Generative models of observations under interventions have been a vibrant topic of interest across machine learning and the sciences in recent years. For example, in drug discovery, there is a need to model the effects of diverse interventions on cells in order to characterize unknown biological mechanisms of action. We propose the Sparse Additive Mechanism Shift Variational Autoencoder, SAMS-VAE, to combine compositionality, disentanglement, and interpretability for perturbation models. SAMS-VAE models the latent state of a perturbed sample as the sum of a local latent variable capturing sample-specific variation and sparse global variables of latent intervention effects. Crucially, SAMS-VAE sparsifies these global latent variables for individual perturbations to identify disentangled, perturbation-specific latent subspaces that are flexibly composable. We evaluate SAMS-VAE both quantitatively and qualitatively on a range of tasks using two popular single cell sequencing datasets.In order to measure perturbation-specific model-properties, we also introduce a framework for evaluation of perturbation models based on average treatment effects with links to posterior predictive checks. SAMS-VAE outperforms comparable models in terms of generalization across in-distribution and out-of-distribution tasks, including a combinatorial reasoning task under resource paucity, and yields interpretable latent structures which correlate strongly to known biological mechanisms. Our results suggest SAMS-VAE is an interesting addition to the modeling toolkit for machine learning-driven scientific discovery.

When David Fajgenbaum was a twenty-five-year-old medical student, at the University of Pennsylvania, he started to feel so tired that he could barely stand. Fajgenbaum, a former college quarterback, could still bench-press three hundred and seventy-five pounds; he was known for doing pullups on a tree near his workplace. But now he was desperately ill. The lymph nodes in his groin and neck swelled. Small red dots--blood moles--emerged on his chest, and he woke up soaked in sweat. One day, at the hospital where he was doing his rotation, he stumbled down the hall into the emergency room, and doctors told him that his liver, bone marrow, and kidneys were failing. Fluid had leaked out of his blood vessels, into his abdomen and around his heart; bleeding in his retina temporarily blinded him in his left eye.

Generative models of observations under interventions have been a vibrant topic of interest across machine learning and the sciences in recent years. For example, in drug discovery, there is a need to model the effects of diverse interventions on cells in order to characterize unknown biological mechanisms of action. We propose the Sparse Additive Mechanism Shift Variational Autoencoder, SAMS-VAE, to combine compositionality, disentanglement, and interpretability for perturbation models. SAMS-VAE models the latent state of a perturbed sample as the sum of a local latent variable capturing sample-specific variation and sparse global variables of latent intervention effects. Crucially, SAMS-VAE sparsifies these global latent variables for individual perturbations to identify disentangled, perturbation-specific latent subspaces that are flexibly composable.

Deep learning is solving biology's deepest secrets at breathtaking speed. Just a month ago, DeepMind cracked a 50-year-old grand challenge: protein folding. A week later, they produced a totally transformative database of more than 350,000 protein structures, including over 98 percent of known human proteins. Structure is at the heart of biological functions. The data dump, set to explode to 130 million structures by the end of the year, allows scientists to foray into previous "dark matter"--proteins unseen and untested--of the human body's makeup.

Artificial intelligence (AI) is a broad and evolving scientific field, and the value it can deliver at various stages of the drug discovery process is now widely accepted in the pharmaceutical industry. This blog seeks to demystify the application of AI in drug discovery, focusing on its key challenges, opportunities and successes. Over one million scientific articles are published every year in the biomedical domain alone, and every new year brings new methods for data collection and more detailed data modalities. While scientists have access to an exponentially increasing amount of knowledge and data, biological data is messy and incomplete; it may contain conflicting or contradicting evidence, suppositions, biases, uncertainty, gaps in knowledge or misclassifications. This prevents us from understanding the full biology landscape and complicates decision making.

A University of Michigan-led research team has uncovered a neural network that enables Drosophila melanogaster fruit flies to convert external stimuli of varying intensities into a "yes or no" decision about when to act. The research, described in Current Biology, helps to decode the biological mechanism that the fruit fly nervous system uses to convert a gradient of sensory information into a binary behavioral response. The findings offer up new insights that may be relevant to how such decisions work in other species, and could possibly even be applied to help artificial intelligence machines learn to categorize information. Senior study author Bing Ye, PhD, a faculty member at the University of Michigan Life Science Institute (LSI), believes the mechanism uncovered could have far-reaching applications. "There is a dominant idea in our field that these decisions are made by the accumulation of evidence, which takes time," Ye said.

Today, the search for biological mechanisms at molecular scale can leverage an unprecedented amount of information. With the recent development of high-throughput technologies, data collection has received an enormous impulse that has radically changed the perspective toward molecular biology. The main protagonist of this shift is omic data--namely, experimental profiles with large coverage over multiple biological domains. Several levels of knowledge have become associated with emerging omic technologies [1–3]. The most widespread to date include DNA sequencing (genomics), microarrays and RNA sequencing (transcriptomics), DNA methylation and histone modifications (epigenomics), and protein or metabolite mass spectrometry (proteomics and metabolomics). As technology moves forward, its associated costs decrease, and a growing wealth of data is being generated.

Artificial Intelligence (AI) has enabled the development of high-performance automatic learning techniques in recent years. However, these techniques are often applied task by task, which implies that an intelligent agent trained for one task will perform poorly on other tasks, even very similar ones. To overcome this problem, researchers at the University of Liège (ULiège) have developed a new algorithm based on a biological mechanism called neuromodulation. This algorithm makes it possible to create intelligent agents capable of performing tasks not encountered during training. This novel and exceptional result is presented this week in the magazine PLOS ONE.