Bulletin of the American Physical Society
APS March Meeting 2017
Volume 62, Number 4
Monday–Friday, March 13–17, 2017; New Orleans, Louisiana
Session S5: Machine Learning for Modeling and Control of Biological Systems IFocus
|
Hide Abstracts |
Sponsoring Units: DBIO GSNP Chair: Ilya Nemenman, Emory University Room: 264 |
Thursday, March 16, 2017 11:15AM - 11:51AM |
S5.00001: Dynamic control and model inference of signal activated gene regulation Invited Speaker: Gregor Neuert Despite robust efforts over the years, it has proven very difficult to identify mathematical models that would improve biological insight by predicting complex biological responses, as needed to accelerate the design of medical treatments. This problem remains unsolved because large-scale models, with hundreds of unknown reaction rates, may be too complex to be supported by existing experimental techniques or data sets, and therefore provide little quantitative insight. At the other extreme, overly simple models ignore the intricacies of real biological processes and are equally limited in their ability to predict real phenomena. Models are often further limited by the fact that most experimental analyses only probe average equilibrium characteristics of cell populations and ignore potentially useful information contained in measurable fluctuations in space, time and environment, and from one cell to another. This is a fundamental problem in all of biology, because models and parameters that are identified from measurements of population of cells do not capture the variability in biological processes and therefore these models are profoundly misleading. In essence, models inferred from population averages can fit the population data very well but they cannot predict. The key to overcoming these limitations is to generate single-cell and single-molecule experimental data sets of high quality and reproducibility that capture the variability in biological processes [1-3]. Because single-cell data contain information hidden in population averages, our cellular systems identification methodology of integrating quantitative single-cell experiments with stochastic mathematical models is maximally predictive [3-7]. Our approach is very broad and may be applied to any measurement that detects variability and any biological process that exhibits variability. [1] van Werven, Neuert, et al., Cell, 150 (6), 1170--1181, 2012. [2] Bumgarner, Neuert, et al., Molecular Cell, 45 (4), 470-482, 2012. [3] Neuert, Munsky, et al., Science, 339(6119):584--587, 2013. [4] Munsky, Neuert, et al., Science, 336(6078):183--187, 2012. [5] Munsky, Neuert. Physical Biology, 12(4):045004, 2015. [6] Munsky, Fox, Neuert. Methods, 85:12--21, 2015. [7] Fox, Neuert, Munsky, The Journal of chemical physics, 145(7). 074101, 2016. [Preview Abstract] |
Thursday, March 16, 2017 11:51AM - 12:03PM |
S5.00002: Controllability of energy landscapes by varying correlations between minima Sai Teja Pusuluri, Alex H. Lang, Pankaj Mehta, Horacio E. Castillo Neural network models have been used recently to generate complex energy landscapes for modeling various biological processes, including protein folding, HIV evolution and cellular reprogramming [1-5]. In these constructions, the minima of the energy landscape correspond to memory patterns of the neural network. A good understanding of the static and dynamic properties of energy landscapes can be helpful in gaining better insight into those processes. Here, we demonstrate that the correlations between memory patterns strongly affect some of those properties, including the basin sizes of energy minima, the density of metastable states, the stability of global energy minima against perturbations, and the switching rates between global energy minima in the presence of an external bias. \\ \\$[1]$ Kanter, I & Sompolinsky, H. (1987) Physical Review A 35, 380–392. { } [2] Pusuluri, S. T, Lang, A. H, Mehta, P, & Castillo, H. E. arXiv:1505.03889 { }[3] Lang, A. H, Li, H, Collins, J. J, & Mehta, P. (2014) PLoS computational biology 10, e1003734.{ } [4] A. L. Ferguson, J. K. Mann, S. Omarjee, T. Ndung’u, B. D. Walker, and A. K. Chakraborty, Immunity 38, 606 (2013). { } [5] H. Frauenfelder, S. G. Sligar, P. G. Wolynes Science 254.5038 (Dec 13, 1991) [Preview Abstract] |
Thursday, March 16, 2017 12:03PM - 12:15PM |
S5.00003: PCA meets RG Serena Bradde, William Bialek A system with many degrees of freedom can be characterized by a covariance matrix; principal components analysis (PCA) focuses on the eigenvalues of this matrix, hoping to find a lower dimensional description. But when the spectrum is nearly continuous, any distinction between components that we keep and those that we ignore becomes arbitrary; it then is natural to ask what happens as we vary this arbitrary cutoff. We argue that this problem is analogous to the momentum shell renormalization group (RG). Following this analogy, we can define relevant and irrelevant operators, where the role of dimensionality is played by properties of the eigenvalue density. These results also suggest an approach to the analysis of real data. As an example, we study neural activity in the vertebrate retina as it responds to naturalistic movies, and find evidence of behavior controlled by a nontrivial fixed point. Applied to financial data, our analysis separates modes dominated by sampling noise from a smaller but still macroscopic number of modes described by a non--Gaussian distribution. [Preview Abstract] |
