Biology is a vast subject with dozens of subdisciplines. Some of them like computational biology, bioinformatics, and systems biology have pioneered the processing of biological data using machine learning approaches, data analysis methods, large-scale data mining, predictive models, Bayesian models, and computer science algorithms like genetic algorithms.
Let's talk about six cutting-edge uses of machine learning in biology.
Scientists saved millions of lives by creating novel vaccines and drugs in record time. Machine learning played a prominent role in many aspects of the COVID-19 response — vaccine design, drug discovery, and disease modeling.
Scientists had to find the part of the virus that was most likely to stimulate an immune response. The virus contained RNA genetic material surrounded by a sheath of spike proteins. The spike proteins let the virus bind to receptors on our cells. So, most vaccines targeted them. If the vaccine created antibodies that neutralized the spike proteins, the virus could not attack cells.
A Stanford University research team was one of the first to study vulnerable spots on the virus. They used two machine learning models named NetMHCpan and MARIA.
NetMHCpan was a simple feed-forward neural network. It had a single hidden layer that predicted receptor and protein interactions. Its inputs were the spike protein amino acid sequences and cell surface MHC protein sequences. Its outputs were the probability that they'd bind.
MARIA was a recurrent neural network model for predicting cancer immunity responses. It was repurposed for COVID-19.
Meanwhile, a DeepMind team used AlphaFold — a protein-folding neural network that we’ll meet again later — to predict the 3D structures of the SARS-CoV-2 virus’s proteins from its genetic sequence. Understanding protein structures was another key step in designing vaccines.
Most smartphones today support hardware-accelerated neural network training and inference. Machine learning software like TensorFlow and Core ML can run neural networks on smartphones. Model quantization has enabled complex models to fit a smartphone's memory. Artificial intelligence at the edge is finally a reality.
Companies like Butterfly Network and Livongo are running deep neural networks on smartphones for innovative medical diagnostics.
For example, Butterfly sells iQ+, a smart portable ultrasound device for consumers. It connects to a person's smartphone to display ultrasound imagery in real-time. They can monitor their womb, heart, lungs, bladder, and other organs. The app measures and diagnoses various health conditions using neural networks.
Real-time images or raw data from a portable device or camera are processed by one or more deep neural networks. A convolutional neural network (CNN) is particularly good at detecting the visual features that are characteristic of each health condition. What happens next depends on the purpose of the network. A classifier network tells the person the nature of the condition it found. A detection or segmentation network reports the location, shape, volume, or pixels of a condition.
Sometimes, a condition can only be detected by analyzing the changes in images or other measurements over time. A recurrent neural network or transformer network is good at detecting such sequence changes.
But machine learning and deep learning are capable of a lot more, such as correlating across different kinds of data. For example, a deep learning model can be trained to associate ultrasound imagery features with clinical information extracted from electronic or paper health records using intelligent document processing.
Such a model can be deployed to a patient's or doctor's smartphone to assist them by suggesting clinical diagnoses based on ultrasound images.Conversely, if the model notices a diagnosis in a new patient's health records, it verifies that their ultrasound matches the expected image and alerts the doctor if there's a mismatch.
If you are a biotech scientist or professional with an innovative device idea, this is a great time to commercialize it. Machine learning has enabled device-side intelligent processing of images, videos, text, and speech. A problem that you assumed is too difficult may actually be easily solved now by hiring the right machine learning expert.
Drug discovery used to be a resource-intensive, time-consuming, trial-and-error process. Biochemists had to manually apply thousands of chemicals to infected tissue samples under controlled conditions. After a few days, the cultures had to be examined one by one manually for desired effects. Automated examination was a difficult problem due to the variety of images.
Advances in computer vision and deep learning over the last decade have intelligently automated, accelerated, and scaled this process by an order of magnitude.
For example, consider a single gene disorder like sickle cell anemia. It's the effect of an abnormal protein due to a mutated gene. Proteins are essential to the normal functioning of our organs and cells. They function as enzymes, hormones, antibodies, and transport at the cellular level.
Genes, DNA, and proteins are closely related.
The nucleic acids, DNA and RNA, are long polymers made up of nucleotides. Each nucleotide consists of a nucleobase molecule represented by its first letter A, G, C, T, or U. This is why we describe DNA using DNA sequences of these letters. Human DNA consists of around 6 billion nucleotide pairs in total packed into 46 chromosomes of different DNA lengths.
Unlike chromosomes and DNA, a gene is not another physical thing. It’s merely an abstract label for a particular segment of DNA. A gene refers to a DNA segment that has been experimentally determined to encode for a protein. Each protein is a long molecule made up of some combination of the 20 amino acids. Each group of three nucleotides in a gene represents one of these 20 amino acids.
In this way, a gene acts as a template to build a particular protein. The process — involving RNA transcription and translation — is called gene expression. A mutated gene means its nucleotide sequence is abnormal. That results in an abnormal amino acid sequence and hence an abnormal protein.
Drug discovery aims to find a chemical that disables the production of an abnormal protein by disabling the expression of its mutated gene. It can be scaled up by using robots to simultaneously apply thousands of drugs to samples. The samples are monitored using automated fluorescent microscopy.
