Jesper Ferkinghoff-Borg - Research

Identifying cellular network states using deep learning   

Deep learning is reshaping the field of AI and has been shown to outperform traditional  techniques on a variety of tasks, including visual and auditive recognition modeling. To the best of our knowledge we are the first in the world to develop and apply these methodologies to identify cellular network states from High Content Screening HCS-imaging (using LindingLabs PerkinElmer Opera HCS-platform). We apply deep-learning to identify the connection between kinase activities and cell morphology and use this information to  decipher the driving mechanisms behind cell invasion, migration, metastasis and resistance development. We characterize each cell by a ~600 dimensional morphological feature vector, entailing shapes, textures and intensities.The HTC-platform facilitates the imaging and phenotyping of millions of such cells, implying that our recognition modeling problem belongs to the realm of big data.

Data-integration and biological forecasting

As part of LindingLabs goal to develop personalized precision medicine for cancer and other complex diseases, I head the developing of the models and algorithms required to integrate multiplexed and multi-omics data. This includes global mass spectrometry, HCS-imaging, exome sequencing and RNA expression levels as well as migration and proliferation characterization on different cell-lines under different treatment conditions, siRNA-knockdown and/or other perturbative methods. In each case, we derive probabilistic models to account for the heterogeneous and heteroscedastic nature of the data, thus avoiding the use of arbitrary cut-offs usually applied in systems-biological data-processing. On this basis we have extended bayesian non-negative matrix factorization to identify coupled phospho-proteomic and genetic signatures driving metastasis. We are currently extending these as well as the deep-learning approaches to integrate HCS- and RNA-expression data  in the analysis. We have recently demonstrated that a global network modeling approach can identify mutations which are potentially oncogenic due to their network attacking properties.

Machine learning in structural biology

Proteins are the primary operational unit in cellular networks. Protein function is intimately linked to  its structure and dynamics. We are interested in understanding the determinants of protein stability, binding and dynamics, the mechanisms gearing protein allostery, folding and misfolding and their relation to human diseases, in particular cancer and amyloidosis.

We have applied Hidden Markov models, directional statistics and generalized ensemble techniques to perform protein structure determination and design, particular in conjunction with low-resolution data such as SAXS, NMR and chemical shift measurements. These approaches have been expounded in the recently published book “Bayesian methods in Structural Bioinformatics” (Springer, April 2012), which I have co-edited with Prof. K. Mardia and T. Hamelryck.  Lately, we have in the LindingLab developed a gaussian process methodology to predict binding affinities between kinases and inhibitor-drugs. We are in the process of extending these kernel-approaches in the Lab to predict other types binding interaction partners and motifs. We are also addressing protein binding, allostery and (mis-)folding from a more biophysical perspective (see below).

Publications

1. Hamelryck T et al.A Bayesian formulation of the protein folding problem. Geometry Driven Statistics 121, 356 (2015).
2. Boomsma W., Tian P, Frellsen J, Ferkinghoff-Borg J. et al. Equilibrium simulations of proteins using molecular fragment replacement and NMR chemical shifts. Proc. Natl. Acad Sci. 111, 13852-13857 (2014).
3. Olsson S., Vogeli, Cavalli A, Boomsma W, Ferkinghoff-Borg J, Lindorff-Larsen K, Hamelryck T. Probabilistic Determination of Native State Ensembles of Proteins.  J. Chemical Theory and Computation 8, 3484-3491(2014).
4. Valentin, J. B., Andreetta, C., Boomsma, W., Bottaro, S., Ferkinghoff‐Borg, J., Frellsen, et al. Formulation of probabilistic models of protein structure in atomic detail using the reference ratio method. Proteins: Structure, Function, and Bioinformatics 82, 288-299 (2013).
5. Harder, T., Borg, M., Bottaro, S., Boomsma, W., Olsson, S., Ferkinghoff-Borg, J., Hamelryck, T. An efficient null model for conformational fluctuations in proteins. Protein Structure, 20, 1028-1039 (2012).
6. Bottaro, S., Boomsma, W., Johansson, K.E., Andreetta, C., Hamelryck, T., Ferkinghoff-Borg, Subtle Monte Carlo updates in dense molecular systems. J. Chem. Theory Comput. (2012), 8, 695–702.
7. Borg, M., Hamelryck, T. Ferkinghoff-Borg J. On the physical relevance and statistical interpretation of knowledge based potentials. In T. Hamelryck, K. Mardia and J. Ferkinghoff-Borg (eds). Bayesian methods in structural bioinformatics. Statistics for Biology and Health. Springer-Verlag, Berlin, Heidelberg. (2012).
8. Frellsen, J., Mardia, KV., Borg, M., Ferkinghoff-Borg, J. and Hamelryck, T. Towards a probabilistic model of protein structure: The reference ratio method. In T. Hamelryck, K. Mardia and J. Ferkinghoff-Borg (eds). Bayesian methods in structural bioinformatics. Statistics for Biology and Health. Springer-Verlag, Berlin, Heidelberg. (2012).
9. Olsson, S., Boomsma, W., Frellsen, J., Bottaro, S., Harder, T., Ferkinghoff-Borg, J., Hamelryck, Generative probabilistic models extend the scope of inferential structure determination. J. Magn. Reson., 213(1), 182-6 (2011).
10. T. Hamelryck, M. Borg, M. Paluszewski, J. Paulsen, J. Frellsen, C. Andreetta, W. Boomsma, S. Bottaro, J. Ferkinghoff-Borg, Potentials of mean force for protein structure prediction vindicated and generalized. PLoS ONE 5(11): e13714 (2010).
11. K. Stovgaard, C. Andreetta, J. Ferkinghoff-Borg and T. Hamelryck. Calculation of accurate small angle X-ray scattering curves from coarse-grained protein models. BMC Bioinformatics 11:429 (2010).
12. J. Frellsen, I. Moltke, M. Thiim, K. V. Mardia, J. Ferkinghoff-Borg and T. Hamelryck. A probabilistic model of RNA conformational space. PloS Comput. Biol. 5(6), e1000406 (2009).
13. W. Boomsma, K.V. Mardia, C. C. Taylor, J. Ferkinghoff-Borg, A. Krogh and T. Hamelryck. A generative probabilistic model of local protein structure.  Proc Natl. Acad. Sci USA 26,, 8932-8937 (2008)

