Identification of causal drivers behind regulatory gene networks is crucial in understanding gene function. Here, we develop a method for the large-scale inference of gene–gene interactions in observational population genomics data that are both directed (using local genetic instruments as causal anchors, akin to Mendelian Randomization) and specific (by controlling for linkage disequilibrium and pleiotropy). Analysis of genotype and whole-blood RNA-sequencing data from 3072 individuals identified 49 genes as drivers of downstream transcriptional changes (Wald P < 7 × 10−10), among which transcription factors were overrepresented (Fisher’s P = 3.3 × 10−7). Our analysis suggests new gene functions and targets, including for SENP7 (zinc-finger genes involved in retroviral repression) and BCL2A1 (target genes possibly involved in auditory dysfunction). Our work highlights the utility of population genomics data in deriving directed gene expression networks. A resource of trans-effects for all 6600 genes with a genetic instrument can be explored individually using a web-based browser.

This talk investigates three challenges in ML and radiology. Firstly, ML models trained on specialized medical data generalize poorly out-of-domain. Secondly, automatic labelling with NLP only provides noisy ground truth. Thirdly, classical non-ML based image registration runs inefficiently. Some ML based recipes are presented to alleviate the aforementioned issues.

In many cases, especially with medical images, it is prohibitively challenging to produce a sufficiently large training sample of pixel-level annotations to train deep neural networks for semantic image segmentation. On the other hand, some information is often known about the contents of images. We leverage information on whether an image presents the segmentation target or whether it is absent from the image to improve segmentation performance by augmenting the amount of data usable for model training. Specifically, we propose a semi-supervised framework that employs image-to-image translation between weak labels (e.g., presence vs. absence of cancer), in addition to fully supervised segmentation on some examples. We conjecture that this translation objective is well aligned with the segmentation objective as both require the same disentangling of image variations. Building on prior image-to-image translation work, we re-use the encoder and decoders for translating in either direction between two domains, employing a strategy of selectively decoding domain-specific variations. For presence vs. absence domains, the encoder produces variations that are common to both and those unique to the presence domain. Furthermore, we successfully re-use one of the decoders used in translation for segmentation. We validate the proposed method on synthetic tasks of varying difficulty as well as on the real task of brain tumor segmentation in magnetic resonance images, where we show significant improvements over standard semi-supervised training with autoencoding.

Gene interaction graphs aim to capture various relationships between genes and can be used to create more biologically-intuitive models for machine learning. There are many such graphs available which can differ in the number of genes and edges covered. In this work, we attempt to evaluate the biases provided by those graphs through utilizing them for 'Single Gene Inference' (SGI) which serves as, what we believe is, a proxy for more relevant prediction tasks. The SGI task assesses how well a gene's neighbors in a particular graph can 'explain' the gene itself in comparison to the baseline of using all the genes in the dataset. We evaluate seven major gene interaction graphs created by different research groups on two distinct datasets, TCGA and GTEx. We find that some graphs perform on par with the unbiased baseline for most genes with a significantly smaller feature set.

Abstract: Proteins are the cogs of cellular machinery. Understanding how these molecules interact and predicting structures of their complexes is unavoidable step we need to pass to understand emergence of the phenotype of an organism from its genotype. Rigid-body docking approaches the problem by assuming two proteins rigid and sampling rotations and translations of the ligand with respect to the receptor. The key idea, that allows the sampling to be performed in a reasonable timeframe is representation of the interaction energy between the two proteins as correlations of 3d protein representations. However, these representations themselves are either pretrained or handcrafted. In this work we present an algorithm that allows learning protein representations for docking in the end-to-end fashion.

In North American hospitals, pharmacists review all medication orders to ensure adequate medication use. As pharmacists move to a practice model involving direct patient care, the activity of drug order verification could benefit from a certain degree of automation to free up resources to perform clinical tasks. Potential applications of machine learning include detecting atypical orders and focusing pharmacist attention on unusual patterns instead of verifying every single order. We prototyped a machine learning model to predict the next prescribed drug during hospitalization, and we compared model predictions with pharmacist ratings on a sample of orders. This presentation will discuss health system pharmacy and the data it generates, the clinical problem we faced, how we developed a model that shows potential at helping pharmacists detect atypical orders, and how the predictions of this model correlate with pharmacist ratings.

Machine learning research has had and will continue to have impact in healthcare and medicine. This talk will discuss a few directions which can potentially have a transformative impact. Including clinical decision support, biased feature construction/representations, and unsupervised representation learning. With each area I will consider a pragmatic approach for any method gaining adoption. The intricacies of each potential use case and the climate of adoption all need to be considered in order to maximize impact.

Abstract: We are developing multimodal intraoperative imaging tools based on an imaging platform combining Raman, fluorescence and diffuse reflectance (multi-spectral and SFDI) with the objective of gathering large spectral datasets across organ sites/pathologies and patients for the development of classification models using machine learning techniques. The combinaison of Raman, fluorescence and diffuse reflectance spectroscopic images offers the potential to provide a breadth of molecular features that can be employed for surface as well as for depth-resolved tissue characterisation. Using a versatile wide-field spectroscopic imaging plateform, databases of tissue-specific spectroscopic information can be acquired for future artificial intelligence applications. Here we will present recent results relating with the development of each imaging modality as well as progress towards the development of the multimodal imaging plateform and the wide-field intraoperative imaging probes to be used for data gathering in the scope of clinical studies.

Abstract: Machine learning algorithms are sensitive to meaningless (or "adversarial") perturbations. This is reminiscent of cellular decision-making where ligands (called "antagonists") prevent correct signalling, like in early immune recognition. We draw a formal analogy between neural networks used in machine learning and models of cellular decision-making (adaptive proofreading). We apply attacks from machine learning to simple decision-making models, and show explicitly the correspondence to antagonism by weakly bound ligands. Such antagonism is absent in more nonlinear models, which inspired us to implement a biomimetic defence in neural networks filtering out adversarial perturbations. We then apply a gradient-descent approach from machine learning to different cellular decision-making models, and we reveal the existence of two regimes characterized by the presence or absence of a critical point. The critical point causes the strongest antagonists to lie close to the threshold. This is validated in the loss landscapes of robust neural networks and cellular decision-making models, and observed experimentally for immune cells. For both regimes, we explain how associated defence mechanisms shape the geometry of the loss landscape, and why different adversarial attacks are effective in different regimes. Our work connects evolved cellular decision-making to machine learning, and motivates the design of a general theory of adversarial perturbations, both for in vivo and in silico systems.

Abstract: In this paper, the authors applied a deep learning approach for predicting the synergy of drug combinations. DeepSynergy uses compound information as well as genomic information as inputs to a regression task in order to find specific drug combinations that have maximal efficacy in specific genomic context, in their case in given cancer cell lines.

Abstract: In this study, they build a gene regulatory network (GRN) for the brain, linking transcription factors, enhancers, and target genes from merging of the QTLs, generalized element-activity correlations, and Hi-C data. They integrate this GRN to a deep Boltzmann machine, by manually adding links between the visible nodes of the network given the relations from the GRN, to perform disease prediction for psychiatric disorders.