Publications

Machine learning interatomic potentials (MLIPs) have become powerful tools to extend molecular simulations beyond the limits of quantum methods, offering near-quantum accuracy at much lower computational cost. Yet, developing reliable MLIPs remains difficult because it requires generating high-quality datasets, preprocessing atomic structures, and carefully training and validating models. In this work, we introduce an Automated Machine Learning Pipeline (AMLP) that unifies the entire workflow from dataset creation to model validation. AMLP employs large-language-model agents to assist with electronic-structure code selection, input preparation, and output conversion, while its analysis suite (AMLP-Analysis), based on ASE supports a range of molecular simulations. The pipeline is built on the MACE architecture and validated on acridine polymorphs, where, with a straightforward fine-tuning of a foundation model, mean absolute errors of 1.7 meV/atom in energies and 7.0 meV/Å in forces are achieved. The fitted MLIP reproduces DFT geometries with sub-Å accuracy and demonstrates stability during molecular dynamics simulations in the microcanonical and canonical ensembles.

We introduce a size-consistent and orbital-invariant formalism for constructing correlation functionals based on the adiabatic connection for density functional theory (DFT). By constructing correlation energy matrices for the weak and strong correlation limits in the space of occupied orbitals, our method, which we call orbital-based size-consistent matrix interpolation (OSMI), avoids previous difficulties in the construction of size-consistent adiabatic connection functionals. We design a simple, nonempirical adiabatic connection and a one-parameter strong-interaction limit functional, and we show that the resulting method reproduces the correlation energy of the uniform electron gas over a wide range of densities. When applied to subsets of the GMTKN55 thermochemistry database, OSMI is more accurate on average than MP2 and nonempirical density functionals. Most notably, OSMI provides excellent predictions of the barrier heights we tested, with average errors of less than 2 kcal mol

The growing availability of biological data resources has transformed research, yet their effective use remains challenging: selecting appropriate sources requires domain knowledge, data are fragmented across databases, and synthesizing results into reliable conclusions is labor-intensive. Although large language models promise to address these barriers, their impact in biomedicine has been limited by unsupported statements, incorrect claims, and lack of provenance. We introduce Alvessa, an evidence-grounded agentic research assistant designed around verifiability. Alvessa integrates entity recognition, orchestration of pre-validated biological tools, and data-constrained answer generation with statement-level verification against retrieved records, explicitly flagging unsupported claims and guiding revision when reliability criteria are not met. We evaluate Alvessa on dbQA from LAB-Bench and GenomeArena, a benchmark of 720 questions spanning gene and variant annotation, pathways, molecular interactions, miRNA targets, drug-target evidence, protein structure, and gene-phenotype associations. Alvessa substantially improves accuracy relative to general-purpose language models and performs comparably to coding-centric agents while producing fully traceable outputs. Using adversarial perturbations, we show that detection of fabricated statements depends critically on access to retrieved evidence. We further demonstrate application to drug discovery, where evidence-grounded synthesis enables identification of candidate targets missed or misattributed by literature-centered reasoning alone. Alvessa and GenomeArena are released to the community to support reproducible, verifiable AI-assisted biological research.

Cryogenic electron microscopy (cryo-EM) has emerged as a powerful method for resolving the structure of biological macromolecules. Recently, several computational methods have been developed to study the heterogeneity of molecules in single-particle cryo-EM. In this study, we analyze a publicly available dataset of TRPV1 using five such methods: 3DFlex, 3DVA, cryoDRGN, ManifoldEM, and Bayesian ensemble reweighting. We find significant heterogeneity, but each method produces different results, with some detecting only compositional or conformational heterogeneity. To compare these diverse results, we develop AnaVox to quantitatively determine agreement between heterogeneity methods. Furthermore, applying Bayesian ensemble reweighting combined with molecular dynamics simulations supports the presence of these rarer states within the sample. This study shows that although current methods reveal the presence of heterogeneity, their stochasticity and potential bias present challenges for their routine use. However, with future development, these tools will enable the use of cryo-EM data for quantitative biophysical investigations.

Cryogenic electron microscopy (cryo-EM) has emerged as a powerful method for resolving the structure of biological macromolecules. Recently, several computational methods have been developed to study the heterogeneity of molecules in single-particle cryo-EM. In this study, we analyze a publicly available dataset of TRPV1 using five such methods: 3DFlex, 3DVA, cryoDRGN, ManifoldEM, and Bayesian ensemble reweighting. We find significant heterogeneity, but each method produces different results, with some detecting only compositional or conformational heterogeneity. To compare these diverse results, we develop AnaVox to quantitatively determine agreement between heterogeneity methods. Furthermore, applying Bayesian ensemble reweighting combined with molecular dynamics simulations supports the presence of these rarer states within the sample. This study shows that although current methods reveal the presence of heterogeneity, their stochasticity and potential bias present challenges for their routine use. However, with future development, these tools will enable the use of cryo-EM data for quantitative biophysical investigations.

