2025-2026

Department of Chemistry, Massachusetts Institute of Technology

Fragment embedding methods have a long history in chemistry. At a basic level, embedding methods divide a large system into many smaller fragments that can be simulated individually and then combined to make predictions about the whole system. In the first part of the talk, we will discuss recent advances in bootstrap embedding (BE) that allow the description of overlapping fragments. Using appropriately-sized fragments, BE is able to describe molecules and solids with complex unit cells containing hundreds of atoms while recovering ~99.9% of the correlation energy. We will discuss applications of the method to describe conducting polymers, 2D-materials and protein-ligand binding.

In the second part of the talk, we will discuss potential applications of quantum computing to chemistry. We will discuss the concrete reasons we should expect that quantum hardware will significantly impact the way computational chemistry operates in the longer term. We will explore the most likely near term applications (some of which involve fragment embedding ideas). We will finally discuss the ultimate limitations of quantum computing – things that one will not ever be able to do with quantum hardware. In particular, we will discuss a recent proof of a “no downfolding” theorem that implies significant limitations for both quantum computing and computational quantum chemistry.

Department of Chemistry and Chemical Biology, Northeastern University

Photoenzymes are emerging protein-based photocatalysts that are repurposed from natural enzymes for non-natural reactions difficult for small-molecule catalysts. They exhibit extraordinary selectivity, scalability, and tunability, and offer a promising new toolbox for solar to chemical energy conversion and chemical synthesis. However, the understanding and design of photoenzymes pose several challenges. First, accurate first-principles simulations of the electronic structure of macromolecules are usually computationally expensive, especially those that involve strong electron correlation. In this talk, I will discuss our computational strategies, including data-driven methods and quantum computing, particularly quantum annealing, to tackle this challenge. Second, existing enzyme design strategies do not consider electronic excited states, and photoenzyme engineering has mainly relied on directed evolution. I will discuss our work on physics-informed computational photoenzyme design, where we combine physics-based simulations and data-driven methods to elucidate the design strategies for photoenzymes and other macromolecular photocatalysts.

Department of Chemistry, Boston College

Accurate predictions of molecular structures and reaction pathways often rely on expensive ab initio electronic-structure calculations, with costs that grow steeply in ab initio molecular dynamics (AIMD). While neural-network potential energy surfaces (PESs) offer relief, they are data-intensive: they typically demand substantial training sets, significant pretraining effort, and careful, representative selection of training structures. This talk presents a data-efficient alternative: on-the-fly construction and exploration of PESs using nonparametric Bayesian surrogates that require no pretraining. Our framework embeds prior physical knowledge through non-constant prior mean functions (e.g., simple physical models or low-level semiempirical Hamiltonian), yielding a physically informed, real-time learner with built-in uncertainty quantification to select the next ab initio queries. We show that this surrogate-driven approach substantially reduces electronic-structure calls for geometry optimization, conformational ensemble refinement, reaction path determination (gas, bulk, and interfaces), and excited-state photodynamics, while producing on-the-fly surrogate PESs that closely track ground-truth surfaces. These surrogates can then be reused as hierarchical priors for higher-level quantum-chemistry tasks or as guidance for steering nuclear motion along dynamical trajectories beyond the Born-Oppenheimer approximation.

Department of Bioengineering, Northeastern University

My lab focuses on developing computational methodologies that reveal how gene regulatory networks control cell state transitions during normal development and disease progression. By integrating top-down bioinformatics with bottom-up systems biology modeling, we aim to elucidate the interplay between gene network topology, single-cell gene expression dynamics, and cell phenotypic transitions. A cornerstone of our work is an ensemble-based mathematical modeling framework, RACIPE (Random Circuit Perturbation), which generates ensembles of models directly from gene circuit topology by randomly sampling kinetic parameters. This framework enables unbiased exploration of gene expression states without specifying kinetic parameters or parameter fitting and reveals biologically relevant states determined by network structure.
In the first part of my presentation, I will describe the methodological foundation and generalizations of RACIPE, including extensions for modeling networks with multiple regulatory types such as microRNA-based regultion, incorporation of intrinsic and extrinsic noise, and applications in circuit motif analysis and gene network optimization. While our studies are primarily computational and numerical, they raise several theoretical questions relevant to statistical physics and nonlinear dynamics that may be of interest to the audience.
In the second part, I will discuss two major applications of our recent work: (1) network coarse-graining methods that capture essential dynamics in a reduced gene circuit without requiring an explicit Hamiltonian, and (2) dissection of reversible and irreversible cell state transitions by using single-cell gene expression profiles and network topology. Together, these studies illustrate how our computational modeling framework reveals the design principles of gene regulation and offers new insights into the mechanisms that govern cell fate determination.

Department of Chemistry, Tufts University

Computing free energy differences is needed in many areas of science and engineering. For physical and chemical processes, free energy determines the relative stability of different states, which makes it a key quantity in understanding the thermodynamics of a process. Although the concept of free energy originates from thermodynamics, it reaches beyond and is frequently encountered in fields such as machine learning (ML) and Bayesian statistics. Because of its broad importance, researchers from various fields have been developing methods to calculate free energy. The benefit of having researchers from various fields working on similar problems is that new ideas and methods can be borrowed and adapted from one field to another. In ML, computing free energy is often not the primary objective but rather an intermediate step required for training or evaluating probabilistic models. When training probabilistic models of data in ML, the repeated computation of free energy becomes a significant computational bottleneck. This challenge has motivated the development of ML methods and training algorithms that circumvent the need for free energy calculations altogether. Interestingly, these models and methods, originally designed to avoid computing free energies, have proven to be effective tools for estimating free energy themselves. In part I of the talk, I will first discuss classical methods for computing free energy. I will then discuss how to leverage recent generative models in ML for efficient free energy estimation. In part II of the talk, I will introduce a Bayesian statistical framework for free energy calculation and show how it can be used to improve the accuracy of free energy estimates by incorporating prior knowledge in a principled manner.

