Artificial Neural Networks and Machine Learning. ICANN 2025 International Workshops and Special Sessions

Table of Contents

1. Frontmatter

2. 2nd AI in Drug Discovery (AIDD) Workshop

   1. Frontmatter

   2. Early-Stage Discovery in the Era of Hard-To-Drug Targets and Giga-Scale Chemical Spaces

      Dmitri Kireev

      Abstract

      As lead discovery increasingly targets hard-to-drug proteins, the expansion of chemical space presents unprecedented opportunities for hit identification – yet scalable, effective technologies to exploit these vast spaces remain underdeveloped. We describe recent advances to our FRASE-based hit-finding platform (FRASE-bot), including integration of an AI-powered 3D pharmacophore screening across multi-billion-compound libraries, a Hit-Triage Pretrained Transformer (Hit-TPT), and alchemical binding free energy (ABFE) simulations. We also introduce emerging strategies for leveraging phenotypic data to support both hit identification and lead optimization. The platform's utility is demonstrated across several case studies, including our winning entries in CACHE Challenges #1 and #2.

   3. Comparative Analysis of Chemical Structure String Representations for Neural Machine Translation

      Kohulan Rajan, Achim Zielesny, Christoph Steinbeck

      Abstract

      In this work, we present a comparative analysis of SMILES, DeepSMILES, and SELFIES string representations for chemical structures in neural machine translation tasks in cheminformatics. Using transformer-based models, we systematically evaluated their effectiveness in translating between these representations and the corresponding linguistic IUPAC nomenclature. The experimental results demonstrate comparable performance for all three string representations, with SMILES achieving a marginally higher accuracy (99.30% with stereochemical information, 99.21% without) compared to its alternatives. In scaling experiments with 1, 10, and 50 million compounds, the performance differences remained small, though the performance gap narrowed with larger datasets. These findings suggest that researchers can confidently continue using SMILES for neural machine translation tasks with transformers, which benefits from their extensive support in existing chemical libraries, tools, and databases, rather than adopting newer representations. This work has a significant impact on developing more efficient chemical language models in drug discovery, material science, and chemical database curation.

   4. ADMETrix: ADMET-Driven De Novo Molecular Generation

      Nikolaos Mourdoukoutas, Aigli Korfiati, Vasilis Pitsikalis

      Abstract

      We introduce ADMETrix, a de novo drug design framework that combines the generative model REINVENT with ADMET AI, a geometric deep learning architecture for predicting pharmacokinetic and toxicity properties. To our knowledge, this is the first integration enabling real-time generation of small molecules optimized across multiple ADMET endpoints. We evaluate our method in two settings: (i) multi-parameter optimization of ADMET and physicochemical properties, and (ii) scaffold hopping to reduce toxicity while preserving key pharmacophoric features. Using the GuacaMol benchmark, we provide the first systematic evaluation of REINVENT in a multi-objective ADMET context, demonstrating its advantages in generating drug-like, biologically relevant molecules. The code is available at https://github.com/n-mourdou/ADMETrix.

   5. Dimension-Augmented Anisotropy in Graph Neural Diffusion

      Tatiana Sycheva, Maxim Beketov, Ivan Smolyar

      Abstract

      We consider Graph Anisotropic Diffusion (GAD), a recently proposed [4] model of graph neural networks, that can be trained to predict desired properties of the graph by performing learnable diffusion of node features on it. In contrast with similar methods, GAD introduces anisotropy of said diffusion by incorporating filters built from the graph’s Fiedler vector. In present work we attempt to improve this approach by increasing the dimension of the space in which GAD runs – that is, adding filters built from other low-frequency eigenmodes of the graph (eigenvectors of its Laplacian). We report the performance of such “dimension-augmented” GAD in predicting the chemical properties of small organic molecules from the ZINC dataset [5].

