Projects
Active Projects
Applying cutting-edge ML methods to accelerate compound screening, perform in-silico predictions of pharmacokinetic properties and toxicity of drug candidates, improve drug efficacy and effectiveness, and design new drug candidates.
More about NEDD
The Neuro-explicit AI for Drug Discovery (NEDD) project was launched in 2022 at Saarland University with the goal of combining the expertise of the Saarland Informatics Campus (SIC) with the Helmholtz Institute for Pharmaceutical Research Saarland (HIPS). The novel interdisciplinary collaboration across campus integrates pharmaceutical research with the latest advances in the area of machine learning.
Our mission is to apply cutting-edge machine learning methods to accelerate compound screening, perform in-silico predictions of pharmacokinetic properties and toxicity of drug candidates, improve drug efficacy and effectiveness, and design new drug candidates.
Integrating expert and domain knowledge into deep reinforcement learning agents through multi-level abstractions and causal modeling. We want to foster interpretable, more robust DRL agents that can generalize across environments.
More about MAC-MERLin
MAC-MERLin aims to integrate established expert and domain knowledge into artificial neural networks by leveraging multi-level abstractions in combination with causal modeling. Building on the established concepts of causality and abstraction, the project improves the interpretability of Deep Reinforcement Learning (DRL) agents and aligns them more closely with human understanding and expert insights, while at the same time making them more robust, better able to generalize across environments, and better tuned to real-world complexity.
The project addresses four fundamental challenges that currently make the practical deployment of neural-network decision making difficult: modeling, robustness of DRL agents, their ability to generalize to new environments, and their interpretability. By closing the gap between the theoretical and practical aspects of DRL, MAC-MERLin contributes to applications across smart cities and power grids, intelligent supply chains and production systems, autonomous driving, drug discovery, medical treatment and bioengineering.
Laying the groundwork for the CERTAIN Centre—a Saarland-anchored, European-scale hub for Trusted AI that delivers functional guarantees and earns user trust even in safety-critical applications.
More about ToCERTAIN
The ToCERTAIN project aims to strengthen the development of trustworthy artificial intelligence (AI) that offers guarantees about its functionality (Trusted AI) and thus creates the necessary trust among users for the widespread, secure and reliable use of AI, even in critical applications. The current state of research and application of AI systems is still very inadequate in this regard, although industry is actively demanding these characteristics and corresponding certification.
The multitude of aspects that contribute to trustworthy artificial intelligence are being developed in part at DFKI, but above all in close and intensive cooperation with partners from research and industry – at the site, in the greater region and throughout Europe. This collaboration is to be intensified and consolidated in the long term, with ToCERTAIN laying the foundation as an ambitious research project and establishing a highly visible beacon in Europe and beyond with the CERTAIN Centre and its closely networked partners.
Completed Projects
Novel methods for the analysis of stochastic chemical reaction networks that exhibit multimodal behaviour.
More about MULTIMODE
In MULTIMODE we develop novel methods for the analysis of stochastic chemical reaction networks. We mostly focus on networks that exhibit multimodal behaviour and a countable finite number of modes — regions of the state space in which the process spends most of its time. Multimodality plays a central role in probabilistic decisions found in cellular processes, which can explain the phenotypic variability of monoclonal cells.
The analysis of such models entails several interesting tasks:
- the location and relative strength of the modes,
- a model decomposition along the system modes, and
- reconstruction of the mode-conditioned distributions.
Deep reinforcement learning, trained by self-play on a discrete-event digital twin, to control the line buffer of the Ford Saarlouis assembly plant in real time.
More about MoDigPro
MoDigPro — Modernisierung und Digitalisierung der Produktionssteuerung des Ford Werkes in Saarlouis zur störungsfreien kontinuierlichen Produktion durch multidimensionale Algorithmen und künstliche Intelligenz — was a cooperation with the Ford Motor Company and the Max Planck Institute for Informatics.
In a car manufacturing process, the order of production is crucial for every production step; especially for the final assembly, where all pre-produced parts of the car are put together.
While more recent factories have a more modern structure, such that there is a generous number of options to optimize this production order, in the Ford factory in Saarlouis, the options to optimize the production order are limited by the line buffer, which can store and take out cars from one of 27 different lines. Currently, the optimization in Saarlouis is done manually by following some rules of thumbs.
In this project, we compensate for this structural disadvantage with intelligent control of the line buffer. We use deep reinforcement learning to train neural networks by self-play. These neural networks are supposed to take over control of the line buffer.
Since deep reinforcement learning is known to be data hungry, training the neural networks while they are deployed in the factory is not an option. Instead, we use discrete event simulation to construct a digital twin of the line buffer and the final assembly. We can use this simulation to train the agents for optimizing car production in Saarlouis.
- Complex Networks
Analysis and Simulation of Complex Networks
Stochastic spreading processes on large complex networks: efficient simulation, analysis and control of the emerging dynamics, with applications from epidemics to neural relational inference.
More about Complex Networks
We consider stochastic spreading processes on large complex networks and develop novel methods for efficient simulation, analysis, and control of the emerging dynamics. Topics include rejection-based simulation of non-Markovian agents, reductions to Markov population models, mean-field and pair approximations, and learning vaccine allocation strategies from simulation.
A central line of work is the COVID-19 case-study programme: how contact structure, overdispersion and individual variation shape the spread of epidemics, and how mobility restrictions interact with overdispersion.
- DNA Methylation
Stochastic Modelling of DNA Methylation Dynamics
Stochastic models of epigenetic modifications of the DNA.
More about DNA Methylation
Together with the group of Jörn Walter we develop stochastic models that describe epigenetic modifications of the DNA. We focus on DNA methylation, which plays an important role during embryogenesis and is linked to certain diseases.
We fit hidden Markov models to (oxidative) hairpin bisulfite sequencing data and analyse active and passive demethylation. A central goal is an improved understanding of the role of the Tet enzymes, their interplay with the chromatin structure, and other methylation enzymes.