Thursday, March 16, 2017 12:15PM - 12:27PM |
S5.00004: Inferring phenomenological models for dynamics of Purkinje neurons Catalina Rivera, David Hofmann, Ilya Nemenman Purkinje neurons are typically described by multi-compartmental models that try to reproduce their complex dendritic structure. These models are very hard to solve computationally, and due to the high number of parameters they are likely to overfit, and therefore are not predictive. Here we build an effective phenomenological model to describe the inter-spike interval probability distribution of a highly complex Purkinje neuron model data set as a function of the injected current. From a hierarchical set of Markov models we select the simplest model able to explain the data, where each state in the Markov model represents an effective state of the neuron. This procedure allows us to construct a coarse-grained model of the system in an automated manner directly from data, without having to build a microscopically accurate description of the system first. We found that a Markov model with about 10 states provides a good fit for the data generated by a morphologically accurate model with about 1000 compartments. [Preview Abstract] |
Thursday, March 16, 2017 12:27PM - 12:39PM |
S5.00005: Imaging localized force displacements in cells and tissues from substrate displacements Joshua Chang, Tom Chou We develop a method to reconstruct, from measured displacements of the underlying elastic substrate, the spatially dependent forces that cells or tissues impart on them. Since these sources of force typically arise from focal adhesions, with are localized or compact," and discontinuous, we solve this inverse problem using methods of optimization useful for image segmentation. In addition to the standard quadratic data mismatch terms (that defines least-squares fitting), we motivate a term in the objective function which penalizes variations in the reconstructed stress eld while preserving boundaries and physical knowledge. By minimizing the objective function subject to physical constraints, we are able to efficiently reconstruct stress fields with localized structure from simulated and experimental substrate displacements. We provide a numerical algorithm for setting up the discretized problem and under loose regularity conditions for the underlying stress tensor, provide bounds on the reconstruction error. [Preview Abstract] |
Thursday, March 16, 2017 12:39PM - 12:51PM |
S5.00006: Multichannel microformulators for massively parallel machine learning and automated design of biological experiments John Wikswo, Aditya Kolli, Harish Shankaran, Matthew Wagoner, Jerome Mettetal, Ronald Reiserer, Gregory Gerken, Clayton Britt, David Schaffer Genetic, proteomic, and metabolic networks describing biological signaling can have 10$^2$ to 10$^3$ nodes. Transcriptomics and mass spectrometry can quantify 10$^4$ different dynamical experimental variables recorded from $in$ $vitro$ experiments with a time resolution approaching 1 s. It is difficult to infer metabolic and signaling models from such massive data sets, and it is unlikely that causality can be determined simply from observed temporal correlations. There is a need to design and apply specific system perturbations, which will be difficult to perform manually with 10 to 10$^2$ externally controlled variables. Machine learning and optimal experimental design can select an experiment that best discriminates between multiple conflicting models, but a remaining problem is to control in real time multiple variables in the form of concentrations of growth factors, toxins, nutrients and other signaling molecules. With time-division multiplexing, a microfluidic MicroFormulator ($\mu$F) can create in real time complex mixtures of reagents in volumes suitable for biological experiments. Initial 96-channel $\mu$F implementations control the exposure profile of cells in a 96-well plate to different temporal profiles of drugs; future experiments will include challenge compounds. [Preview Abstract] |
Thursday, March 16, 2017 12:51PM - 1:03PM |
S5.00007: Discovering fine structure in big biological data Alexandre Day, Pankaj Mehta, Amir Erez, Grégoire Altan-Bonnet Modern experimental methods allow for an unprecedented high-dimensional single-cell resolution of complex tissues. We propose a novel and robust method to perform deep unsupervised learning to discover and analyze the structure of such data. Our approach harvest the power of neural networks to self-consistently identify and validate sub-structures in the data. We test our method on multi-dimensional protein distributions and show that while we are able to reproduce some experimental benchmarks, our approach is able to uncover previously unidentified structures in the data. We use our approach as a basis to modelling the population distribution of dynamical biological datasets. [Preview Abstract] |
Thursday, March 16, 2017 1:03PM - 1:15PM |
