arXiv subjects

Michael M. Danziger

Publications and source records attributed to Michael M. Danziger.

15 recordsLinked to original sources

arXiv2026

BMFM-RNA: whole-cell expression decoding improves transcriptomic foundation models

Transcriptomic foundation models pretrained with masked language modeling can achieve low pretraining loss yet produce poor cell representations for downstream tasks. We introduce whole-cell expression decoding (WCED), where models reconstruct the entire gene vocabulary from a single CLS token embedding, even with limited inputs, creating a maximally informative bottleneck. WCED consistently outperforms MLM on all downstream metrics despite higher reconstruction error during training. Gene-level error tracking reveals that both methods preferentially learn genes whose expression co-varies with stable transcriptional programs rather than those driven by transient factors. We further add hierarchical cross-entropy loss that exploits Cell Ontology structure for zero-shot annotation at multiple granularity levels. Models trained with these objectives achieve best overall performance across CZI benchmarks, on zero-shot batch integration and linear probing cell-type annotation. Methods are implemented in biomed-multi-omic ( https://github.com/BiomedSciAI/biomed-multi-omic ), an open-source framework for transcriptomic foundation model development.

q-bio.GN↗

arXiv2025

MAMMAL -- Molecular Aligned Multi-Modal Architecture and Language

Large language models applied to vast biological datasets have the potential to transform biology by uncovering disease mechanisms and accelerating drug development. However, current models are often siloed, trained separately on small-molecules, proteins, or transcriptomic data, limiting their ability to capture complex, multi-modal interactions. Effective drug discovery requires computational tools that integrate multiple biological entities while supporting prediction and generation, a challenge existing models struggle to address. For this purpose, we present MAMMAL - Molecular Aligned Multi-Modal Architecture and Language - a versatile method applied to create a multi-task foundation model that learns from large-scale biological datasets across diverse modalities, including proteins, small-molecules, and omics. MAMMAL's structured prompt syntax supports classification, regression, and generation tasks while handling token and scalar inputs and outputs. Evaluated on eleven diverse downstream tasks, it reaches a new state of the art (SOTA) in nine tasks and is comparable to SOTA in two tasks, all within a unified architecture, unlike prior task-specific models. Additionally, we explored Alphafold 3 binding prediction capabilities on antibody-antigen and nanobody-antigen complexes showing significantly better classification performance of MAMMAL in 3 out of 4 targets. The model code and pretrained weights are publicly available at https://github.com/BiomedSciAI/biomed-multi-alignment and https://huggingface.co/ibm/biomed.omics.bl.sm.ma-ted-458m

q-bio.QM↗

arXiv2023

Mastering Percolation-like Games with Deep Learning

Though robustness of networks to random attacks has been widely studied, intentional destruction by an intelligent agent is not tractable with previous methods. Here we devise a single-player game on a lattice that mimics the logic of an attacker attempting to destroy a network. The objective of the game is to disable all nodes in the fewest number of steps. We develop a reinforcement learning approach using deep Q-learning that is capable of learning to play this game successfully, and in so doing, to optimally attack a network. Because the learning algorithm is universal, we train agents on different definitions of robustness and compare the learned strategies. We find that superficially similar definitions of robustness induce different strategies in the trained agent, implying that optimally attacking or defending a network is sensitive the particular objective. Our method provides a new approach to understand network robustness, with potential applications to other discrete processes in disordered systems.

