Alex Bronstein

Machine learning has transformed vision, language, games-even the prediction of a protein's folded shape-and every one of those victories quietly relied on gifts financial markets withhold: abundant data, stable rules, a legible signal, and a target that does not move because it was predicted. Medium-horizon asset return forecasting is starved of all four, which is why the sophistication of a model says so little about whether it will survive a live market. Asset managers, ever keen on the cutting edge, keep adopting these methods in the pursuit of alpha-sometimes successfully, often not. We argue that what divides the models that endure from those that vanish is less the power of the box than the rigor built around and into it-governance by construction, not as an afterthought: controls and economic constraints designed into the model from the outset, uncertainty quantified so the system knows when to abstain, frictions and capacity priced before the first trade, and every claimed edge made auditable. Borrowing medicine's discipline, where trust rests on a protocol of evidence and vigilance rather than on a fully understood mechanism, we suggest that investors should shift the burden of proof from fully explaining the model to governing it-and close with the questions an allocator should ask before believing that any machine has learned to beat the market.

AlphaFold3 predicts highly accurate protein structures from sequence but tends to collapse to a single dominant conformation, even when the underlying structure is inherently heterogeneous. Moreover, its predictions are oblivious to experimental conditions that can alter local sequence conformation. In this work, we show that AlphaFold3 can be guided to match data obtained by nuclear magnetic resonance (NMR) spectroscopy, X-ray crystallography and cryogenic electron microscopy (cryo-EM) experiments and combinations thereof. Our approach can also incorporate data that explicitly report on dynamics, such as site-resolved order parameters. We demonstrate that this methodology generates compact structural ensembles whose ensemble-averaged observables agree with experiment, with fewer distance restraint violations than traditionally resolved NMR structures and with unmodeled alternate conformations uncovered in electron density. This methodology paves the way for experimentally aware predictive models that generate structural ensembles consistent with the measurements, potentially over multiple modalities, and that can be further refined toward thermodynamically grounded ensembles by incorporating energetics.

β-turns are among the most common structural motifs in proteins, yet their conformational dynamics and sequence determinants remain incompletely understood. Here we present a data-driven classification and dynamic analysis of β-turn conformations using large-scale molecular dynamics trajectories from the mdCATH database. Clustering of backbone dihedral angles using a cross-bond Ramachandran representation identifies six β-turn types, including a previously uncharacterized hybrid I/I' cluster that combines geometric features of canonical type I and I' conformations. Time-resolved analysis indicates that this hybrid state acts as a transient intermediate state of β-turns. Transitions observed in molecular dynamics simulations closely match NMR ensembles and altlocs detected in X-ray crystal structures, with the most dominant exchanges occurring between type I and II, and between type I' and II' turns. Sequence analysis shows that each turn type exhibits characteristic amino acid preferences at the central residues (i+1 and i+2). Within these overall preferences, specific residue pairs display distinct biases toward static or dynamic behavior. Targeted in silico substitutions that interchange dynamic- and static-enriched residue pairs shift the conformational behavior of turns accordingly, providing direct support for these sequence-dynamics relationships. Analysis of flanking secondary-structure environments reveals that structural context further modulates turn flexibility, with strand- and coil-associated turns exhibiting higher dynamic propensity than helix-associated turns. Together, these results reveal how sequence composition and structural context jointly shape the conformational landscape of β-turns.

Although X-ray crystallography captures the ensemble of conformations present within the crystal lattice, models typically depict only the most dominant conformation, obscuring the existence of alternative states. Applying the electron density-guided AlphaFold3 approach to β ₂ -Microglobulin highlights how ensembles of alternate backbone conformations can be systematically modeled directly from crystallographic maps. This study also highlights how the detection of conformational ensembles is affected by the local quality of electron density and subtle variations in crystallization conditions and lattice packing. These results demonstrate that density-guided AlphaFold3 can uncover conformational heterogeneity missed by conventional refinement, offering a robust, systematic framework to capture the full structural landscape of proteins in crystals and enhancing the interpretive power of macromolecular crystallography. Synopsis: Electron-density-guided AlphaFold3 reveals previously unmodeled conformational heterogeneity in β ₂ -Microglobulin and shows how crystal packing influences ensemble detection in X-ray crystallography.

Recent work has established learned k-space acquisition patterns as a promising direction for improving reconstruction quality in accelerated Magnetic Resonance Imaging (MRI). Despite encouraging results, most existing research focuses on acquisition patterns optimized for a single dataset or modality, with limited consideration of their transferability across imaging domains. In this work, we demonstrate that the benefits of learned k-space sampling can extend beyond the training domain, enabling superior reconstruction performance under domain shifts. Our study presents two main contributions. First, through systematic evaluation across datasets and acquisition paradigms, we show that models trained with learned sampling patterns exhibitimproved generalization under cross-domain settings. Second, we propose a novel method that enhances domain robustness by introducing acquisition uncertainty during training-stochastically perturbing k-space trajectories to simulate variability across scanners and imaging conditions. Our results highlight the importance of treating kspace trajectory design not merely as an acceleration mechanism, but as an active degree of freedom for improving domain generalization in MRI reconstruction.

