Molecular design encompasses tasks ranging from de-novo design to structural alteration of given molecules or fragments. For the latter, state-of-the-art methods predominantly function as "Instance Optimizers'', expending significant compute restarting the search for every input structure. While model-based approaches theoretically offer amortized efficiency by learning a policy transferable to unseen structures, existing methods struggle to generalize. We identify a key failure mode: the high variance arising from the heterogeneous difficulty of distinct starting structures. To address this, we introduce GRXForm, adapting a pre-trained Graph Transformer model that optimizes molecules via sequential atom-and-bond additions. We employ Group Relative Policy Optimization (GRPO) for goal-directed fine-tuning to mitigate variance by normalizing rewards relative to the starting structure. Empirically, GRXForm generalizes to out-of-distribution molecular scaffolds without inference-time oracle calls or refinement, achieving scores in multi-objective optimization competitive with leading instance optimizers.

Generative molecular optimization aims to design molecules with properties surpassing those of existing compounds. However, such candidates are rare and expensive to evaluate, yielding sample efficiency essential. Additionally, surrogate models introduced to predict molecule evaluations, suffer from distribution shift as optimization drives candidates increasingly out-of-distribution. To address these challenges, we introduce Joint Self-Improvement, which benefits from (i) a joint generative-predictive model and (ii) a self-improving sampling scheme. The former aligns the generator with the surrogate, alleviating distribution shift, while the latter biases the generative part of the joint model using the predictive one to efficiently generate optimized molecules at inference-time. Experiments across offline and online molecular optimization benchmarks demonstrate that Joint Self-Improvement outperforms state-of-the-art methods under limited evaluation budgets.

Protein sequence optimization under tight oracle budgets requires methods that explore vast combinatorial spaces while making each evaluation informative. Existing reinforcement learning and off-policy generative approaches often degrade under surrogate noise, and position-agnostic mutation proposals risk disrupting functionally critical residues. We introduce SILO, a trajectory-level self-improvement imitation framework for oracle-budgeted protein design. SILO uses a hierarchical edit policy that decomposes each mutation into a position choice followed by a residue choice. In each active-learning round, the policy samples candidate trajectories via incremental stochastic beam search without replacement (SBS), and a UCB-based proxy ensemble, combined with an alanine-scan fitness score (AFS), selects candidates with functionally relevant edits for in silico oracle evaluation. The policy is then updated by next-action cross-entropy imitation on the round's best oracle-labeled trajectories, avoiding value-function estimation. Across eight reproduced protein fitness landscapes and five strong baselines from prior work, SILO achieves the highest maximum and top-100 mean fitness on 8 of 8 landscapes within our evaluations, often exhibiting faster early-stage improvement. In low-data and noisy-proxy stress tests on two landscapes per setting, SILO remains competitive or best when several baselines degrade. Ablations show that SBS with AFS account for much of the gains, with iterative imitation providing additional improvement. Code is available at: https://github.com/grimmlab/SILO.git

Current methods for end-to-end constructive neural combinatorial optimization usually train a policy using behavior cloning from expert solutions or policy gradient methods from reinforcement learning. While behavior cloning is straightforward, it requires expensive expert solutions, and policy gradient methods are often computationally demanding and complex to fine-tune. In this work, we bridge the two and simplify the training process by sampling multiple solutions for random instances using the current model in each epoch and then selecting the best solution as an expert trajectory for supervised imitation learning. To achieve progressively improving solutions with minimal sampling, we introduce a method that combines round-wise Stochastic Beam Search with an update strategy derived from a provable policy improvement. This strategy refines the policy between rounds by utilizing the advantage of the sampled sequences with almost no computational overhead. We evaluate our approach on the Traveling Salesman Problem and the Capacitated Vehicle Routing Problem. The models trained with our method achieve comparable performance and generalization to those trained with expert data. Additionally, we apply our method to the Job Shop Scheduling Problem using a transformer-based architecture and outperform existing state-of-the-art methods by a wide margin.

Generative deep learning has become pivotal in molecular design for drug discovery, materials science, and chemical engineering. A widely used paradigm is to pretrain neural networks on string representations of molecules and fine-tune them using reinforcement learning on specific objectives. However, string-based models face challenges in ensuring chemical validity and enforcing structural constraints like the presence of specific substructures. We propose to instead combine graph-based molecular representations, which can naturally ensure chemical validity, with transformer architectures, which are highly expressive and capable of modeling long-range dependencies between atoms. Our approach iteratively modifies a molecular graph by adding atoms and bonds, which ensures chemical validity and facilitates the incorporation of structural constraints. We present GraphXForm, a decoder-only graph transformer architecture, which is pretrained on existing compounds and then fine-tuned using a new training algorithm that combines elements of the deep cross-entropy method and self-improvement learning. We evaluate GraphXForm on various drug design tasks, demonstrating superior objective scores compared to state-of-the-art molecular design approaches. Furthermore, we apply GraphXForm to two solvent design tasks for liquid-liquid extraction, again outperforming alternative methods while flexibly enforcing structural constraints or initiating design from existing molecular structures.