   6. Uni-Mol Docking V2: Towards Realistic and Accurate Binding Pose Prediction

      Eric Alcaide, Zhifeng Gao, Guolin Ke, Yaqi Li, Linfeng Zhang, Hang Zheng, Gengmo Zhou

      Abstract

      In recent years, machine learning (ML) methods have emerged as promising alternatives for molecular docking, offering the potential for high accuracy without incurring prohibitive computational costs. However, recent studies have indicated that these ML models may overfit to quantitative metrics while neglecting the physical constraints inherent in the problem. In this work, we present Uni-Mol Docking V2, which demonstrates a remarkable improvement in performance, accurately predicting the binding poses of 77+% of ligands in the PoseBusters benchmark with an RMSD value of less than 2.0 Å, and 75+% passing all quality checks. This represents a significant increase from the 62% achieved by the previous Uni-Mol Docking model. Notably, our Uni-Mol Docking approach generates chemically accurate predictions, circumventing issues such as chirality inversions and steric clashes that have plagued previous ML models. Furthermore, we observe enhanced performance in terms of high-quality predictions (RMSD values of less than 1.0 Åand 1.5 Å) and physical soundness when Uni-Mol Docking is combined with more physics-based methods like Uni-Dock. Our results represent a significant advancement in the application of artificial intelligence for scientific research, adopting a holistic approach to ligand docking that is well-suited for industrial applications in virtual screening and drug design. The code, data and service for Uni-Mol Docking are publicly available for use and further development in https://github.com/dptech-corp/Uni-Mol.

   7. MolEncoder: Improved Masked Language Modeling for Molecules

      Fabian P. Krüger, Nicklas Österbacka, Mikhail Kabeshov, Ola Engkvist, Igor Tetko

      Abstract

      Predicting molecular properties is an important challenge in drug discovery. Machine learning methods, particularly those based on transformer architectures, have become increasingly popular for this task by learning molecular representations directly from chemical structure [1, 2]. Motivated by progress in natural language processing, many recent approaches apply models of the BERT (Bidirectional Encoder Representations from Transformers) architecture [3] to molecular data using SMILES as the input format [4‐9].

      In this study, we revisit core design assumptions that originate in natural language processing but are often carried over to molecular tasks without modification. We explore how variations in masking strategies, pretraining dataset size, and model size influence downstream performance in molecular property prediction.

      Our findings suggest that common practices inherited from natural language processing do not always yield optimal results in this setting. In particular, we observe that increasing the masking ratio can lead to significant improvements, while scaling up the model or dataset size results in stagnating gains despite higher computational cost (Fig. 1). Building on these observations, we develop MolEncoder, a BERT-style model that achieves improved performance on standard benchmarks while remaining more efficient than existing approaches.

      These insights highlight meaningful differences between molecular and textual learning settings. By identifying design choices better suited to chemical data, we aim to support more effective and efficient model development for researchers working in drug discovery and related fields.

   8. Consensus Prediction of Chemical Reactions with OCHEM-R Platform

      Igor V. Tetko, Guillaume Godin, Kevin M. Jablonka, Adrian Mirza, Luc Patiny

      Abstract

      Here we describe the OCHEM-R platform, developed to propose chemical pathways based on a consensus prediction of retrosynthetic pathways using eight different methods. The developed software allows users to visualize predicted reactions and identify similar reactions using retrieval augmented generation (RAG) . The platform is publicly available at https://ochem.eu.

3. Special Session: Neural Networks for Graphs and Beyond

   1. Frontmatter

   2. HeNCler: Node Clustering in Heterophilous Graphs via Learned Asymmetric Similarity

      Sonny Achten, Zander Op de Beeck, Francesco Tonin, Volkan Cevher, Johan A. K. Suykens

      Abstract

      Clustering nodes in heterophilous graphs is challenging as traditional methods assume that effective clustering is characterized by high intra-cluster and low inter-cluster connectivity. To address this, we introduce HeNCler—a novel approach for Heterophilous Node Clustering. HeNCler learns a similarity graph by optimizing a clustering-specific objective based on weighted kernel singular value decomposition Our approach enables spectral clustering on an asymmetric similarity graph, providing flexibility for both directed and undirected graphs. By solving the primal problem directly, our method overcomes the computational difficulties of traditional adjacency partitioning-based approaches. Experimental results show that HeNCler significantly improves node clustering performance in heterophilous graph settings, highlighting the advantage of its asymmetric graph-learning framework.

   3. Visualization and Analysis of the Loss Landscape in Graph Neural Networks

      Samir Moustafa, Lorenz Kummer, Simon Fetzel, Nils M. Kriege, Wilfried N. Gansterer

      Abstract

      Graph Neural Networks (GNNs) are powerful models for graph-structured data, with broad applications. However, the interplay between GNN parameter optimization, expressivity, and generalization remains poorly understood. We address this by introducing an efficient learnable dimensionality reduction method for visualizing GNN loss landscapes, and by analyzing the effects of over-smoothing, jumping knowledge, quantization, sparsification, and preconditioner on GNN optimization. Our learnable projection method surpasses the state-of-the-art PCA-based approach, enabling accurate reconstruction of high-dimensional parameters with lower memory usage. We further show that architecture, sparsification, and optimizer’s preconditioning significantly impact the GNN optimization landscape and their training process and final prediction performance. These insights contribute to developing more efficient designs of GNN architectures and training strategies.