cs.LG↗

arXiv2020

Recovery Coupling in Multilayer Networks

The increased complexity of infrastructure systems has resulted in critical interdependencies between multiple networks---communication systems require electricity, while the normal functioning of the power grid relies on communication systems. These interdependencies have inspired an extensive literature on coupled multilayer networks, assuming that a component failure in one network causes failures in the other network, a hard interdependence that results in a cascade of failures across multiple systems. While empirical evidence of such hard coupling is limited, the repair and recovery of a network requires resources typically supplied by other networks, resulting in well documented interdependencies induced by the recovery process. If the support networks are not functional, recovery will be slowed. Here we collected data on the recovery time of millions of power grid failures, finding evidence of universal nonlinear behavior in recovery following large perturbations. We develop a theoretical framework to address recovery coupling, predicting quantitative signatures different from the multilayer cascading failures. We then rely on controlled natural experiments to separate the role of recovery coupling from other effects like resource limitations, offering direct evidence of how recovery coupling affects a system's functionality. The resulting insights have implications beyond infrastructure systems, offering insights on the fragility and senescence of biological systems.

physics.soc-ph↗

arXiv2020

High-resolution human mobility data reveal race and wealth disparities in disaster evacuation patterns

Major disasters such as extreme weather events can magnify and exacerbate pre-existing social disparities, with disadvantaged populations bearing disproportionate costs. Despite the implications for equity and emergency planning, we lack a quantitative understanding of how these social fault lines translate to different behaviors in large-scale emergency contexts. Here we investigate this problem in the context of Hurricane Harvey, using over 30 million anonymized GPS records from over 150,000 opted-in users in the Greater Houston Area to quantify patterns of disaster-inflicted relocation activities before, during, and after the shock. We show that evacuation distance is highly homogenous across individuals from different types of neighborhoods classified by race and wealth, obeying a truncated power-law distribution. Yet here the similarities end: we find that both race and wealth strongly impact evacuation patterns, with disadvantaged minority populations less likely to evacuate than wealthier white residents. Finally, there are considerable discrepancies in terms of departure and return times by race and wealth, with strong social cohesion among evacuees from advantaged neighborhoods in their destination choices. These empirical findings bring new insights into mobility and evacuations, providing policy recommendations for residents, decision makers, and disaster managers alike.

physics.soc-ph↗

arXiv2019

Critical stretching of mean-field regimes in spatial networks

We study a spatial network model with exponentially distributed link-lengths on an underlying grid of points, undergoing a structural crossover from a random, Erdős--Rényi graph to a $2D$ lattice at the characteristic interaction range $ζ$. We find that, whilst far from the percolation threshold the random part of the incipient cluster scales linearly with $ζ$, close to criticality it extends in space until the universal length scale $ζ^{3/2}$ before crossing over to the spatial one. We demonstrate this {\em critical stretching} phenomenon in percolation and in dynamical processes, and we discuss its implications to real-world phenomena, such as neural activation, traffic flows or epidemic spreading.

physics.soc-ph↗

arXiv2019

Faster calculation of the percolation correlation length on spatial networks

The divergence of the correlation length $ξ$ at criticality is an important phenomenon of percolation in two-dimensional systems. Substantial speed-ups to the calculation of the percolation threshold and component distribution have been achieved by utilizing disjoint sets, but existing algorithms of this sort cannot measure the correlation length. Here, we utilize the parallel axis theorem to track the correlation length as nodes are added to the system, allowing us to utilize disjoint sets to measure $ξ$ for the entire percolation process with arbitrary precision in a single sweep. This algorithm enables direct measurement of the correlation length in lattices as well as spatial network topologies, and provides an important tool for understanding critical phenomena in spatial systems.

physics.soc-ph↗

arXiv2017

Critical field-exponents for secure message-passing in modular networks

We study secure message-passing in the presence of multiple adversaries in modular networks. We assume a dominant fraction of nodes in each module have the same vulnerability, i.e., the same entity spying on them. We find both analytically and via simulations that the links between the modules (interlinks) have effects analogous to a magnetic field in a spin system in that for any amount of interlinks the system no longer undergoes a phase transition. We then define the exponents $δ$, which relates the order parameter (the size of the giant secure component) at the critical point to the field strength (average number of interlinks per node), and $γ$, which describes the susceptibility near criticality. These are found to be $δ=2$ and $γ=1$ (with the scaling of the order parameter near the critical point given by $β=1$). When two or more vulnerabilities are equally present in a module we find $δ=1$ and $γ=0$ (with $β\geq2$). Apart from defining a previously unidentified universality class, these exponents show that increasing connections between modules is more beneficial for security than increasing connections within modules. We also measure the correlation critical exponent $ν$, and the upper critical dimension $d_c$, finding that $νd_c=3$ as for ordinary percolation, suggesting that for secure message-passing $d_c =6$. These results provide an interesting analogy between secure message-passing in modular networks and the physics of magnetic spin-systems.