Biophysics of protein structures and dynamics

My interest and work in this field entails prediction of protein stability, binding, folding, design and aggregation. I have been a prime developer of the widely used Fold-x algorithm  for these purposes. My work also entails more simplified, coarse-grained polymer models  to understand generic statistical aspects of protein folding and design, and analytical solutions to the geometrical loop closure problem in polymer kinetics used to increase the efficiency of Monte Carlo sampling of bio macro-molecules. In the LindingLab we are currently expanding the scope of Fold-X to predict mutations which cause significant rewiring of cellular networks and thus lead to aberrant cellular behavior. 

Publications

1. Tian P, Jonsson S, Ferkinghoff-Borg J. et al. Robust estimation of diffusion-optimized ensembles for enhanced sampling. Journal of Chemical Theory and Computation 10, 543-553 (2014).
2. Boomsma W., Ferkinghoff-Borg J. and Lindorff-Larsen K. Combining experiments and simulations using the maximum entropy principle. Plos Computational Biology 10, e1003406 (2014).
3. T. Lenaerts, J. Ferkinghoff-Borg, J. Schymkowitz and F. Rousseau. Information theoretical quantification of cooperativity in signalling complexes. BMC Systems Biology 3, 9 (2009).
4. K. Lindorff-Larsen and J. Ferkinghoff-Borg, Similarity measures for protein ensembles. Plos ONE {\bf 4}(1): e4203 (2009).
5. T. Lenaerts, J. Ferkinghoff-Borg, F. Stricher, L. Serrano, F. Rousseau and J. Schymkowitz. Quantifying information transfer by protein domains: Analysis of the Fyn SH2 domain structure. BMC structural biology 8:43-58 (2008).
6. Sánchez IE, Beltrao P, Stricher F, Schymkowitz J, Ferkinghoff-Borg J, Rousseau F, Serrano L. Genome-wide prediction of SH2 domain targets using structural information and the FoldX algorithm. PLoS Comput Biol. 4, e1000052 (2008).
7. A. Norgaard, J. Ferkinghoff-Borg, K. Lindorff-Larsen. Experimental parametrization of an energy function for the simulation of unfolded proteins. Biohpys J. 94:182-92 (2008).
8. A. Amatori, J. Ferkinghoff-Borg, G. Tiana and R.A. Broglia. The denatured state is critical in determining the properties of model proteins designed on different folds. Proteins: Structure, Function and Bioinformatics}, 70:1047-1055 (2008).
9. A. Amatori, J. Ferkinghoff-Borg, G. Tiana and R.A. Broglia and G. Tiana. Thermodynamic features characterizing good and bad folding sequences obtained using a simplified off-lattice protein. Phys. Rev. E 73, 061905 (2006).
10. J. Reumers, J. Schymkowitz, J. Ferkinghoff-Borg, F. Stricher, L. Serrano and F. Rousseau. SNPeffect: a database mapping molecular phenotypic effects of human non-synonymous coding SNPs. J. Reumers, J. Schymkowitz, J. Ferkinghoff-Borg, F. Stricher, L. Serrano and F. Rousseau, Nucl Acids Res 33: D527 (2005).
11. A. Amatori, G. Tiana, L. Sutto, J. Ferkinghoff-Borg, A. Trovato, R.A. Broglia, Design of amino acid sequences to fold into C(alpha)-model proteins. J. Chem. Phys. 123, 54904 (2005).
12. J. Schymkowitz, F. Rousseau, I. C. Martins, J. Ferkinghoff-Borg, F. Stricher and L. Serrano. Predicting of water and metal binding sites and their affinities by using the Fold-X force field. Proc. Natl Acad Sci USA, 102: 10147 (2005).
13. C. Kiel, S. Wohlgemuth, F. Rousseau, J. Schymkowitz, J. Ferkinghoff-Borg, F. Wittinghofer and L. Serrano. Recognizing and Defining true Ras Binding Domains II: In Silico prediction based on homology modeling and energy calculations. J. Mol. Biol. 348: 759 (2005).
14. J. Schymkowitz, J. Borg, F. Stricher, R. Nys, F. Rousseau and L. Serrano. The Foldx web server: an on-line force field. Nucl. Acids. Res. 33: W382 (2005).
15. M. A. Micheelsen, C. Rischel, J. Ferkinghoff-Borg, R. Guerois and L. Serrano. Mean first passage time analysis reveals rate-limiting steps,parallel pathways and dead ends in a simple model of protein folding. Europhysics Letters, 61: 561 (2003).
16. A. Trovato, J. Ferkinghoff-Borg and M.H. Jensen. Compact phases of polymers with hydrogen bonding. Physical Review E 67: 021805 (2003).
17. J. Borg, M. H. Jensen, K. Sneppen and G. Tiana. Hydrogen Bonds in Polymer Folding.  Phys. Rev. Lett. 86: 1031-1033 (2001).