4. Special Session: Neurorobotics

   1. Frontmatter

   2. Pointing-Guided Target Estimation via Transformer-Based Attention

      Luca Müller, Hassan Ali, Philipp Allgeuer, Lukáš Gajdošech, Stefan Wermter

      Abstract

      Deictic gestures, like pointing, are a fundamental form of non-verbal communication, enabling humans to direct attention to specific objects or locations. This capability is essential in Human-Robot Interaction (HRI), where robots should be able to predict human intent and anticipate appropriate responses. In this work, we propose the Multi-Modality Inter-TransFormer (MM-ITF), a modular architecture to predict objects in a controlled tabletop scenario with the NICOL robot, where humans indicate targets through natural pointing gestures. Leveraging inter-modality attention, MM-ITF maps 2D pointing gestures to object locations, assigns a likelihood score to each, and identifies the most likely target. Our results demonstrate that the method can accurately predict the intended object using monocular RGB data, thus enabling intuitive and accessible human-robot collaboration. To evaluate the performance, we introduce a patch confusion matrix, providing insights into the model’s predictions across candidate object locations.

      Code available at: https://github.com/lucamuellercode/MMITF.

   3. Keypoint-Based Diffusion for Robotic Motion Planning on the NICOL Robot

      Lennart Clasmeier, Jan-Gerrit Habekost, Connor Gäde, Philipp Allgeuer, Stefan Wermter

      Abstract

      We propose a novel diffusion-based action model for robotic motion planning. Commonly, established numerical planning approaches are used to solve general motion planning problems, but have significant runtime requirements. By leveraging the power of deep learning, we are able to achieve good results in a much smaller runtime by learning from a dataset generated by these planners. While our initial model uses point cloud embeddings in the input to predict keypoint-based joint sequences in its output, we observed in our ablation study that it remained challenging to condition the network on the point cloud embeddings. We identified some biases in our dataset and refined it, which improved the model’s performance. Our model, even without the use of the point cloud encodings, outperforms numerical models by an order of magnitude regarding the runtime, while reaching a success rate of up to 90% of collision free solutions on the test set.

   4. Real-Time Syllable Recognition in LIBRAS Using Deep Learning for Human-Robot Interaction

      Joelmir Ramos, Nadia Nedjah, Paulo Victor Rorigues de Carvalho

      Abstract

      This work presents a real-time syllable-level recognition system for LIBRAS, the Brazilian Sign Language. The system extracts 2D hand landmarks using MediaPipe and a Gaussian Temporal Smoothing technique to reduce frame-wise jitter. Two deep learning models are implemented for classification: a Multilayer Perceptron (MLP) and a Convolutional Neural Network (CNN). A dataset of 27,456 samples covering all 26 LIBRAS syllables was constructed for training and evaluation. Experiments were conducted on both a desktop workstation and a Raspberry Pi 4 to assess classification accuracy and inference latency. The CNN model achieves an average accuracy of 97.4%, with an inference latency of approximately 50 ms on desktop and 195 ms on Raspberry Pi, meeting the typical requirements for Human-Robot Interaction (HRI) systems. Furthermore, the proposed system was successfully deployed on the humanoid robotic platform 14-bis, demonstrating real-time syllable detection in a practical HRI scenario. These results confirm the feasibility of deploying lightweight LIBRAS classifiers on low-cost embedded platforms, enabling inclusive, scalable, and real-time applications in assistive and educational robotics.

