Potestio Lab

18/02/2025

🥳🥳🥳 We are pleased to announce the publication of our paper, “Density of States in Neural Networks: An In-Depth Exploration of Learning in Parameter Space” in Transactions on Machine Learning Research.

In this work, we apply advanced sampling techniques from statistical physics—specifically, the Wang-Landau algorithm—to map the entire loss landscape of neural networks. Our findings provide a detailed characterization of the density of states across network configurations, shedding light on the interplay between data structure, network architecture, and class imbalance.

This research is the result of a collaborative effort among the teams at the University of Trento, INFN-TIFPA, and the Donders Institute, and we are grateful for the support of our colleagues and institutions.

To read the full paper, please visit our OpenReview page: https://openreview.net/forum?id=BLDtWlFKhn

Learning in neural networks critically hinges on the intricate geometry of the loss landscape associated with a given task. Traditionally, most research has focused on finding specific weight...

07/01/2025

PAPER ALERT!

It is with the utmost pride that we advertise the publication on JPCB of the paper “A Multiscale Analysis of the CzrA Transcription Repressor Highlights the Allosteric Changes Induced by Metal Ion Binding”, by Marta Rigoli, Raffaello Potestio and Roberto Menichetti.

https://pubs.acs.org/doi/full/10.1021/acs.jpcb.4c06106

In this paper we make use of the recently developed mapping entropy optimisation workflow, or MEOW, to investigate the allosteric behaviour of the metal-operated CzrA transcription repressor.

Our approach highlights subtle differences in the behaviour of the molecule between the apo and the zinc-bound state; these features do not manifest on major, large-amplitude rearrangements of the protein structure, hence they are rather difficult to highlight.

The MEOW approach, implemented in the EXCOGITO software suite, attributes to atoms and residues a relevance score based on their dynamical and energetic properties, and allows one to rationalise their functional behaviour.

https://pubs.acs.org/doi/full/10.1021/acs.jcim.4c00490

The method, which is based on concepts from coarse-graining theory, allowed us to make better sense of the allosteric mechanism of CzrA through a novel, intrinsically multi-body analysis of plain MD simulations.

23/07/2024

PAPER ALERT!

It is a monstrous pleasure to announce the publication of “EXCOGITO, an Extensible Coarse-Graining Toolbox for the Investigation of Biomolecules by Means of Low-Resolution Representations”.

https://pubs.acs.org/doi/full/10.1021/acs.jcim.4c00490

Here we illustrate our software to identify highly relevant degrees of freedom of a complex system based on the idea that low-resolution representations of it can be more or less informative. Optimising on them allows one to highlight key correlations and important elements.

The method is illustrated in detail and the software is showcased. We go through the theory of the mapping entropy optimisation workflow (MEOW) and other mapping-based tools. We illustrate them through their application to proteins.

The method is more general though, and it has been applied to other systems, such as binary spin systems, models of the stock market, and Hopfield networks.

https://pubs.acs.org/doi/full/10.1021/acs.jctc.0c00676
https://journals.aps.org/pre/abstract/10.1103/PhysRevE.106.044101
https://arxiv.org/abs/2405.14531

A preprint version of the paper is available on the arXiv at the address https://arxiv.org/abs/2403.08097

18/03/2024

Ready for... work!

On March 8 we held the very informal kick-off meeting of project COMMODORE - Control, optimisation, and modulation of DNA damage through multi-scale modelling of the impact of ionizing radiation. The project, supported by the CARITRO Foundation, aims at characterising the properties of radiation-induced cleavage of the DNA double strand, especially focussing on their modulation through biological features and environmental parameters, e.g. genetic sequence and temperature.

The project is led by prof. Raffaello Potestio from the Physics Dept. and Alessandra Bisio from the CIBIO Dept., and it involves other young researchers from their groups.

https://pressroom.unitn.it/comunicato-stampa/fondazione-caritro-e-unitrento-insieme-la-ricerca

Fondazione Caritro Università di Trento Alessandra Bisio Raffaello Potestio CIBIO - UniTrento

15/03/2024

Today we participated in the Euregio Joint Computational Physics Initiative workshop, gathering researchers from universities of Innsbruck Bozen and Trento.

Our PhD stud's Manuel Micheloni and Margherita Mele delivered great talks!

05/02/2024

On the road again!

Our team has reached Bressanone-Brixen for the 8th Workshop "Physics of Biomolecules: Structure, Dynamics and Function"

https://sites.unimi.it/bressanone/

Looking forward to a four-days-long full immersion of talk, chats, and coffee-fuelled scientific interactions!

15/12/2023

Let it snow!

On December 13, 2023 our group gathered for a social dinner before the winter break. It was a great opportunity to further tighten personal interactions and human connections in our team, and enjoy an easy-going the incipient holiday atmosphere!

29/11/2023

📢📢📢 Postdoctoral position available at the University of Trento - Statistical and Biological Physics group.

More information here:

The postdoctoral position (1.5 Years) is aimed at developing a computational model  of selective polymeric sponges for metal ions capture (more details - call - application page). This research is Funded by the European Union under NextGenerationEU. PRIN 2022RYP9YT_004 "SCOPE: Selective c...

22/11/2023

Splendid job today by Margherita Mele and Roberto Menichetti, who presented their research work at the ALPACA workshop in ECT*!

https://indico.ectstar.eu/event/184/

21/11/2023

Kudos to our (ex 🥲) PhD student Giovanni Mattiotti, who brilliantly graduated yesterday with a fantastic thesis on (many!) multi-scale methods to investigate biological macromolecules! So long Joe! 🥳🥳🥳💪💪💪

Incidentally, Joe's latest preprint is online since a few days here! https://arxiv.org/abs/2311.08076