The procurement and production planning of horticultural production and retail companies faces many uncertainties, including seasonality and perishability, and is often organized through tactical pre-order and operational re-order planning. We present a stochastic model formulation for this problem and develop a deterministic mixed-integer linear programming (MILP) approximation to determine pre-order quantities, taking uncertain re-order opportunities into account, as well as a newsvendor adaptation for deciding about short-term re-orders. In doing this we consider the typical characteristics of horticultural products and their sales season, and integrate an advanced machine learning technique to factor adequate forecasts into the solution approach. Our model considers target alpha- and beta-service levels, uncertain and limited re-order options with specific costs, and minimum re-order shares. Reflecting the perishability of the products focused, we track the age distribution of stock. To evaluate the results, we set up a simulation comparing our modeling approach with a practical and a literature benchmark using actual data from three case companies. Additionally, we provide sensitivity analyses using a large set of varied scenarios to derive further managerial insights. We show that our approach outperforms the benchmarks in terms of profit and is able to significantly reduce product waste. It is also able to meet target service levels while providing robust solutions that maintain flexibility for in-season adaptations.

Process synthesis in chemical engineering is a complex planning problem due to vast search spaces, continuous parameters and the need for generalization. Deep reinforcement learning agents, trained without prior knowledge, have shown to outperform humans in various complex planning problems in recent years. Existing work on reinforcement learning for flowsheet synthesis shows promising concepts. We further develop those concepts and present a general deep reinforcement learning approach for flowsheet synthesis. We demonstrate the adaptability of an agent to the general task of separating binary azeotropic mixtures. The agent is trained to setup the discrete process topology alongside choosing continuous specifications for the individual flowsheet elements (e.g., distillation columns and recycles). Without prior knowledge, it learns within one training cycle to craft flowsheets for multiple chemical systems, considering different feed compositions and conceptual approaches. The agent discovers autonomously fundamental process engineering paradigms as heteroazeotropic distillation or curved-boundary distillation.

Linear mixed models (LMMs) are a commonly used method for genome-wide association studies (GWAS) that aim to detect associations between genetic markers and phenotypic measurements in a population of individuals while accounting for population structure and cryptic relatedness. In a standard GWAS, hundreds of thousands to millions of statistical tests are performed, requiring control for multiple hypothesis testing. Typically, static corrections that penalize the number of tests performed are used to control for the family-wise error rate, which is the probability of making at least one false positive. However, it has been shown that in practice this threshold is too conservative for normally distributed phenotypes and not stringent enough for non-normally distributed phenotypes. Therefore, permutation-based LMM approaches have recently been proposed to provide a more realistic threshold that takes phenotypic distributions into account. In this work, we discuss the advantages of permutation-based GWAS approaches, including new simulations and results from a re-analysis of all publicly available Arabidopsis phenotypes from the AraPheno database.

Machine learning methods for extracting patterns from high-dimensional data are very important in the biological sciences. However, in certain cases, real-world applications cannot confirm the reported prediction performance. One of the main reasons for this is data leakage, which can be seen as the illicit sharing of information between the training data and the test data, resulting in performance estimates that are far better than the performance observed in the intended application scenario. Data leakage can be difficult to detect in biological datasets due to their complex dependencies. With this in mind, we present seven questions that should be asked to prevent data leakage when constructing machine learning models in biological domains. We illustrate the usefulness of our questions by applying them to nontrivial examples. Our goal is to raise awareness of potential data leakage problems and to promote robust and reproducible machine learning-based research in biology.

Motivation:Permutation-based significance thresholds have been shown to be a robust alternative to classical Bonferroni significance thresholds in genome-wide association studies (GWAS) for skewed phenotype distributions. The recently published method permGWAS introduced a batch-wise approach to efficiently compute permutation-based GWAS. However, running multiple univariate tests in parallel leads to many repetitive computations and increased computational resources. More importantly, traditional permutation methods that permute only the phenotype break the underlying population structure. Results:We propose permGWAS2, an improved method that does not break the population structure during permutations and uses an elegant block matrix decomposition to optimize computations, thereby reducing redundancies. We show on synthetic data that this improved approach yields a lower false discovery rate for skewed phenotype distributions compared to the previous version and the commonly used Bonferroni correction. In addition, we re-analyze a dataset covering phenotypic variation in 86 traits in a population of 615 wild sunflowers (Helianthus annuus L.). This led to the identification of dozens of novel associations with putatively adaptive traits, and removed several likely false-positive associations with limited biological support. Availability and implementation:permGWAS2 is open-source and publicly available on GitHub for download: https://github.com/grimmlab/permGWAS.