   5. Generating and Customizing Robotic Arm Trajectories Using Neural Networks

      Andrej Lúčny, Matilde Antonj, Carlo Mazzola, Hana Hornáčková, Igor Farkaš

      Abstract

      We introduce a neural network approach for generating and customizing the trajectory of a robotic arm, that guarantees precision and repeatability. To highlight the potential of this novel method, we describe the design and implementation of the technique and show its application in an experimental setting of cognitive robotics. In this scenario, the NICO robot was characterized by the ability to point to specific points in space with precise linear movements, increasing the predictability of the robotic action during its interaction with humans. To achieve this goal, the neural network computes the forward kinematics of the robot arm. By integrating it with a generator of joint angles, another neural network was developed and trained on an artificial dataset created from suitable start and end poses of the robotic arm. Through the computation of angular velocities, the robot was characterized by its ability to perform the movement, and the quality of its action was evaluated in terms of shape and accuracy. Thanks to its broad applicability, our approach successfully generates precise trajectories that could be customized in their shape and adapted to different settings. The code is released at https://github.com/andylucny/nico2/tree/main/generate.

   6. Robotic Calibration Based on Haptic Feedback Improves Sim-to-Real Transfer

      Juraj Gavura, Michal Vavrečka, Igor Farkaš, Connor Gäde

      Abstract

      When inverse kinematics (IK) is adopted to control robotic arms in manipulation tasks, there is often a discrepancy between the end effector (EE) position of the robot model in the simulator and the physical EE in reality. In most robotic scenarios with sim-to-real transfer, we have information about joint positions in both simulation and reality, but the EE position is only available in simulation. We developed a novel method to overcome this difficulty based on haptic feedback calibration, using a touchscreen in front of the robot that provides information on the EE position in the real environment. During the calibration procedure, the robot touches specific points on the screen, and the information is stored. In the next stage, we build a transformation function from the data based on linear transformation and neural networks that is capable of outputting all missing variables from any partial input (simulated/real joint/EE position). Our results demonstrate that a fully nonlinear neural network model performs best, significantly reducing positioning errors.

   7. Towards Bio-inspired Robotic Trajectory Planning via Self-supervised RNN

      Miroslav Cibula, Kristína Malinovská, Matthias Kerzel

      Abstract

      Trajectory planning in robotics is understood as generating a sequence of joint configurations that will lead a robotic agent, or its manipulator, from an initial state to the desired final state, thus completing a manipulation task while considering constraints like robot kinematics and the environment. Typically, this is achieved via sampling-based planners, which are computationally intensive. Recent advances demonstrate that trajectory planning can also be performed by supervised sequence learning of trajectories, often requiring only a single or fixed number of passes through a neural architecture, thus ensuring a bounded computation time. Such fully supervised approaches, however, perform imitation learning; they do not learn based on whether the trajectories can successfully reach a goal, but try to reproduce observed trajectories. In our work, we build on this approach and propose a cognitively inspired self-supervised learning scheme based on a recurrent architecture for building a trajectory model. We evaluate the feasibility of the proposed method on a task of kinematic planning for a robotic arm. The results suggest that the model is able to learn to generate trajectories only using given paired forward and inverse kinematics models, and indicate that this novel method could facilitate planning for more complex manipulation tasks requiring adaptive solutions.

5. 3rd International Workshop on Reservoir Computing

   1. Frontmatter

   2. Impact of Plasticity-Based Reservoir Adaptation on Spectral Radius and Performance of ESNs

      Franziska Weber, Lluís Belanche-Muñoz, Andreas Maier

      Abstract

      Echo state networks (ESNs) are recurrent neural networks belonging to the reservoir computing framework. While ESNs are conceptually simple, their successful application can be challenging. For instance, there are no generally applicable methods for optimally setting important hyperparameters like the reservoir spectral radius. Therefore, the development of strategies for appropriately initializing ESNs is an active field of research. Plasticity-based pretraining is a bio-inspired reservoir optimization approach. We analyze if this approach is able to improve the results of a non-optimized ESN and if the pretraining effects can be explained by the influence on the spectral radius. In our experiments, we evaluate the effects of four synaptic plasticity rules (SP), namely anti-Oja’s, normalized anti-Hebbian, BCM, and dual-threshold BCM, and of intrinsic plasticity (IP) on the Mackey-Glass, NARMA, and Lorenz series. IP significantly improves the ESN’s performance across all three benchmarks whereas this is not the case for any of the considered SP rules. Overall, the influence of plasticity on the spectral radius is not sufficient for explaining the pretraining effects. The cases, in which plasticity significantly worsens the results, however, can be explained by the spectral radius having been moved to a disadvantageous value.

