Columbia University in the City of New York

Tavazoie Lab

Laboratory of Systems Biology


Despite our rapidly improving ability to probe molecular and cellular states, our understanding of biology remains largely descriptive and often overwhelmed by details of unclear significance. Our research is driven by the desire to move beyond the vast scale of these detailed descriptions to discover the underlying organizing principles. These principles reflect fundamental solutions to common challenges revealed by natural selection across billions of years of evolution. In essence, these principles constitute a compact set of concepts and algorithms that enable us to broadly generalize our understanding of biological phenomena, to efficiently manipulate them for medical and industrial purposes, and to engineer biological systems from first principles. As such, our research is focused on systems-level phenomena independent of conventionally defined processes and pathways. We employ diverse experimental systems from bacteria to mammalian cell-lines in order to study general principles that operate across organismal taxa and complexity. Our systems-level studies often necessitate observations, perturbations, or analyses that are beyond the scale and precision of existing methods. We thus develop new enabling technologies and computational methods that significantly advance our ability to study systems-level phenomena.    



Our vision of systems biology originated when the very first microarray gene expression datasets appeared in the late 1990’s. At the time, we wondered whether such large-scale observations, coupled with properly constrained machine-learning, would enable us to discover the inherent organization of gene regulatory networks. Our work demonstrated that, indeed, this strategy can systematically reveal the regulatory vocabulary and grammar in DNA, enabling prediction of global gene expression dynamics from DNA sequence alone. The ability to rapidly rediscover gene regulatory programs that had taken decades to uncover demonstrated the power of this ‘reverse engineering’ paradigm for rapidly advancing biology. Over the years, we have established the unique efficacy of this paradigm across diverse domains of biology including transcriptional & post-transcriptional regulation, chromatin modification dynamics, molecular interaction mapping, fitness landscape characterization, genetic interaction mapping, and genotype-phenotype inference. These studies have revealed fundamental new insights into the genetic and regulatory underpinnings of a variety of phenomena including complex bacterial behaviors, antibiotic resistance & persistence, metazoan development, oncogenesis, metastatic progression, and cellular adaptation to extreme environments. Many of the approaches and technologies we have introduced have made broad impact through their adoption across diverse areas of biology. 

Gene regulation            

Context-dependent regulation of gene expression is fundamental to all cellular behaviors. The expression of a gene is shaped by the convergence of upstream inputs that impinge upon DNA and RNA sequence elements in the vicinity of genes, leading to precise modulation of global mRNA and protein abundances. The richness of gene expression programs—in a given cell across time, across distinct cell types, and in response to diverse stimuli—results from the combinatorial logic of spatially organized nucleic acid elements that bind transcription factors, RNA binding proteins, and microRNAs. The identification of these regulatory elements and elucidation of the rules by which they operate remains a central challenge for modern biology. With the arrival of the very first microarray expression experiments, we reasoned that the massive scale of the data may enable an entirely agnostic approach to discovering the underlying regulatory logic of gene expression. We thus applied unsupervised machine learning approaches to first discover the intrinsic patterns of gene expression (clusters), and then used ab initio motif discovery to reveal highly enriched DNA-sequence motifs, representing transcription factor binding sites through which gene expression is modulated. We showed that this agnostic strategy can systematically reveal the known transcriptional regulatory architecture of the yeast cell-cycle and predict novel regulators that have since been discovered by us and others (Tavazoie et al., Nature Genetics, 1999). We also introduced pathway enrichment analysis that has since become a critical tool for interpreting the biological significance of genome-wide studies.

Motivated by the desire to build predictive models of gene expression, we used supervised machine learning approaches to learn the combinatorial grammar in regulatory regions, revealing the need for precise spatial configuration of multiple transcription factor binding sites. We showed that, in yeast, these models can achieve surprisingly high performance, enabling prediction of global gene expression dynamics directly from DNA sequence alone (Beer & Tavazoie, Cell, 2004). We developed next-generation reverse engineering approaches based on information-theory that reveal transcriptional and post-transcriptional regulatory elements with high sensitivity and specificity in organisms ranging from bacteria to human (Elemento et al., Molecular Cell, 2007; Goodarzi et al., Molecular Cell, 2009). The incorporation of complex representations of sequence and structure allowed us to discover a large regulatory landscape of post-transcriptional regulation by RNA-structural elements in 5’ and 3’ non-coding regions of mRNAs (Goodarzi et al., Nature, 2012). We showed that these regulatory elements and the RNA binding proteins that bind them regulate critical pathways in physiology and disease, including cancer metastasis (Goodarzi et al., Nature, 2014). We are currently using these computational tools, together with CRISPR-based functional profiling technologies, to map the post-transcriptional regulatory landscape of oncogenesis and cancer metastasis. Another area of interest is the application of these approaches to decode the transcriptional and post-transcriptional regulatory networks critical for differentiation and development, including those that establish identity and function in the nervous system (Taylor et al., Cell, 2021). Modern deep learning architectures are an increasingly important tool for building predictive models that reveal the key regulatory parameters in the central dogma including those modulating transcription, mRNA stability, translation, post-translational modifications, localization, and protein stability.

Cellular adaptation

A major focus of our work is to understand how cells adapt to changes in their external environment. We study this inherently systems-level phenomenon across a range of timescales, from rapid transcriptional responses, to multi-generational epigenetic reprogramming, to long-term rewiring of signaling and regulatory networks over evolutionary timescales. Much of our contributions have resulted from questioning long prevailing dogma in the field. By posing problems as open questions, the agnostic nature of our approach allows the system itself to reveal the essential governing principles. This has been crucial to discovering surprising new phenomena such as the ability of microbes to carry out predictive behavior akin to metazoan nervous systems (Tagkopoulos et al. Science, 2008). Often these higher-level principles are obscured by approaches that focus only on a narrow slice of the cell's response.

