Computational Life Science

The Covino group develops theoretical and computational methods to investigate rare molecular events in biological systems. Biomolecules continuously transition between different conformations to perform their functions, but many of these transitions occur on timescales that make them difficult to access with experiments and conventional molecular simulations.

The group combines molecular simulation, machine learning and simulation-based inference to uncover the mechanisms underlying these processes and to connect computational models with experimental data. Method development and biological applications go hand in hand, with a particular focus on approaches designed to exploit massively parallel high-performance computing.

Research Topics

Simulating Rare Events: Path Sampling, the Committor and Machine Learning

Conformational transitions, binding and folding are rare events: a molecule spends almost all of its time in metastable states and transitions between them only during brief, infrequent excursions. In a conventional molecular dynamics simulation, most of the computational time is therefore spent waiting for such transitions to occur.

We develop methods that focus computational resources on the transition itself. At the core of this work is AIMMD (AI for Molecular Mechanism Discovery): a neural network learns the committor function from each sampled trajectory—the probability that a given configuration will reach the target state—and uses it to determine where subsequent simulations should be initiated.

From the learned committor, we derive mechanisms, free energies and rates; using symbolic regression, we translate it into compact mathematical expressions that are physically interpretable. Because AIMMD manages thousands of short, mutually independent simulations, the approach is designed for massively parallel computing. In parallel, we continue to advance the theory of the committor, most recently through a variational formulation based on low-dimensional projections (“sliced committor”).

Simulation-Based Inference: From Experiment to Quantitative Model

Single-molecule experiments—such as optical tweezers and single-molecule Förster resonance energy transfer (smFRET)—observe individual molecules at work, but only indirectly: the measurement apparatus distorts the signal and introduces noise, and the likelihood of a realistic model generally cannot be calculated.

Simulation-based inference (SBI) circumvents this problem. We jointly simulate the molecule and the experimental apparatus, generate large synthetic datasets, and train neural networks to estimate the Bayesian posterior distribution of model parameters directly from the experimental data. This allows us to systematically disentangle the properties of the molecule from artefacts introduced by the measurement apparatus.

Recently, we demonstrated that the folding landscape of a molecule can be reconstructed from only a few seconds of optical-tweezer data—approximately one hundredth of the amount of data previously required. We are currently extending this approach to smFRET using a photon-by-photon likelihood.

cryoSBI: Conformations from Individual Cryo-EM Images

Cryo-electron microscopy freezes millions of individual molecules in the conformational states they occupy at that moment. Each image therefore represents a snapshot of a molecular conformation—but an extremely noisy one.

With cryoSBI, developed in collaboration with the Flatiron Institute in New York, we infer the molecular conformation and its associated uncertainty from each individual image. Once trained, the network requires only milliseconds to perform inference on a single image.

This makes it possible to learn conformational ensembles directly from the data, including rare and short-lived states.

A Digital Twin of the Cell (SCALE)

Within the SCALE (SubCellular Architecture of LifE) Cluster of Excellence at Goethe University, we are working towards building a digital twin of the cell: a computational model that integrates structural data, simulations across multiple scales and microscopy to enable quantitative predictions of cellular processes and comparison with experimental data.

The group’s methods—including rare-event simulation, simulation-based inference and multiscale modelling—provide the building blocks for this effort.

Group Leader

Roberto Covino studied physics at the University of Bologna, where he completed his studies with a thesis on Hawking radiation. He then moved to the University of Trento for his PhD, developing computational methods to investigate how proteins fold into their native structures.

From 2014 to 2020, he worked in the Department of Theoretical Biophysics at the Max Planck Institute of Biophysics in Frankfurt. There, he investigated how cells sense the physical state of their membranes and developed computational approaches combining physics-based models with machine learning.

In 2020, he established an independent research group at the Frankfurt Institute for Advanced Studies (FIAS), where he has been a Senior Fellow since 2024. Following a professorship in Artificial Intelligence in Protein Science at the University of Bayreuth from 2023 to 2024, he was appointed Professor of Computational Life Science at the Institute of Computer Science at Goethe University Frankfurt in 2024. He has been the head of the NIC Research Group Computational Life Science since 2025.

Selected Publications

For a complete list of publications, please visit Roberto Covino’s ORCID profile and Google Scholar profile.

L. Dingeldein, A. Lyons, P. Cossio, M. T. Woodside, R. Covino, „Quantitative and Predictive Folding Models from Limited Single-Molecule Data Using Simulation-Based Inference“, *Phys. Rev. Lett.* **137**, 088402 (2026). DOI 10.1103/yltb-6jkj. Editors' Suggestion.

L. Dingeldein, D. Silva-Sánchez, L. Evans, E. D'Imprima, N. Grigorieff, R. Covino, P. Cossio, „Amortized template matching of molecular conformations from cryoelectron microscopy images using simulation-based inference“, *Proc. Natl. Acad. Sci. U.S.A.* **122**, e2420158122 (2025). DOI 10.1073/pnas.2420158122.

L. Dingeldein, P. Cossio, R. Covino, „Simulation-based inference of single-molecule experiments“, *Curr. Opin. Struct. Biol.* **91**, 102988 (2025). DOI 10.1016/j.sbi.2025.102988.

M. Petersen, S. Lichtinger, R. Covino, „Committors and Reaction Rates from Trial Functions That Violate the Boundary Conditions“, arXiv:2608.02536 (2026). Preprint; keine Journalversion auffindbar.

K. Töpfer, G. Lazzeri, V. Ossanna, F. Renner, G. Lattanzi, R. Covino, B. G. Keller, „A Machine-Learned Symbolic Committor for a Chemical Reaction: Retinal Isomerization“, arXiv:2604.24245 (2026). Preprint.