Stellar astrophysics relies on diverse observational modalities—primarily photometric light curves and spectroscopic data—from which fundamental stellar properties are inferred. While machine learning (ML) has advanced analysis within individual modalities, the complementary information encoded across modalities remains largely underexploited. We present the dual embedding for stellar astronomy (DESA) model, a novel multimodal foundation model that integrates light curves and spectra to learn a unified, physically meaningful latent space for stars. DESA first trains separate modality-specific encoders using a hybrid supervised/self-supervised scheme, and then aligns them through DualFormer, a transformer-based cross-modal integration module tailored for astrophysical data. DualFormer combines cross- and self-attention, a novel dual-projection alignment loss, and a projection-space eigendecomposition that yields physically structured embeddings. We demonstrate that DESA significantly outperforms leading unimodal and self-supervised baselines across a range of tasks. In zero- and few-shot settings, DESA’s learned representations recover stellar color–magnitude and Hertzsprung–Russell diagrams with high fidelity ( R ² = 0.92 for photometric regressions). In full fine-tuning, DESA achieves state-of-the-art accuracy for binary star detection (AUC = 0.99, AP = 1.00) and stellar age prediction (RMSE = 0.94 Gyr). As a compelling case, DESA naturally separates synchronized binaries from young stars—two populations with nearly identical light curves—purely from their embedded positions in UMAP space, without requiring external kinematic or luminosity information. DESA thus offers a powerful new framework for multimodal, data-driven stellar population analysis, enabling both accurate prediction and novel discovery.

AlphaFold3 predicts highly accurate protein structures from sequence, but tends to collapse to a single dominant conformation, even when the underlying structure is inherently heterogeneous. Moreover, its predictions are oblivious to experimental conditions that can alter local sequence conformation. In this work we show that AlphaFold3 can be guided to match data obtained by NMR spectroscopy, X-ray crystallography and cryo-EM experiments. We demonstrate that this methodology can generate ensembles of conformations having less distance restraint violations than traditionally resolved NMR structures and uncover unmodelled alternate conformations detectable in electron density. We show that AlphaFold3 can also be guided by cryo-electron microscopy maps and that doing so in combination with NMR parameters improves model quality. This methodology paves the way for the development of experimentally aware predictive models that capture the ensemble nature of protein structures.

AlphaFold3 predicts highly accurate protein structures from sequence, but tends to collapse to a single dominant conformation, even when the underlying structure is inherently heterogeneous. Moreover, its predictions are oblivious to experimental conditions that can alter local sequence conformation. In this work we show that AlphaFold3 can be guided to match data obtained by NMR spectroscopy, X-ray crystallography and cryo-EM experiments, and combinations thereof. Our approach can also incorporate data that explicitly report on dynamics, such as site-resolved order parameters. We demonstrate that this methodology can generate ensembles of conformations having less distance restraint violations than traditionally resolved NMR structures and uncover unmodelled alternate conformations detectable in electron density. This methodology paves the way for the development of experimentally aware predictive models that capture the ensemble nature of protein structures.

Understanding physiological responses during running is critical for performance optimization, tailored training prescriptions, and athlete health management. We introduce a comprehensive framework -- what we believe to be the first capable of predicting instantaneous oxygen consumption (VO$_{2}$) trajectories exclusively from consumer-grade wearable data. Our approach employs two complementary physiological models: (1) accurate modeling of heart rate (HR) dynamics via a physiologically constrained ordinary differential equation (ODE) and neural Kalman filter, trained on over 3 million HR observations, achieving 1-second interval predictions with mean absolute errors as low as 2.81\,bpm (correlation 0.87); and (2) leveraging the principles of precise HR modeling, a novel VO$_{2}$ prediction architecture requiring only the initial second of VO$_{2}$ data for calibration, enabling robust, sequence-to-sequence metabolic demand estimation. Despite relying solely on smartwatch and chest-strap data, our method achieves mean absolute percentage errors of approximately 13\%, effectively capturing rapid physiological transitions and steady-state conditions across diverse running intensities. Our synchronized dataset, complemented by blood lactate measurements, further lays the foundation for future noninvasive metabolic zone identification. By embedding physiological constraints within modern machine learning, this framework democratizes advanced metabolic monitoring, bridging laboratory-grade accuracy and everyday accessibility, thus empowering both elite athletes and recreational fitness enthusiasts.