five years deep learning is going to do better than humans.” One of the reasons Dr Hinton said this is that diagnostic imaging is a natural fit for deep learning algorithms that excel at learning from labeled and unlabeled image data. It has been six years since he made that claim and while deep learning has been widely used in disease detection, the employment of radiologists has not been impacted in the way he suggested.
One case of this accuracy improvement was recently published in The Lancet by a Korean team of researchers (7). In their research on breast cancer detection they showed that a deep learning model alone had 88.8% sensitivity, whereas radiologists alone had 75.3% accuracy. When radiologists were aided by the deep learning model, the accuracy increased by 9.5% to 84.8%.
In just the last few years, models have been published that match or exceed the accuracy of expert humans at diagnosing many important diseases including: pneumonia , skin cancer , diabetic retinopathy , age-related macular degeneration, heart arrhythmia – just to name a few (8,9,10,11,12). In diagnostic medicine, trained deep learning models clearly have had a profound effect in accuracy improvement. However, even with all of the advancements provided by these models medical professionals will not be replaced anytime soon. The models are meant to support and augment the diagnoses of the medical professional, providing them with additional data to support their decisions for care. In the coming months and years, we are likely to see a dramatic increase in the standard of care doctors provide when aided by these models.
The next step beyond diagnosing an illness is deciding how to treat it. Precision medicine is an emerging approach to patient care that focuses on understanding and treating disease by integrating data from an individual to make patient-tailored decisions. It takes into account every patient’s unique genetics and biochemistry to select the best treatment for that individual, providing the greatest benefit with the fewest side effects. In order to achieve this lofty goal, models must be trained on terabytes of genomic, image, and other forms of data – deep learning excels at this kind of challenge.
The evolution of precision medicine has been heavily influenced by the exponential increase in the amount of biologic data that can now be collected for each individual patient, in large part due to the advent of new technologies in the fields of medicine, genetics, metabolics, and imaging, among others. Using data generated from previous patients treated for a disease, deep learning models can identify future patients who may benefit from a specific treatment. Deep learning models can also be trained on data acquired through electronic medical records to identify patients with conditions that may benefit from participation in novel treatment studies (13).
A recent study by researchers at the University of Shanghaidemonstrated that response to cancer treatments could be predicted based on genomics (14). If the methodology in research such as theirs is clinically applied it would help avoid unnecessary treatments for patients that are unresponsive to certain drugs in favor of the most effective treatment based on the patient’s genome. Deep learning models can even take this one step further by helping design drugs that are tailored for the biological response of each individual. These deep learning models have helped pave the way for pharmacogenomics – a new research field – which helps inform personalized drug design (15).
There are a multitude of areas where deep learning and biological research, medical diagnostics and precision medicine have joined forces to make significant advancements, in this insights piece, we have only highlighted a few areas that demonstrate the breadth of the problem domains addressable by deep learning. We are at a remarkable moment in history when a set of new technologies is coming together to change the world.
Subscribe to our newsletter and get updates on emerging computational life science trends straight in your inbox.
The integration of AI and multiomics data insight generation is accelerating gene therapy development. Learn more.
DeepMind researchers released a large language model (LLM) that uses fewer parameters yet achieves better performance. Read more to understand how this is relevant to your gene therapy programs.