Accurately forecasting demand is a potential competitive advantage, especially when dealing with perishable products. The multi-billion dollar horticultural industry is highly affected by perishability, but has received limited attention in forecasting research. In this paper, we analyze the applicability of general compared to dataset-specific predictors, as well as the influence of external information and online model update schemes. We employ a heterogeneous set of horticultural data, three classical, and twelve machine learning-based forecasting approaches. Our results show a superiority of multivariate machine learning methods, in particular the ensemble learner XGBoost. These advantages highlight the importance of external factors, with the feature set containing statistical, calendrical, and weather-related features leading to the most robust performance. We further observe that a general model is unable to capture the heterogeneity of the data and is outperformed by dataset-specific predictors. Moreover, frequent model updates have a negligible impact on forecasting quality, allowing long-term forecasting without significant performance degradation.

Process synthesis in chemical engineering is a complex planning problem due to vast search spaces, continuous parameters and the need for generalization. Deep reinforcement learning agents, trained without prior knowledge, have shown to outperform humans in various complex planning problems in recent years. Existing work on reinforcement learning for flowsheet synthesis shows promising concepts. We further develop those concepts and present a general deep reinforcement learning approach for flowsheet synthesis. We demonstrate the adaptability of an agent to the general task of separating binary azeotropic mixtures. The agent is trained to set up the discrete process topology alongside choosing continuous specifications for the individual flowsheet elements (e.g., distillation columns and recycles). Without prior knowledge, it learns within one training cycle to craft flowsheets for multiple chemical systems, considering different feed compositions and conceptual approaches. The agent discovers autonomously fundamental process engineering paradigms as heteroazeotropic distillation or curved-boundary distillation.

Sustainable weed management strategies are critical to feeding the world’s population while preserving ecosystems and biodiversity. Therefore, site-specific weed control strategies based on automation are needed to reduce the additional time and effort required for weeding. Machine vision-based methods appear to be a promising approach for weed detection, but require high quality data on the species in a specific agricultural area. Here we present a dataset, the Moving Fields Weed Dataset (MFWD), which captures the growth of 28 weed species commonly found in sorghum and maize fields in Germany. A total of 94,321 images were acquired in a fully automated, high-throughput phenotyping facility to track over 5,000 individual plants at high spatial and temporal resolution. A rich set of manually curated ground truth information is also provided, which can be used not only for plant species classification, object detection and instance segmentation tasks, but also for multiple object tracking.

The constructive approach within Neural Combinatorial Optimization (NCO) treats a combinatorial optimization problem as a finite Markov decision process, where solutions are built incrementally through a sequence of decisions guided by a neural policy network. To train the policy, recent research is shifting toward a 'self-improved' learning methodology that addresses the limitations of reinforcement learning and supervised approaches. Here, the policy is iteratively trained in a supervised manner, with solutions derived from the current policy serving as pseudo-labels. The way these solutions are obtained from the policy determines the quality of the pseudo-labels. In this paper, we present a simple and problem-independent sequence decoding method for self-improved learning based on sampling sequences without replacement. We incrementally follow the best solution found and repeat the sampling process from intermediate partial solutions. By modifying the policy to ignore previously sampled sequences, we force it to consider only unseen alternatives, thereby increasing solution diversity. Experimental results for the Traveling Salesman and Capacitated Vehicle Routing Problem demonstrate its strong performance. Furthermore, our method outperforms previous NCO approaches on the Job Shop Scheduling Problem.

Summary: Time series forecasting is a research area with applications in various domains, nevertheless without yielding a predominant method so far. We present ForeTiS, a comprehensive and open source Python framework that allows rigorous training, comparison, and analysis of state-of-the-art time series forecasting approaches. Our framework includes fully automated yet configurable data preprocessing and feature engineering. In addition, we use advanced Bayesian optimization for automatic hyperparameter search. ForeTiS is easy to use, even for non-programmers, requiring only a single line of code to apply state-of-the-art time series forecasting. Various prediction models, ranging from classical forecasting approaches to machine learning techniques and deep learning architectures, are already integrated. More importantly, as a key benefit for researchers aiming to develop new forecasting models, ForeTiS is designed to allow for rapid integration and fair benchmarking in a reliable framework. Thus, we provide a powerful framework for both end users and forecasting experts.Availability: ForeTiS is available at https://github.com/grimmlab/ForeTiS. We provide a setup using Docker, as well as a Python package at https://pypi.org/project/ForeTiS/. Extensive online documentation with hands-on tutorials and videos can be found at https://foretis.readthedocs.io.