Kinetics and biomechanics: from molecules to intra- and inter-cellular communication

I am interested in data-driven modeling of kinetic processes with specific application to intra- and inter-cellular signaling processes. We have applied ordinary and partial differential equation models to understand processes ranging from turbulence, crystal formation, protein fibrillation and aggregation, intra-cellular dynamics, quorum sensing and biofilm formation. In the LindingLab we use gaussian processes and dynamic models to analyse wound closure in order to identify genes playing a key role in cellular migration and proliferation and with a particular view on identifying critical steps in cancer metastasis.

Publications

1. Garde C, Welch M, Ferkinghoff-Borg J, Sams T. Microbial Biofilm as a Smart Material. Sensors 15 (2):4229-4241 (2015).
2. Ferkinghoff-Borg J., Sams. T.  Size of Quorum sensing communities. Molecular Biosystems 10, 103-109 (2014).
3. Claussen, A. , Jakobsen, T. H., Bjarnsholt, T., Givskov, M., Welch M. Ferkinghoff- Borg, J. ; Sams T.,  Kinetic Model for Sigma Binding to the Querom Sensing Regulator LasR. International Journal of Molecular Sciences 14(7), 13360-13376 (2013).
4. J. Ferkinghoff-Borg, J. Fonslet, C. B. Andersen, S. Krishna, S. Pigolotti, H. Yagi, Y. Goto, D. Otzen and M. H. Jensen. Stop and go kinetics in amyloid fibrillation. Phys. Rev. E (Rapid Communication) 82:010901(R) (2010).
5. C. Garde, T. Bjarnsholt, M. Givskov, T. H. Jakobsen, M. Hentzer, A. Claussen, K. Sneppen, J. Ferkinghoff-Borg and T. Sams. Quorum Sensing regulation in Aeromonas hydrophila. J. Mol. Biol. {\bf 396}: 849-859 (2010).
6. J. Ferkinghoff-Borg, J. Mathiesen, M. H. Jensen and P. Olesen,  Diffusion, Fragmentation and Merging: Rate Equations, Distributions and Critical Points, Special issue of Physica D on coagulation and fragmentation 122, Issue 1-2, 88 (2006).
7. J. Ferkinghoff-Borg, M. H. Jensen, J. Mathiesen and P. Olesen. Scale Free Cluster Distributions from Conserving Merging-Fragmentation Processes. Europhys. Letters 73, 422 (2006).
8. P. Olesen, J. Ferkinghoff-Borg, M. H. Jensen and J. Mathiesen. Diffusion, Fragmentation and Coagulation Processes: Analytical and Numerical Results. Phys. Rev. E 72: 031103 (2005).
9. J. Mathiesen, J. Ferkinghoff-Borg, M.H. Jensen, M. Levinsen, P. Olesen, D. Dahl-Jensen and A. Svenson. Dynamics of Crystal Formation in the Greenland NorthGRIP Ice Core. J. Glaciology {\bf 50}, 325 (2004)
10. M. Ander, P. Beltrao, B. Di Ventura, J. Ferkinghoff-Borg, M. Foglierine, A. Kaplan, C. Lemerle, I. Tomas-Oliveira and L. Serrano. SmartCell: a framework to simulate cellular processes that combines stochastic approximation with diffusion and localization: analysis of simple gene networks.  Systems Biology 1: 129 (2004).
11. J. Ferkinghoff-Borg, M.H. Jensen, J. Mathiesen, P. Olesen and K. Sneppen.  Competition between Diffusion and Fragmentation: An Important Evolutionary Process of Nature. Phys. Rev. Lett. {\bf 91}, 266103 (2003).
12. M.S. Johansen, P. Alstr\o m, J. Borg and M.T. Levinsen, Corrective measures in turbulent pipe flows and extended self-similarity. Eur. Phys. J. B. {\bf 11} 665-676 (1999).

