Deep Force, last year’s winning hackathon project, returns with the goal of further improving the classification of single-molecule force spectroscopy data. By refining our auto-encoder-inspired machine learning models developed during last year’s Hackathon, we aim to facilitate force curve analysis for the community. This year, we will not only optimize our approach but also benchmark our model against existing software solutions to assess performance, accuracy, and usability. We will first optimize last year’s best models by improving simulations, exploring semi-supervised strategies, and fine-tuning hyperparameters to boost classification accuracy and robustness. Next, we will benchmark our models against established force spectroscopy analysis tools, evaluating performance in terms of accuracy, processing speed, and usability, and identifying areas for further improvement. Finally, we will refine GUI to provide smoother interaction, add features for easier manual correction and re-training, optimize handling of large datasets, and prepare comprehensive documentation and tutorials to support wider adoption. Together, these efforts aim to deliver a powerful, user-friendly tool that significantly advances the analysis of force spectroscopy data and accelerates discoveries in molecular biophysics.

Project lead: Jure Majnik

Inhibitory neurons are the second most abundant cell type in the adult brain and one of their main roles is to “put a break” on neural activity. Indeed, impairments to inhibitory cells can lead to overexcitation that plays a key role in neurological disease such as for example epilepsy. The precise balance between excitatory and inhibitory cells is thus crucial for healthy brain function.

In mammals, the balance between excitatory and inhibitory cells gets established during postnatal development and comprises three steps (see illustration): 1) Migration, 2) Cell death and 3) Establishment of synaptic connections. This hackathon project will focus on the quantification of the central step - inhibitory cell death, using 2 photons fluorescence imaging volumes acquired daily (close to 1mm3, thousands of cells, see illustration). The  goal of the project will be to detect cell death by accurately tracking neurons across days in 3D and analyze features that are predictive of an interneuron’s death (or survival) - such as for example its anatomical location, morphology or activity. Besides the application in this project, the developed code library will be useful for future projects studying early brain development using longitudinal volumetric imaging.

Group leader: Felix Rico (Dynamo)

Mechanobiology is an emerging field focused on understanding how mechanical forces influence biological function, from molecules to tissues, in both health and disease.  Atomic force microscopy (AFM) is one of the most widely used, robust and standardized method for mechanical tests at the nano and microscales. However, AFM data analysis remains a major bottleneck, as it requires processing of thousands of individual recordings (force curves) per map. This involves fitting a viscoelastic contact model to each individual force curve that, although automated, remains computationally intensive, taking hours to analyze each map.

This project aims to address this computational challenge. Participants will build upon an existing Python-based analysis tool (pyFMlab) to develop methods that substantially accelerate the curve-fitting process. The project will explore and compare strategies ranging from conventional optimization techniques (e.g., parallel computing) to advanced machine learning approaches. Enhancements to the user interface will also play a crucial role in streamlining interaction with the dataset and improving usability.

Participants will work with both experimental AFM mechanical maps and simulated force curves. The expected outcome is a reduction of analysis time by at least an order of magnitude, enabling faster and more scalable interpretation of AFM data in mechanobiology.

Project leads: V. Marquioni-Monteiro, F.Bansept, F.Labourel, L. Espinosa

Keywords: Modeling, Trafficking, Random Process

Biological membranes are crucial gatekeepers of the cell, balancing the need to import vital molecules like glucose with the imperative to block harmful agents such as toxins or viruses. To optimize this selective uptake, cells embed transmembrane proteins that act as molecular “gateways.” But there’s a catch: too many proteins can overcrowd the membrane, potentially impairing its function.This project explores membrane crowding using a physical process called random sequential adsorption (RSA)—a model where objects are randomly added to a surface only if they don’t overlap. Here, we reinterpret this model biologically, treating each protein as a “hole” drilled into the membrane, while accounting for two crucial biological realities: protein degradation and varying membrane geometries (e.g., curved or irregular cell shapes). Participants will simulate these dynamics using computational models and analyze how turnover and shape influence overall membrane packing. The dataset involves synthetic membrane surfaces and stochastic rules for adsorption and degradation. Through modeling and visualization, the outcome will shed light on how cells might tune their membrane composition over evolutionary time.

Project lead: Tom Orjollet--Lacomme

To better understand how mice learn, we designed an experiment where a mouse in an arena is rewarded according to its behaviour. More specifically, the mouse is moving freely in the arena and if it performs the correct set of actions it will be rewarded with water. This would enable us to answer questions such as:

• Can a mouse learn the correct behaviour to trigger the reward?
• What behaviours will the mouse exhibit during its learning phase?

To run the experiment, we have set up the arena and a camera recording the mouse from below. We now need a way to track the mouse in real time such that the rewards are triggered at the right time and place.

Moreover, we want to quantitatively analyse the mouse movements and actions in the arena to better characterize its behaviour.

We have put in place a codebase in Python mainly relying on the library OpenCV to do the detection and tracking of the mouse. However, for now, our code suffers from the following issues:

1. The mouse can sometimes be lost by our tracker.
2. The tracking of the mouse is not working properly in the first seconds of the recording.
3. The camera lens distorts the image which biases the shape of mouse trajectory.

Because of these problems, we often see that::

• The reward is not correctly triggered
• The behaviour analysis is incorrect

Finally, the code has to run at good time efficiency so that it can be run in realtime in a closed-loop system.

As mentioned above, we have a pre-existing code that the hacker would be able to use or start from scratch if they prefer.

Project Leaders: Florian Saby et Aitor González

Abstract:

Going through the relevant literature of a given field can sometimes be cumbersome. Finding that paper you saw the other time in which there is that specific result can take way too much time.

In this project we want to develop a python tool that uses deep learning and large language models to help answering these problems and many more!

Currently we are focused on developing algorithms to search bibliography from natural queries. We also want to automatically classify the references of a given bibliography. But, we are open to any ideas about bibliography manipulation and exploration. The input data will be Bibtex files with abstracts and PDF files sets.