The determination of liquid phase equilibria plays an important role in chemical process simulation. This work presents a generalization of an approach called the convex envelope method (CEM), which constructs all liquid phase equilibria over the whole composition space for a given system with an arbitrary number of components. For this matter, the composition space is discretized and the convex envelope of the Gibbs energy graph is computed. Employing the tangent plane criterion, all liquid phase equilibria can be determined in a robust way. The generalized CEM is described within a mathematical framework and it is shown to work numerically with various examples of up to six components from the literature.

Motivation Permutation-based significance thresholds have been shown to be a robust alternative to Bonferroni-based significance thresholds in genome-wide association studies (GWAS). However, the implementation of permutation-based thresholds is computationally demanding. The recently published method permGWAS introduced a batch-wise approach using 4D tensors to efficiently compute permutation-based GWAS. However, running multiple univariate tests in parallel leads to many repetitive computations and increased computational resources. More importantly, the previous version of permGWAS does not take into account the population structure when permuting the phenotype. Results We propose permGWAS2 , an improved and accelerated version that uses a block matrix decomposition to optimize computations, thereby reducing redundant computations. It also introduces an alternative permutation strategy that takes into account the population structure during permutation. We show that this improved framework provides a more streamlined approach to performing permutation-based GWAS with a lower false discovery rate compared to the previous version and the traditional Bonferroni correction. Availability permGWAS2 is open-source and publicly available on GitHub for download: https://github.com/grimmlab/permGWAS .

AbstractSummaryPredicting complex traits from genotypic information is a major challenge in various biological domains. With easyPheno, we present a comprehensive Python framework enabling the rigorous training, comparison and analysis of phenotype predictions for a variety of different models, ranging from common genomic selection approaches over classical machine learning and modern deep learning-based techniques. Our framework is easy-to-use, also for non-programming-experts, and includes an automatic hyperparameter search using state-of-the-art Bayesian optimization. Moreover, easyPheno provides various benefits for bioinformaticians developing new prediction models. easyPheno enables to quickly integrate novel models and functionalities in a reliable framework and to benchmark against various integrated prediction models in a comparable setup. In addition, the framework allows the assessment of newly developed prediction models under pre-defined settings using simulated data. We provide a detailed documentation with various hands-on tutorials and videos explaining the usage of easyPheno to novice users.Availability and implementationeasyPheno is publicly available at https://github.com/grimmlab/easyPheno and can be easily installed as Python package via https://pypi.org/project/easypheno/ or using Docker. A comprehensive documentation including various tutorials complemented with videos can be found at https://easypheno.readthedocs.io/.Supplementary informationSupplementary data are available at Bioinformatics Advances online.

Assigning inpatients to hospital beds impacts patient satisfaction and the workload of nurses and doctors. The assignment is subject to unknown inpatient arrivals, in particular for emergency patients. Hospitals, therefore, need to deal with uncertainty on actual bed requirements and potential shortage situations as bed capacities are limited. This paper develops a model and solution approach for solving the patient bed-assignment problem that is based on a machine learning (ML) approach to forecasting emergency patients. First, it contributes by improving the anticipation of emergency patients using ML approaches, incorporating weather data, time and dates, important local and regional events, as well as current and historical occupancy levels. Drawing on real-life data from a large case hospital, we were able to improve forecasting accuracy for emergency inpatient arrivals. We achieved up to 17% better root mean square error (RMSE) when using ML methods compared to a baseline approach relying on averages for historical arrival rates. We further show that the ML methods outperform time series forecasts. Second, we develop a new hyper-heuristic for solving real-life problem instances based on the pilot method and a specialized greedy look-ahead (GLA) heuristic. When applying the hyper-heuristic in test sets we were able to increase the objective function by up to 5.3% in comparison to the benchmark approach in [ 40 ]. A benchmark with a Genetic Algorithm shows also the superiority of the hyper-heuristic. Third, the combination of ML for emergency patient admission forecasting with advanced optimization through the hyper-heuristic allowed us to obtain an improvement of up to 3.3% on a real-life problem.

Supplementary Table 1. Clinical characteristics of patients included in the study (T790M positive/negative and the different pools of plasma from healthy donors); Supplementary Table 2. Mutant admixtures used as part of the analytical validation. These samples were run using three technical replicates; Supplementary Table 3. Analytical validation. Supplementary Table 4. Summary of patient demographics and T790M results for the validation cohort; Supplementary Figure 1. Patient histology for T790M positive and negative patients (A) and disease distribution by stage (B) for all the clinical samples.