S5.00008: Identifying mechanisms for superdiffusive dynamics in cell trajectories Giuseppe Passucci, Megan Brasch, James Henderson, M Lisa Manning Self-propelled particle (SPP) models have been used to explore features of active matter such as motility-induced phase separation, jamming, and flocking, and are often used to model biological cells. However, many cells exhibit super-diffusive trajectories, where displacements scale faster than $t^{1/2}$ in all directions, and these are not captured by traditional SPP models. We extract cell trajectories from image stacks of mouse fibroblast cells moving on 2D substrates and find super-diffusive mean-squared displacements in all directions across varying densities. Two SPP model modifications have been proposed to capture super-diffusive dynamics: Levy walks and heterogeneous motility parameters. In mouse fibroblast cells displacement probability distributions collapse when time is rescaled by a power greater than $\frac{1}{2}$, which is consistent with Levy walks. We show that a simple SPP model with heterogeneous rotational noise can also generate a similar collapse. Furthermore, a close examination of statistics extracted directly from cell trajectories is consistent with a heterogeneous mobility SPP model and inconsistent with a Levy walk model. Our work demonstrates that a simple set of analyses can distinguish between mechanisms for anomalous diffusion in active matter. [Preview Abstract] |
(Author Not Attending)
|
S5.00009: Using Maximum Entropy to Find Patterns in Genomes Sophia Liu, Adam Hockenberry, Andrea Lancichinetti, Michael Jewett, Luis Amaral The existence of over- and under-represented sequence motifs in genomes provides evidence of selective evolutionary pressures on biological mechanisms such as transcription, translation, ligand-substrate binding, and host immunity. To accurately identify motifs and other genome-scale patterns of interest, it is essential to be able to generate accurate null models that are appropriate for the sequences under study. There are currently no tools available that allow users to create random coding sequences with specified amino acid composition and GC content. Using the principle of maximum entropy, we developed a method that generates unbiased random sequences with pre-specified amino acid and GC content. Our method is the simplest way to obtain maximally unbiased random sequences that are subject to GC usage and primary amino acid sequence constraints. This approach can also be easily be expanded to create unbiased random sequences that incorporate more complicated constraints such as individual nucleotide usage or even di-nucleotide frequencies. The ability to generate correctly specified null models will allow researchers to accurately identify sequence motifs which will lead to a better understanding of biological processes. [Preview Abstract] |
Thursday, March 16, 2017 1:27PM - 1:39PM |
S5.00010: An experimental-computational platform for investigating microbial interactions and dynamics in communities with two codependent species Miguel Fuentes-Cabrera, John D. Anderson, Jared Wilmoth, Marta Ginovart, Clara Prats, Xavier Portell-Canal, Scott Retterer Microbial interactions are critical for governing community behavior and structure in natural environments. Examination of microbial interactions in the lab involves growth under ideal conditions in batch culture; conditions that occur in nature are, however, characterized by disequilibrium. Of particular interest is the role that system variables play in shaping cell-to-cell interactions and organization at ultrafine spatial scales. We seek to use experiments and agent-based modeling to help discover mechanisms relevant to microbial dynamics and interactions in the environment. Currently, we are using an agent-based model to simulate microbial growth, dynamics and interactions that occur on a microwell-array device developed in our lab. Bacterial cells growing in the microwells of this platform can be studied with high-throughput and high-content image analyses using brightfield and fluorescence microscopy. The agent-based model is written in the language Netlogo, which in turn is "plugged into" a computational framework that allows submitting many calculations in parallel for different initial parameters; visualizing the outcomes in an interactive phase-like diagram; and searching, with a genetic algorithm, for the parameters that lead to the most optimal simulation outcome. [Preview Abstract] |
Thursday, March 16, 2017 1:39PM - 2:15PM |
S5.00011: Inference for single molecules Invited Speaker: Steve Presse Bursts in experimental progress have helped drive the punctuated development of successive fields of Mathematics and Statistics. Most recently, the development of new imaging methods -- that often exploit fluorescence probes to enhance contrast -- have provided data at length and time scales previously inaccessible. While modeling fluorescence data has contributed to bringing data-driven methods into the mainstream of the physical sciences, more complex systems, such as live cells, demand model adaptability and improvements brought to commonly used data-driven methods (such as Hidden Markov Models) have reached a point of diminishing returns. Here I discuss some recent work in my lab, both parametric and nonparametric, toward gaining deeper insight from indirect observations of microscopic processes, often through fluorescent probes. [Preview Abstract] |
Follow Us |
Engage
Become an APS Member |
My APS
Renew Membership |
Information for |
About APSThe American Physical Society (APS) is a non-profit membership organization working to advance the knowledge of physics. |
© 2024 American Physical Society
| All rights reserved | Terms of Use
| Contact Us
Headquarters
1 Physics Ellipse, College Park, MD 20740-3844
(301) 209-3200
Editorial Office
100 Motor Pkwy, Suite 110, Hauppauge, NY 11788
(631) 591-4000
Office of Public Affairs
529 14th St NW, Suite 1050, Washington, D.C. 20045-2001
(202) 662-8700