physics.soc-ph↗

arXiv2017

Dynamic interdependence and competition in multilayer networks

From critical infrastructure, to physiology and the human brain, complex systems rarely occur in isolation. Instead, the functioning of nodes in one system often promotes or suppresses the functioning of nodes in another. Despite advances in structural interdependence, modeling interdependence and other interactions between dynamic systems has proven elusive. Here we define a broadly applicable dynamic dependency link and develop a general framework for interdependent and competitive interactions between general dynamic systems. We apply our framework to studying interdependent and competitive synchronization in multi-layer oscillator networks and cooperative/competitive contagions in an epidemic model. Using a mean-field theory which we verify numerically, we find explosive transitions and rich behavior which is absent in percolation models including hysteresis, multi-stability and chaos. The framework presented here provides a powerful new way to model and understand many of the interacting complex systems which surround us.

cond-mat.stat-mech↗

arXiv2017

Spreading of localized attacks in spatial multiplex networks

Many real-world multilayer systems such as critical infrastructure are interdependent and embedded in space with links of a characteristic length. They are also vulnerable to localized attacks or failures, such as terrorist attacks or natural catastrophes, which affect all nodes within a given radius. Here we study the effects of localized attacks on spatial multiplex networks of two layers. We find a metastable region where a localized attack larger than a critical size induces a nucleation transition as a cascade of failures spreads throughout the system, leading to its collapse. We develop a theory to predict the critical attack size and find that it exhibits novel scaling behavior. We further find that localized attacks in these multiplex systems can induce a previously unobserved combination of random and spatial cascades. Our results demonstrate important vulnerabilities in real-world interdependent networks and show new theoretical features of spatial networks.

physics.soc-ph↗

arXiv2016

Color-avoiding percolation

Many real world networks have groups of similar nodes which are vulnerable to the same failure or adversary. Nodes can be colored in such a way that colors encode the shared vulnerabilities. Using multiple paths to avoid these vulnerabilities can greatly improve network robustness. Color-avoiding percolation provides a theoretical framework for analyzing this scenario, focusing on the maximal set of nodes which can be connected via multiple color-avoiding paths. In this paper we extend the basic theory of color-avoiding percolation that was published in [Krause et. al., Phys. Rev. X 6 (2016) 041022]. We explicitly account for the fact that the same particular link can be part of different paths avoiding different colors. This fact was previously accounted for with a heuristic approximation. We compare this approximation with a new, more exact theory and show that the new theory is substantially more accurate for many avoided colors. Further, we formulate our new theory with differentiated node functions, as senders/receivers or as transmitters. In both functions, nodes can be explicitly trusted or avoided. With only one avoided color we obtain standard percolation. With one by one avoiding additional colors, we can understand the critical behavior of color avoiding percolation. For heterogeneous color frequencies, we find that the colors with the largest frequencies control the critical threshold and exponent. Colors of small frequencies have only a minor influence on color avoiding connectivity, thus allowing for approximations.