Many tasks in single-particle cryo-electron microscopy (cryo-EM), such as 2D/3D classification and homo/heterogeneous reconstruction, require optimizing model parameters to minimize the discrepancy between observed data and a forward model. The standard Mean Squared Error (MSE) loss function is computationally efficient but suffers from a non-convex rugged loss landscape, particularly for high-resolution heterogeneity inference. In this work, we investigate the practical utility of Sliced Wasserstein (SW) distances. We implement exact W2 estimators (inverse-CDF and greedy matching) of projections alongside a computationally efficient proxy based on the L2 norm of CDFs, a formulation akin to the sliced Cramér–von Mises distance. We establish the latter as a robust, fully differentiable workhorse for the cryo-EM forward model. We evaluate its performance against the MSE in joint inference tasks recovering pose, CTF parameters, and conformational heterogeneity. Our results demonstrate that SW significantly broadens the basin of attraction, enabling robust gradient-based optimization from distant initializations where MSE fails. Using a helical spiral toy model, we highlight how SW losses are sensitive to per-particle contrast, where background noise level miscalibration can induce geometric bias in the inferred structure. We show that this bias is manageable through a joint optimization strategy that treats background contrast as a learnable parameter. Finally, we validate the approach on a synthetic dataset using the Zernike3D framework, showing that the SW loss works and yields an accurate landscape representations, comparable with MSE. These findings establish SW as a powerful tool for navigating the rugged landscapes of cryo-EM forward model parameters

Parkinson’s Disease (PD) is characterized by the loss of dopaminergic neurons in the substantia nigra pars compacta (SNpc), which is associated with changes in microglia function. While age remains the biggest risk factor, the underlying molecular cause of PD onset and its concurrent neuroinflammation are not well understood. Many identified PD risk genes have been directly linked to dopamine neuron impairment, while others are linked to immune cell function. In this study, we found that the PD risk gene FAM49B is critically expressed in microglia of the human SNpc and is downregulated with age and PD. We utilized human and murine microglia cells to demonstrate the role of FAM49B in regulating fundamental microglial functions such as cytoskeletal maintenance, migration, surface adherence, energy homeostasis, autophagy, and, importantly, inflammatory response. Downregulation of microglial FAM49B, as observed in the SNpc of aging individuals, led to significant alterations in these cellular functions, which are associated with increased microglial activation. Thus, our study highlights novel cell-type-specific roles of FAM49B and provides a potential mechanism for susceptibility to neuroinflammation, and reactive gliosis observed in both PD and normal aging.

We present the Score-based Autoencoder for Multiscale Inference (SAMI), a method for unsupervised representation learning that combines the theoretical frameworks of diffusion models and VAEs. By unifying their respective evidence lower bounds, SAMI formulates a principled objective that learns representations through score-based guidance of the underlying diffusion process. The resulting representations automatically capture meaningful structure in the data: it recovers ground truth generative factors in our synthetic dataset, learns factorized, semantic latent dimensions from complex natural images, and encodes video sequences into latent trajectories that are straighter than those of alternative encoders, despite training exclusively on static images. Furthermore, SAMI can extract useful representations from pre-trained diffusion models with minimal additional training. Finally, the explicitly probabilistic formulation provides new ways to identify semantically meaningful axes in the absence of supervised labels, and its mathematical exactness allows us to make formal statements about the nature of the learned representation. Overall, these results indicate that implicit structural information in diffusion models can be made explicit and interpretable through synergistic combination with a variational autoencoder.

Singularity swap quadrature (SSQ) is an effective method for the evaluation at nearby targets of potentials due to densities on curves in three dimensions. While highly accurate in most settings, it is known to suffer from catastrophic cancellation when the kernel exhibits both near-vanishing numerators and strong singularities, as arises with scalar double layer potentials or tensorial kernels in Stokes flow or linear elasticity. This precision loss turns out to be tied to the interpolation basis, namely monomial (for open curves) or Fourier (for closed curves). We introduce a simple yet powerful remedy: target-specific translated monomial and Fourier bases that explicitly incorporate the near-vanishing behavior of the kernel numerator. We combine this with a stable evaluation of the constant term which now dominates the integral, significantly reducing cancellation. We show that our approach achieves close to machine precision for prototype integrals, and up to ten orders of magnitude lower error than standard SSQ at extremely close evaluation distances, without significant additional computational cost.

Neural activity is usually interpreted by imposing external labels (e.g., stimuli or position during locomotion) and decoding within that space (e.g. replay). While powerful, such supervision can mask structure in the data that do not correspond to the label. Unsupervised methods, in turn, often assume smooth latent dynamics and miss genuine discontinuities. We introduce a conceptually simple, computationally efficient latent variable model that infers both (i) the latent variables organizing population activity and (ii) whether their dynamics are continuous or fragmented in time. Fitting reduces to an expectation-maximization (EM) procedure that alternates two operations familiar to systems neuroscience—tuning-curve estimation and label decoding—without requiring external labels. Applied to rodent hippocampal spike recordings, the model reveals distinct population patterns at the same physical position that supervised spatial decoding fails to detect. While learned latents exhibit place-field-like tuning, their reactivation patterns are better distinguished by behavioral states. The model further identifies a continuity-fragmentation axis that characterizes population activities across sleep-wake brain states that is modulated by cholinergic inputs. By not relying on externally imposed spatial labels, our approach exposes structure that supervised approaches obscure and provides a powerful tool for datasets lacking behavioral tracking.