Monte Carlo techniques for inferences in high-dimensional models: bio-structures, data-processing and deep-learning.

Typically, statistical inference in high-dimensional models or in large data-set of heterogeneous and heteroscedastic nature is notoriously difficult. Probabilistic representation of  

the data generated in LindingLab, from exome sequencing,  global mass-spectrometry to HCS-screening, is on the other hand of central importance for rigorous data-integration, validation and biological forecasting. The systematic validation and comparison of these global models involves a high-dimensional inference problem, which formally amounts to estimating and comparing a multidimensional integral over the degrees of freedom, the so-called partition function of the model. We have developed new Monte Carlo techniques (Munnin and BayesGE) to make high-dimensional inferences tractable, and successfully applied these to problems in respectively protein structure determination and dynamics, analysis of global mass-spectrometry data, large-scale pharmacological screens (see also preceeding sections and section below) and to evaluation of partition functions in deep-learning models.

Publications

1. Frellsen J., Winther O., Ghahramani Z. and Ferkinghoff-Borg J.,  Bayesian generalised ensemble Markov chain Monte Carlo. AISTAT 2016, The 19th International Conference on Artificial Intelligence and Statistics, accepted for publication (2016).
2. Boomsma, W., Frellsen, J., Harder, T., Bottaro, S., Johansson, KE., Tian, P., Stovgaard, K., Andreetta, C., Olsson, S., Valentin, J., Antonov, L., Christensen, A., Borg, M., Jensen, J., Lindorff-Larsen, K., Ferkinghoff-Borg, J., Hamelryck, T.  A framework for Markov chain Monte Carlo simulation and inference of protein structure. J. Comput. Chem. 34, 1697-1705 (2013).
3. Bottaro, S., Boomsma, W., Johansson, K.E., Andreetta, C., Hamelryck, T., Ferkinghoff-Borg. Subtle Monte Carlo updates in dense molecular systems. J. Chem. Theory Comput. (2012), 8, 695–702.
4. Ferkinghoff-Borg J. Monte Carlo Methods for Inferences in High-dimensional systems. In T. Hamelryck, K. Mardia and J. Ferkinghoff-Borg (eds). Bayesian methods in structural bioinformatics. Statistics for Biology and Health. Springer-Verlag, Berlin, Heidelberg. (2012).
5. P. M. Hansen, J. K Dreyer, J. Ferkinghoff-Borg and Lene Oddershede. Novel optical and statistical methods reveal colloid-wall interactions inconsistent with DLVO and Lifshitz theories. J. Colloid Interface Sci. 287, 561 (2005).
6. J. Ferkinghoff-Borg, Optimized Monte Carlo Analysis for Generalized Ensembles. Eur. Phys. J. B. 29, 481-484 (2002).

Bayesian framework for large-scale dose-response analysis

We are interested in evolving drug development for cancer and other complex diseases from targeting genes to systems-level interventions. For this it is essential to extend existing pharmacological methodologies in order to deal more effectively with data sparsity and uncertainties in large-scale combinatorial screens. To this end we have developed a novel Bayesian dose-response framework to analyse a large-scale scanning of a 100x100  oncological drug matrix in melanoma and pancreatic cancer. We have identified a set of strain-specific small molecule combinations with high-confidence temporal synergy, which we show have capabilities to lower tumor burden in vivo. Our approach can likely be extended to enable systematic discovery of complex treatment strategies for cancer and other diseases.