The microscope images of these thousands of samples are analyzed by convolutional neural networks trained on large datasets of such images. They look for minor visual changes in shape, morphology, area, counts, and other such traits collectively called the sample's phenotype. The neural networks classify the samples as diseased or treated or segment out desired components for further processing using segmentation neural networks like U-net.
In this way, companies like Recursion are using machine learning and deep learning to speed up drug discovery by a factor of at least 10x. Classical computer vision techniques like thresholding and watershed segmentation that involved manually adjusted parameter values for each type of cell are no longer necessary. Moreover, deep learning does not suffer from the manual feature selection and accuracy problems that hobbled classical pattern recognition methods like support vector machines, decision trees, and random forests.
Cancer therapy is often a combination of approaches involving chemotherapy, immunotherapy, and others. However, individual responses to therapies can differ widely. A treatment approach that worked for one person or one type of cancer may not work for another person or a variant of the same cancer. Such data is said to have high variability.
Machine learning is helping design personalized cancer therapies instead.
Notable Labs is one such company designing them using automated drug sensitivity screening. They receive blood and bone marrow samples from partner hospitals and apply a large number of FDA-approved drugs and drug combinations to the samples. The samples are observed using high-throughput flow cytometry that involves illuminating cells with multiple lasers.
White blood cell counts and phenotype traits like cell surface, cell membrane integrity, and morphology are measured by convolutional neural networks. The observed cell changes are correlated against the drugs and clinical observations using hierarchical clustering, principal component analysis, and analysis of variance.
They enable the company to detect drugs that showed maximum sensitivity and efficacy on that particular patient's cancer cells. The patient's treatment is then started using that drug or drug combination. Their outcome analysis showed regression of cancer in many patients and proved the effectiveness of the approach.
Reverie Labs is another company working on automated drug discovery for cancers. Their platform is capable of generating potential drug molecules based on desired molecular properties. They use generative techniques like genetic algorithms and generative neural networks to come up with new drug ideas.
Drugs are large molecules consisting of a number of smaller molecules held together by chemical bonds. Graphs are a natural way of modeling them — atoms are nodes and bonds are edges. Various graph neural networks can then be used to model and predict their molecular properties and behaviors.
Graph convolutional networks were popular for this kind of molecular graph processing till 2019. However, just like regular convolutional networks, they are good at detecting and aggregating local features but not higher-order graph connections and properties. So, graph transformer networks are becoming popular in the field. Thanks to their self-attention mechanism, they can detect relationships over larger areas of the graph while paying attention to the surrounding chemical context of each atom.
These techniques are applicable to any disease or disorder. If you are a biotech scientist or a medical specialist who has novel therapeutic ideas, we urge you to discuss your ideas with a machine learning expert. You may be surprised just how many difficult problems have become much easier now.
A genome is the set of all genes. Genomics is the study of health and other impacts of the genome.
Gene expression is the process of creating proteins from genetic instructions. Gene expression results in phenotype traits that affect everything from cells to body structure.
Thanks to quicker and cheaper next-generation sequencing methods, genome datasets and genetic signatures are increasingly being used in clinical decisions.
However, gene expression data structure cannot yet be exploited by neural networks. Worse, it's very high-dimensional data which makes any kind of data analysis noisy. So, the focus has been on reducing its dimensionality.
Machine learning techniques like autoencoders and variational autoencoders are used to reduce dimensionality. These unsupervised learning approaches have middle layers that create low-dimensional representations of gene expression data. These representations — also known as embeddings — can be used for prediction and classification tasks more efficiently.
Proteins perform a critical role at the cellular and higher levels. They act as enzymes, antibodies, hormones, and much more. The set of all proteins expressed in a body is its proteome. Proteomics is the study of the proteome.
Proteins are large polymers made up of long chains of the 20 amino acids. But it's not merely the protein sequence that affects its behavior. The entire 3D structure and its orientation also affect its behavior. Protein folding is the process where an amino acid chain attains a natural folded 3D structure to become a protein capable of performing its biological functions correctly.
Since proteins function as antibodies and enzymes, it's often necessary to engineer them to fight diseases and disorders.
However, predicting how an amino acid chain folds had been an unsolved problem for many years until 2020 when DeepMind announced their AlphaFold neural network architecture. It was able to predict protein folding with high accuracy.
AlphaFold uses two concepts to achieve this. First, contact between amino acid residues — the part of an amino acid molecule that makes it unique from other amino acids — is a good predictor of the protein structure. Second, defining a potential energy metric for this contact and optimizing it using machine learning methods predicts the final structure accurately.
AlphaFold uses a convolutional neural network to predict the distance between pairs of amino acid residues of an input sequence. These distances are used to formulate a potential function. Minimizing this potential function using gradient descent results in a protein structure that has the lowest energy and matches real-life protein structures.
Since graph convolutional networks and graph transformer networks have proved effective in other bioinformatics and modeling tasks, we can perhaps anticipate their use in improved versions of AlphaFold in the future.
For decades, the processing of images, videos, audio, and text were considered difficult barriers. Many innovations in biology and life sciences were never undertaken because the scientists and researchers who thought them up did not know how to overcome these problems.
But deep learning running on inexpensive cloud and smartphone hardware has dismantled these barriers. The data engineering aspects of biological innovations are a solved problem today. You scientists, researchers, and founders can focus on biology while letting data experts like us help you out with the rest. Contact us with your innovation idea and let's bring it to life together.