   3. Benchmarking Nonlinear Readouts in Linear Reservoir Networks

      Giacomo Lagomarsini, Andrea Ceni, Claudio Gallicchio

      Abstract

      Recent theoretical advances have demonstrated the universality of linear Reservoir Computing (RC) models equipped with nonlinear readouts, showing their potential to approximate arbitrary input-output mappings. However, practical insights into the selection and performance of nonlinear readouts are limited. This paper addresses this gap by systematically benchmarking a spectrum of nonlinear readouts within linear RC frameworks. Our results reveal the practical trade-offs in accuracy and efficiency across tasks, offering insights on how to train performant RC systems with linear recurrence. These findings provide valuable guidelines for designing efficient recurrent architectures that combine theoretical guarantees with state-of-the-art performance in sequential data processing.

   4. Investigating Time-Scales in Deep Echo State Networks for Natural Language Processing

      Corrado Baccheschi, Alessandro Bondielli, Alessandro Lenci, Alessio Micheli, Lucia Passaro, Marco Podda, Domenico Tortorella

      Abstract

      Reservoir Computing (RC) enables efficiently-trained deep Recurrent Neural Networks (RNNs) by removing the need to train the hierarchy of representations of the input sequences. In this paper, we analyze the performance and the dynamical behavior of RC models, specifically Deep Bidirectional Echo State Networks (Deep-BiESNs), applied to Natural Language Processing (NLP) tasks. We compare the performance of Deep-BiESNs against fully-trained NLP baseline models on six common NLP tasks: three sequence-to-vector tasks for sequence-level classification and three sequence-to-sequence tasks for token-level labeling. Experimental results demonstrate that Deep-BiESNs achieve comparable or superior performance to these baseline models. We then adapt the class activation mapping technique for explainability to analyze the dynamical properties of these deep RC models, highlighting how the hierarchy of representations in Deep-BiESNs layers contributes to forming the class prediction in the different NLP tasks. Investigating time scales in deep RNN layers is highly relevant for NLP because language inherently involves dependencies that occur over various temporal horizons. The findings not only underscore the potential of Deep ESNs as a competitive and efficient alternative for NLP applications, but also contribute to a deeper understanding of how to effectively model such architectures for addressing other NLP challenges.

   5. A Spectral Interpretation of Redundancy in a Graph Reservoir

      Anna Bison, Alessandro Sperduti

      Abstract

      Reservoir computing has been successfully applied to graphs as a preprocessing method to improve the training efficiency of Graph Neural Networks (GNNs). However, a common issue that arises when repeatedly applying layer operators on graphs is over-smoothing, which consists in the convergence of graph signals toward low-frequency components of the graph Laplacian. This work revisits the definition of the reservoir in the Multiresolution Reservoir Graph Neural Network (MRGNN), a spectral reservoir model, and proposes a variant based on a Fairing algorithm originally introduced in the field of surface design in computer graphics. This algorithm provides a pass-band spectral filter that allows smoothing without shrinkage, and it can be adapted to the graph setting through the Laplacian operator. Given its spectral formulation, this method naturally connects to GNN architectures for tasks where smoothing, when properly controlled, can be beneficial, such as graph classification. The core contribution of the paper lies in the theoretical analysis of the algorithm from a random walks perspective. In particular, it shows how tuning the spectral coefficients can be interpreted as modulating the contribution of redundant random walks. Exploratory experiments based on the MRGNN architecture illustrate the potential of this approach and suggest promising directions for future research.

   6. Shaping Attractor Landscapes in Boolean Liquid State Machines via STDP and Global Plasticity

      Jérémie Cabessa, Alessandro E. P. Villa

      Abstract

      Small Boolean Liquid State Machines (B-LSMs) offer a simplified yet expressive biologically inspired model of recurrent computation, in which network attractor dynamics can be systematically analyzed. In their untrained form, B-LSMs exhibit complex, often chaotic dynamics with short-lived memory traces. This study investigates how local synaptic plasticity (STDP) and a global plasticity (GP) mechanism jointly shape the attractor landscapes of these networks. Specifically, we show that synaptic modifications can drive B-LSMs to exhibit exponentially many attractors, each corresponding to a potential memory. Such high attractor regimes are attainable through global synaptic crafting. Under noisy background conditions, STDP tends to drive the networks back to low attractor regimes; however, when receiving carefully designed inputs, STDP maintains the networks’ rich attractor dynamics. Overall, our findings highlight the theoretical potential for storing an impressive number of memories in recurrent neural networks, with significant implications for theoretical neuroscience and neuromorphic computing.

6. Backmatter