Department of Chemistry, University of Massachusetts Boston

By assembling supramolecular materials through synthetic chemical reaction networks, it is becoming possible to mimic some of the remarkable behaviors of biological systems. An influx of reagents can sustain materials that function dynamically–responding to stimuli, converting free energy into mechanical work, and self-healing after perturbation. However, these dissipative materials operate far from equilibrium and require the continuous consumption of energy, introducing fundamental challenges in designing and controlling how effectively such functions are executed. In these systems, molecular-scale chemical reaction networks govern the dissipative assembly and disassembly of mesoscale structures. As a result, material function–and even material lifetime–is often transient, with experiments revealing assembly and response processes that unfold over finite timescales. This tight coupling of dynamics across scales places these materials beyond the scope of traditional thermodynamic descriptions and motivates new approaches to nonequilibrium statistical mechanics. From the geometry of their irreversible dynamics, two promising design principles emerge.

First, dissipative materials must balance how fast a function is executed against how much energy is consumed in the process. In the first part of the talk, I will introduce thermodynamic speed limits on dissipation that quantify this tradeoff. I will discuss recent work suggesting that these bounds can serve as a design principle spanning length and timescales. These speed limits share theoretical connections with regression analysis and, by extension, existing machine learning techniques, offering a potential framework for optimizing energy efficiency and the timing of structure formation in dissipative materials.

Second, controlling the dynamics of these materials requires designing their nonequilibrium response from the underlying chemical kinetics. In the second part of the talk, I will discuss how open reaction networks can transiently amplify small fluctuations in concentrations and produce large, spontaneous increases in the yield of assembled material. After a small dose of reagent, the chemical kinetics do not simply relax toward a steady state; instead, they can evolve further away from equilibrium and spontaneously assemble thermodynamically unstable products. For a model supramolecular system, this mechanism can produce 400% more material than the steady state yield. By deriving the conditions for this transient amplification, I will show that steady states that break detailed balance can support stronger growth and higher yields. The precise conditions for this behavior arise from the nonnormal stability of these steady states and provide a route to exploit fluctuations in dissipative materials.

Department of Chemistry, Tufts University

A robust ability to target specific protein–protein interactions (PPIs) and protein surfaces, both intracellular and extracellular, would provide control of diverse cellular functions for both fundamental research and therapeutic intervention. However, PPIs and protein surfaces are challenging targets for small molecules. Cyclic peptides offer a promising solution, owing to their inherently large surface area and ability to easily mimic functional groups and structures at protein interfaces. Cyclic peptides can also be proteolytically stable and membrane-permeable, providing access to high-value intracellular targets.

Unfortunately, minimal structural information is available for cyclic peptides in solution, making it difficult to perform structure-based design or understand why sequence-similar cyclic peptides can display vastly different binding affinities, membrane-permeability, and other properties. Critically, NMR studies show that most cyclic peptides reported thus far, including the approved cyclic peptide drugs, adopt multiple conformations in solution, existing as structural ensembles. Robust experimental methods to structurally characterize each conformation in a structural ensemble do not currently exist.

Since structural information for most cyclic peptides is impractical to obtain by experiments, computational approaches provide an optimal alternative. Molecular dynamics (MD) simulation, particularly explicit-solvent MD, has long been used to provide high-quality predictions of the structures, dynamics, and functions of peptides and proteins. However, the ring strain in small cyclic peptides often leads to high free energy barriers between conformations that prevent sampling of all relevant structures by standard MD. Larger cyclic peptides can also present significant issues, because they often have vast conformational ensembles. Therefore, running standard, plain MD simulation in an explicit solvent at 300K typically provides insufficient conformational sampling of cyclic peptides even after weeks of simulation. I will describe enhanced sampling methods and how they can be tailored to cyclic peptides to efficiently sample their conformations and accurately characterize their solution structural ensembles using explicit-solvent MD.

In addition, I will describe how we develop StrEAMM (Structural Ensembles Achieved by Molecular dynamics and Machine learning) models for cyclic peptides by combining MD simulation and machine learning. We can now provide simulation-quality cyclic peptide structure predictions in seconds. We expect this capability to rapidly predict cyclic peptide structures, enabling researchers to understand the structural basis for the diverse properties of cyclic peptides and greatly accelerating the development of this unique class of molecules.

Department of Chemistry, Brandeis University

In this seminar, we will present computational methods from our lab for modeling enzymatic reactions and bioimaging probes. In the first part of the seminar, we will focus on the identification of the minimum free energy pathways, which is essential to fully understand an enzyme reaction. To reduce the steep computational cost of these free energy simulations, we have adapted the multiple time-step integration algorithm and machine learning potentials in these simulations. We will showcase the power of our simulation protocols with chorismate mutase and CRISPR-Cas9 enzymes. In the second part, we will look at small molecule bioimaging probes. Specifically, we will show how to analyze the interactions between chromophore orbitals and substituent (or solvent) orbitals, and explain how electron-donating and withdrawing groups could modulate the chromophore emission wavelengths.

Department of Chemical Engineering, Massachusetts Institute of Technology

Department of Chemistry, Boston University

Department of Chemistry, Massachusetts Institute of Technology

School of Science, Westlake University

Department of Chemistry, Boston University