Atomistic simulations of matter, especially those that leverage first-principles (ab initio) electronic structure theory, provide a microscopic view of the world, underpinning much of our understanding of chemistry and materials science. Over the last decade or so, machine-learned force fields have transformed atomistic modeling by enabling simulations of ab initio quality over unprecedented time and length scales. However, early machine-learning (ML) force fields have largely been limited by (i) the substantial computational and human effort required to develop and validate potentials for each particular system of interest and (ii) a general lack of transferability from one chemical system to the next. Here, we show that it is possible to create a general-purpose atomistic ML model, trained on a public dataset of moderate size, that is capable of running stable molecular dynamics for a wide range of molecules and materials. We demonstrate the power of the MACE-MP-0 model—and its qualitative and at times quantitative accuracy—on a diverse set of problems in the physical sciences, including properties of solids, liquids, gases, chemical reactions, interfaces, and even the dynamics of a small protein. The model can be applied out of the box as a starting or “foundation” model for any atomistic system of interest and, when desired, can be fine-tuned on just a handful of application-specific data points to reach ab initio accuracy. Establishing that a stable force-field model can cover almost all materials changes atomistic modeling in a fundamental way: experienced users obtain reliable results much faster, and beginners face a lower barrier to entry. Foundation models thus represent a step toward democratizing the revolution in atomic-scale modeling that has been brought about by ML force fields.

The rapid adoption of machine learning in various scientific domains calls for the development of best practices and community agreed-upon benchmarking tasks and metrics. We present Matbench Discovery as an example evaluation framework for machine learning energy models, here applied as pre-filters to first-principles computed data in a high-throughput search for stable inorganic crystals. We address the disconnect between (1) thermodynamic stability and formation energy and (2) retrospective and prospective benchmarking for materials discovery. Alongside this paper, we publish a Python package to aid with future model submissions and a growing online leaderboard with adaptive user-defined weighting of various performance metrics allowing researchers to prioritize the metrics they value most. To answer the question of which machine learning methodology performs best at materials discovery, our initial release includes random forests, graph neural networks, one-shot predictors, iterative Bayesian optimizers and universal interatomic potentials. We highlight a misalignment between commonly used regression metrics and more task-relevant classification metrics for materials discovery. Accurate regressors are susceptible to unexpectedly high false-positive rates if those accurate predictions lie close to the decision boundary at 0 eV per atom above the convex hull. The benchmark results demonstrate that universal interatomic potentials have advanced sufficiently to effectively and cheaply pre-screen thermodynamic stable hypothetical materials in future expansions of high-throughput materials databases.

The binding of transcription factors to short recognition sequences plays a pivotal role in controlling the expression of genes. The sequence and shape characteristics of binding sites influence DNA binding specificity and have also been implicated in modulating the activity of transcription factors downstream of binding. To quantitatively assess the transcriptional activity of tens of thousands of designed synthetic sites in parallel, we developed a synthetic version of STARR-seq (synSTARR-seq). We used the approach to systematically analyze how variations in the recognition sequence of the glucocorticoid receptor (GR) affect transcriptional regulation. Our approach resulted in the identification of a novel highly active functional GR binding sequence and revealed that sequence variation both within and flanking GR’s core binding site can modulate GR activity without apparent changes in DNA binding affinity. Notably, we found that the sequence composition of variants with similar activity profiles was highly diverse. In contrast, groups of variants with similar activity profiles showed specific DNA shape characteristics indicating that DNA shape may be a better predictor of activity than DNA sequence. Finally, using single cell experiments with individual enhancer variants, we obtained clues indicating that the architecture of the response element can independently tune expression mean and cell-to cell variability in gene expression (noise). Together, our studies establish synSTARR as a powerful method to systematically study how DNA sequence and shape modulate transcriptional output and noise. The expression level of genes is controlled by transcription factors, which are proteins that bind to genomic response elements that contain their recognition DNA sequence. Importantly, genes are not simply turned on but need to be expressed at the right level. This is, at least in part, assured by the sequence composition of genomic response elements. Here, we studied how the recognition DNA sequence influences gene regulation by a transcription factor called the glucocorticoid receptor. Specifically, we developed a method to test the activity of variants in a highly parallelized setting where everything is kept identical except for the sequence of the binding site. The systematic analysis of tens of thousands of sequence variants facilitated the identification of a previously unknown sequence variant with high activity. Moreover, we report how sequence variation of the response element influences cell-to-cell variability in expression levels. Finally, we observe similar activity profiles for distinct sequence variants that share similar three-dimensional DNA shape characteristics arguing that the three-dimensional perception of DNA by the glucocorticoid receptor, modulates its activity towards individual target genes.

Non-targeted metabolomics holds great promise for advancing precision medicine and biomarker discovery. However, identifying compounds from tandem mass spectra remains a challenging task due to the incomplete nature of spectral reference libraries. Augmenting these libraries with simulated mass spectra can provide the necessary references to resolve unmatched spectra, but generating high-quality data is difficult. In this study, we present FIORA, an open-source graph neural network designed to simulate tandem mass spectra. Our main contribution lies in utilizing the molecular neighborhood of bonds to learn breaking patterns and derive fragment ion probabilities. FIORA not only surpasses state-of-the-art fragmentation algorithms, ICEBERG and CFM-ID, in prediction quality, but also facilitates the prediction of additional features, such as retention time and collision cross section. Utilizing GPU acceleration, FIORA enables rapid validation of putative compound annotations and large-scale expansion of spectral reference libraries with high-quality predictions.

SynCoTrain is a PU-learning framework using dual GNN classifiers to predict material synthesizability. It leverages co-training to mitigate model bias and enhance generalizability.

Genome segmentation methods are powerful tools to obtain cell type or tissue-specific genome-wide annotations and are frequently used to discover regulatory elements. However, traditional segmentation methods show low predictive accuracy and their data-driven annotations have some undesirable properties. As an alternative, we developed ModHMM, a highly modular genome segmentation method. Inspired by the supra-Bayesian approach, it incorporates predictions from a set of classifiers. This allows to compute genome segmentations by utilizing state-of-the-art methodology. We demonstrate the method on ENCODE data and show that it outperforms traditional segmentation methods not only in terms of predictive performance, but also in qualitative aspects. Therefore, ModHMM is a valuable alternative to study the epigenetic and regulatory landscape across and within cell types or tissues.