physics.soc-ph↗

arXiv2015

Two distinct transitions in spatially embedded multiplex networks

Multilayer infrastructure is often interdependent, with nodes in one layer depending on nearby nodes in another layer to function. The links in each layer are often of limited length, due to the construction cost of longer links. Here, we model such systems as a multiplex network composed of two or more layers, each with links of characteristic geographic length, embedded in 2-dimensional space. This is equivalent to a system of interdependent spatially embedded networks in two dimensions in which the connectivity links are constrained in length but varied while the length of the dependency links is always zero. We find two distinct percolation transition behaviors depending on the characteristic length, $ζ$, of the links. When $ζ$ is longer than a certain critical value, $ζ_c$, abrupt, first-order transitions take place, while for $ζ<ζ_c$ the transition is continuous. We show that, though in single-layer networks increasing $ζ$ decreases the percolation threshold $p_c$, in multiplex networks it has the opposite effect: increasing $p_c$ to a maximum at $ζ=ζ_c$. By providing a more realistic topological model for spatially embedded interdependent and multiplex networks and highlighting its similarities to lattice-based models, we provide a new direction for more detailed future studies.

physics.soc-ph↗

arXiv2015

Optimal redundancy against disjoint vulnerabilities in networks

Redundancy is commonly used to guarantee continued functionality in networked systems. However, often many nodes are vulnerable to the same failure or adversary. A "backup" path is not sufficient if both paths depend on nodes which share a vulnerability.For example, if two nodes of the Internet cannot be connected without using routers belonging to a given untrusted entity, then all of their communication-regardless of the specific paths utilized-will be intercepted by the controlling entity.In this and many other cases, the vulnerabilities affecting the network are disjoint: each node has exactly one vulnerability but the same vulnerability can affect many nodes. To discover optimal redundancy in this scenario, we describe each vulnerability as a color and develop a "color-avoiding percolation" which uncovers a hidden color-avoiding connectivity. We present algorithms for color-avoiding percolation of general networks and an analytic theory for random graphs with uniformly distributed colors including critical phenomena. We demonstrate our theory by uncovering the hidden color-avoiding connectivity of the Internet. We find that less well-connected countries are more likely able to communicate securely through optimally redundant paths than highly connected countries like the US. Our results reveal a new layer of hidden structure in complex systems and can enhance security and robustness through optimal redundancy in a wide range of systems including biological, economic and communications networks.

physics.soc-ph↗

arXiv2014

Robustness of a Network Formed of Spatially Embedded Networks

We present analytic and numeric results for percolation in a network formed of interdependent spatially embedded networks. We show results for a treelike and a random regular network of networks each with $(i)$ unconstrained interdependent links and $(ii)$ interdependent links restricted to a maximum length, $r$. Analytic results are given for each network of networks with unconstrained dependency links and compared with simulations. For the case of two spatially embedded networks it was found that only for $r>r_c\approx8$ does the system undergo a first order phase transition. We find that for treelike networks of networks $r_c$ significantly decreases as $n$ increases and rapidly reaches its limiting value, $r=1$. For cases where the dependencies form loops, such as in random regular networks, we show analytically and confirm through simulations, that there is a certain fraction of dependent nodes, $q_{max}$, above which the entire network structure collapses even if a single node is removed. This $q_{max}$ decreases quickly with $m$, the degree of the random regular network of networks. Our results show the extreme sensitivity of coupled spatial networks and emphasize the susceptibility of these networks to sudden collapse. The theory derived here can be used to find the robustness of any network of networks where the profile of percolation of a single network is known.

physics.soc-ph↗

arXiv2013

Spatially localized attacks on interdependent networks: the existence of a finite critical attack size

Many real world complex systems such as infrastructure, communication and transportation networks are embedded in space, where entities of one system may depend on entities of other systems. These systems are subject to geographically localized failures due to malicious attacks or natural disasters. Here we study the resilience of a system composed of two interdependent spatially embedded networks to localized geographical attacks. We find that if an attack is larger than a finite (zero fraction of the system) critical size, it will spread through the entire system and lead to its complete collapse. If the attack is below the critical size, it will remain localized. In contrast, under random attack a finite fraction of the system needs to be removed to initiate system collapse. We present both numerical simulations and a theoretical approach to analyze and predict the effect of local attacks and the critical attack size. Our results demonstrate the high risk of local attacks on interdependent spatially embedded infrastructures and can be useful for designing more resilient systems.

physics.soc-ph↗