Colloquia and Seminars

Nanopore sequencing is emerging as a powerful tool for storing digital information in DNA molecules. This technique offers several advantages over traditional methods, making it an attractive area of research. In this work, we focus on a simplified model of the nanopore sequencing process, represented as a channel.

This channel operates by taking a DNA sequence and analyzing it one segment at a time. It uses a sliding window of a specific length, denoted by ℓ, to scan the sequence. The window is then shifted by δ characters, and the process repeats. The output of the channel called the read vector, is a collection of values where each value represents the sum of the bases (like A, C, G, T) within a particular window.

The channel’s capacity signifies the maximum rate of information transmission through it. Prior research has established capacity values for specific combinations of ℓ and δ. In this study, we delve deeper, demonstrating that when δ < ℓ < 2δ, the channel’s capacity can be expressed as (1/δ) log (1/2 (ℓ + 1 +√((ℓ + 1)2 − 4(ℓ − δ)(ℓ − δ + 1))). Furthermore, we establish an upper bound on the capacity when 2δ is less than ℓ. Finally, to enhance the model’s complexity, we extendit to a two-dimensional scenario and present various findings on its capacity. This extended model brings us closer to mimicking the real intricacies of nanopore sequencing and its potential for DNA storage

OMICS-based screening offers a promising approach for untargeted drug discovery. Mass-spectrometry metabolomics and proteomics were recently used for inferring drug mechanisms of action and off-target effects, analyzing the cellular response to treatment with numerous clinically approved drugs and tool compounds. In this talk, we present a novel high-throughput LC-MS metabolomics screening pipeline and a deep-learning method specifically designed for cellular metabolism.

Our novel Graph Neural Network (GNN) model fits the unique structure of this domain, enabling accurate identification of target pathways. Learning from time- and dose-dependent metabolic responses of cultured cells treated with 76 known metabolic inhibitors, we correctly identified the target pathway within the top three ranked pathways for ~50% of the drugs. Applying this approach to a diversity library of 1,020 drug-like compounds, we discovered five novel inhibitors targeting clinically relevant pathways and enzymes involved in purine and pyrimidine biosynthesis and redox metabolism. Our pipeline is readily scalable for screening thousands of compounds to identify new, clinically relevant metabolic inhibitors.

We will dive deep into our machine learning approach, offering valuable insights and comprehensive ablation studies that highlight its strength and domain-specific innovation.

Despite remarkable advances, visual generative models are still far from faithfully modeling the world, struggling with fundamental aspects such as spatial relations, physics, motion, and dynamic interactions.

In this talk, I present a line of work that tackles these challenges, based on a deep understanding of the inner mechanisms that drive models. I will begin by analyzing state-of-the-art visual generators, gaining insights into the underlying reasons for their limited understanding. Building upon these insights, I will demonstrate methods that significantly enhance both spatial and temporal reasoning in image and video generation, surpassing even resource-intensive proprietary models without relying on additional data or model scaling. I will conclude the talk by discussing open challenges and future directions for advancing faithful world modeling in visual generative models.

Bio:
Hila is a PhD candidate at Tel Aviv University, advised by Prof. Lior Wolf. Her research focuses on understanding, interpreting, and correcting the predictions of deep foundational models. During her PhD, she was a visiting researcher at Google Research, Google DeepMind, and Meta AI, where she led works on video generation.

Hila has received several awards, including the Fulbright Postdoctoral Fellowship, the Eric and Wendy Schmidt Postdoctoral Award, the Deutsch Prize for Outstanding PhD Students, and the Council for Higher Education (VATAT) Award for Outstanding PhD Students.

In this talk, I present three research directions from my PhD.

1. Geometric Learning for Structured Data. Firstly, we introduce a simple, spectral-geometric approach for matrix completion on graphs. Our approach couples the priors implicitly induced by gradient descent with explicitly imposed spectral-geometric priors and achieves strong performance in drug-target interaction prediction and recommendation systems applications. Secondly, we introduce inductive/generalizable solvers for quadratic optimal transport problems, demonstrating superior scalability over transductive methods, with applications in single-cell multi-omic alignment. Finally, on the practical front, we show algorithms for efficient execution of graph machine learning workloads for large-scale recommendation system inference.

2. Statistical Inference via Optimal Transport. In this line of work, we develop new solvers and extensions for vector quantile regression (VQR), an optimal-transport–based statistical framework introduced by Carlier et al. 2016. Firstly, we introduce nonlinear VQR, the first practical approach to model quantile functions of multivariate conditional distributions, and demonstrate applications in creating calibrated high-dimensional confidence sets. Secondly, we propose manifold VQR, and extend the notion of conditional quantile functions for manifold-valued response variables. Finally, we introduce continuous solvers for VQR by solving conditional continuous OT problems. We perform fundamental statistical inference tasks on conditional distributions, i.e., sampling, computing likelihoods, constructing confidence sets, computing order statistics (ranks, quantiles, etc.) from the derived OT maps.

3. Modeling Biomolecules. In this research direction, we make advances in overcoming the "single-sequence, single-structure" dogma in structural biology, and highlight fundamental limitations of protein structure predictors such as AlphaFold. Firstly, we investigate protein structures that exhibit dual conformations in a single crystal, called "altlocs", and identify "stable altlocs", the altloc dual conformations that are thermodynamically stable. We demonstrate that the state-of-the-art protein structure predictors and protein structure generative models fail to recover these dual conformations. We introduce a guided diffusion framework to improve the modeling of biomolecules conditioned on experimental measurements. Secondly, we demonstrate limitations of protein structure prediction algorithms for modeling chimeric proteins. To overcome this, we introduce windowed multiple sequence alignment (MSA), for enriching the MSA of the chimeric proteins to yield improved predictions. 

Submodular functions arise in various fields, including combinatorics, graph theory, information theory, and economics. This thesis addresses two key problems in submodular maximization and presents new algorithmic contributions. The first focuses on the Online Submodular Welfare problem, where bidders with submodular utility functions compete for items arriving sequentially. We propose a randomized algorithm that achieves a tight competitive ratio of 1/4 under adversarial arrivals and improve this to approximately 0.27493 under random arrivals. The second result introduces a Greedy Poisson Process for submodular maximization under matroid constraints, achieving a tight (1 − 1/e − ϵ)-approximation without requiring discretization or rounding. Our algorithm introduces a new framework for submodular maximization under matroid constraints, combining the strengths of both continuous and discrete approaches.

Nonvolatile Memory (NVM) technologies offer new avenues for building high-performance and crash-consistent applications by combining byte-addressable DRAM-like characteristics with non-volatility. However, these features introduce complex challenges in ensuring data consistency, especially under concurrent access scenarios.

This paper introduces PRD, a specialized tool designed to detect persistency races - specific type of concurrency bugs in PM environments that can lead to critical inconsistencies following system failures.

PRD utilizes graph-theoretical analysis along with happens-before and program-dependence analysis to map causal relationships and dependencies among program operations. While these analyses are well-established techniques in data-race detection, PRD efficiently generalizes across multiple thread interleavings to detect potential races within a single program execution, significantly enhancing the speed and efficiency of race detection in NVM environments.