By challenging fundamental notions of cellular adaptation, we have found that eukaryotic cells utilize a noise-driven optimization mechanism to reprogram gene expression in a manner that is independent of conventional gene regulation. In this process that we call stochastic tuning, cells utilize the inherent noise in mRNA transcription to randomly increase or decrease expression of genes and to actively reinforce changes that improve the overall health of the cell. We have shown that this mechanism enables cells to empirically establish adaptive gene expression levels in the absence of conventional hard-wired regulatory input (Freddolino et al., eLife 2018). A such, stochastic tuning is an active form of gene expression optimization that occurs in individual cells—a phenomenon distinct from population-level bet-hedging. We have shown that stochastic tuning occurs in budding yeast, enabling cells to adapt to unfamiliar environments where their conventional regulatory systems are challenged beyond their physiological operating capacity. Since the pioneering work of Jacob and Monod sixty years ago, we have been taught that cellular responses are determined by pre-defined regulatory programs established by specific genetically encoded pathways (e.g. the osmolarity response pathway). The fact that individual cells can, instead, utilize a noise-driven trial & error process to empirically establish gene expression levels goes against our most cherished notions of gene regulation and cellular adaptation. Beyond its role in adaptation of single-cell eukaryotes, stochastic tuning may be the underlying mechanism for observed non-mutational tumor adaptation, increasingly recognized as a key factor in chemotherapy failure. Identifying the molecular components of stochastic tuning, discovering the detailed underlying mechanisms, and exploring the breadth of its pathophysiological significance are major areas of focus for current and future work in the lab.

Molecular interactions

Molecular interactions are the fundamental building blocks of life. Almost every biological process can be understood in terms of specific interactions between key biomolecules such as proteins, DNA, RNA, and metabolites. Mapping and understanding the astronomically complex network of molecular interactions among the millions of distinct components in the cell remains a grand challenge with far-reaching implications across all areas of biology. For example, if we focus on the human proteome (even ignoring differential splice variants), there are some ~109 possible pair-wise interactions. This number grows to ~1012 for all potential human DNA-protein interactions at 15 bp resolution. In a perfect world, a technology would allow us to efficiently, systematically, and quantitatively, measure all these potential interactions, providing a global unbiased architectural view. The advent of yeast two-hybrid technology (Y2H) more than two decades ago, initiated an effort to move in this direction and indeed the emerging maps have formed a critical foundation for modern biology. However, the sensitivity, cost, and labor-intensive nature of Y2H and many alternative variants do not permit routine mapping and monitoring of global molecular interactions as a function of varying molecular and cellular states in vivo. We are developing next-generation technologies to address these limitations.

A major focus has been to monitor in vivo protein-DNA interactions as a function of changing conditions. We have developed a biochemical procedure to efficiently isolate protein bound sites throughout a bacterial genome and to quantify their occupancy using next-generation sequencing. This approach, called IPOD (In vivo Protein Occupancy Display) enables comprehensive global monitoring of protein-DNA interactions as a function of genetic and environmental perturbations (Vora et al., Molecular Cell, 2009). These dynamic global maps can be readily utilized to identify specific genes and regulatory regions that are dynamically activated as a function of any perturbation, revealing the vast regulatory landscape that is obscured when the genome is only interrogated through single ChIP-seq experiments (Freddolino et al., PLoS Biology, 2021). The ability to globally monitor all such sites, without prior bias, has revealed novel architectural features of the genome. For example, we have found that the E. coli chromosome contains hundreds of kilo-base scale regions, bound by nucleoid proteins, that, through their transcriptional silencing effect, appear to function as prokaryotic analogs of eukaryotic heterochromatin.

For decades, the dominant approach to quantifying protein abundance and interactions has been mass-spectrometry (MS). However, MS-based approaches lack the affordability, scalability, and standardization required for routine comprehensive profiling of proteomes at a scale required for advances we envision. With support from an NIH Transformative award, we have developed an alternative approach to proteomics based on attachment of proteins to their encoding mRNA sequences in vivo. This technology, we call In vivo mRNA-display enables us to query abundance, interactions, and localization of proteins through identification of their accompanying nucleic-acid tags (Oikonomou et al., PNAS, 2020). In this way, we recast proteomic analysis as a DNA sequencing problem, bypassing mass-spectrometry bottlenecks, and setting proteomics on the path to benefit from exponentially improving cost and throughput of next-generation sequencing. Furthermore, by preserving the intrinsic post-translational modification states of proteins, in vivo mRNA-display enables more accurate representation of native proteomes, critical for precise inference of protein function and interactions. We are using in vivo mRNA display as the platform to scale a variety of proteomic technologies to enable routine generation of dynamic global maps of protein function and interactions. We are also developing in vivo mRNA display as a versatile foundation for protein engineering and synthetic biology applications.

Areas of active investigation include:

-Cellular adaptation through stochastic tuning of gene expression

-Genetic basis of antibiotic sensitivity, resistance, and persistence

-Microbial adaptation to extreme environments

-Predictive models of transcriptional, post-transcriptional, and translational regulatory networks

-Reverse engineering regulatory networks in cancer and metastasis 

-Regulatory networks of cellular and organismal aging

-DNA-sequencing based next-generation proteomics technologies

-Deep mapping of molecular interactions in networks of proteins, DNA, RNA, and metabolites

-CRISPR-based functional genomic technologies

-Microbiome functional genomics

-Synthetic transcriptomic and proteomic technologies for cellular engineering

Join our team:

We are looking for creative and ambitious people to join our highly interdisciplinary and interactive group. Please send inquiries to: st2744 [at]

-© 2022 Columbia University