Recent advances in machine learning (ML) have shown promise in accelerating the discovery of polymers with desired properties by aiding in tasks such as virtual screening via property prediction. However, progress in polymer ML is hampered by the scarcity of high-quality labeled datasets, which are necessary for training supervised ML models. In this work, we study the use of the very recent ‘Joint Embedding Predictive Architecture’ (JEPA), a type of architecture developed for self-supervised learning (SSL), on polymer molecular graphs to understand whether pretraining with the proposed SSL strategy improves downstream performance when labeled data is scarce. We first pretrain our polymer-JEPA model on a large dataset of conjugated copolymer photocatalysts. The pretrained model is then fine-tuned on two distinct downstream tasks: predicting electron affinity in the same chemical space and classifying phase behavior in diblock copolymers, a different chemical space. Our results indicate that JEPA-based self-supervised pretraining enhances downstream performance, particularly when labeled data is very scarce, achieving improvements across both tested datasets. The method provides performance gains in cross-domain fine-tuning, highlighting its potential to extract general knowledge across different classes of polymers. By leveraging large amounts of unlabeled polymer structures for pretraining, the proposed strategy can further reduce the dependence on extensive labeled datasets.

Advancing protein design is crucial for breakthroughs in medicine and biotechnology. Traditional approaches for protein sequence representation often rely solely on the 20 canonical amino acids, limiting the representation of non-canonical amino acids and residues that undergo post-translational modifications. This work explores discrete diffusion models for generating novel protein sequences using the all-atom chemical representation SELFIES. By encoding the atomic composition of each amino acid in the protein, this approach expands the design possibilities beyond standard sequence representations. Using a modified ByteNet architecture within the discrete diffusion D3PM framework, we evaluate the impact of this all-atom representation on protein quality, diversity, and novelty, compared to conventional amino acid-based models. To this end, we develop a comprehensive assessment pipeline to determine whether generated SELFIES sequences translate into valid proteins containing both canonical and non-canonical amino acids. Additionally, we examine the influence of two noise schedules within the diffusion process—uniform (random replacement of tokens) and absorbing (progressive masking)—on generation performance. While models trained on the all-atom representation struggle to consistently generate fully valid proteins, the successfully generated proteins show improved novelty and diversity compared to their amino acid-based model counterparts. Furthermore, the all-atom representation achieves structural foldability results comparable to those of amino acid-based models. Lastly, our results highlight the absorbing noise schedule as the most effective for both representations. Data and code are available at https://github.com/Intelligent-molecular-systems/All-Atom-Protein-Sequence-Generation.

Predicting enzymatic reactions is crucial for applications in biocatalysis, metabolic engineering, and drug discovery, yet it remains a complex and resource-intensive task. Large Language Models (LLMs) have recently demonstrated remarkable success in various scientific domains, e.g., through their ability to generalize knowledge, reason over complex structures, and leverage in-context learning strategies. In this study, we systematically evaluate the capability of LLMs, particularly the Llama-3.1 family (8B and 70B), across three core biochemical tasks: enzyme commission number prediction, forward synthesis, and retrosynthesis. We compare single-task and multitask learning strategies, employing parameter-efficient fine-tuning via LoRA adapters. Additionally, we assess performance across different data regimes to explore their adaptability in low-data settings. Our results demonstrate that fine-tuned LLMs capture biochemical knowledge, with multitask learning enhancing forward- and retrosynthesis predictions by leveraging shared enzymatic information. We also identify key limitations, for example challenges in hierarchical EC classification schemes, highlighting areas for further improvement in LLM-driven biochemical modeling.

The chemical industry needs to undergo a significant transformation towards more sustainable and circular production systems. To guide this transformation, estimating the environmental impacts of chemical production at early product screening or development stages is highly desirable. This study leverages the molecular structure of the process products with graph neural networks (GNNs) for early-stage environmental impact approximation of chemical processes. Specifically, we use end-to-end GNN models to predict fifteen environmental impact categories, utilizing a CarbonMinds dataset of 51,905 processes producing 791 molecules produced in 91 countries, augmented with country-specific energy mix data. Our analysis begins with a comparison of Quantitative Structure-Property Relationship (QSPR) and GNN models for the climate change impact category. Specifically, we develop three different GNN models: (i) GNN with only molecular structure, (ii) GNN with molecular structure and additional geographical features, and (iii) GNN with molecular structure and additional energy mix features. The results indicate that the three GNN models show an improvement over the QSPR models. Furthermore, benchmarking our GNN models against the existing literature in the climate change impact category reveals that our models perform comparably. We then extend our approach by developing both single- and multi-task GNN models to predict all fifteen impact categories. The findings indicate that multi-task learning can improve model performance in complex environmental impact predictions compared to single-task GNNs. Therefore, we recommend using a multi-task GNN for predicting multiple impact categories, with single-task models applied to fine-tune performance on underperforming categories. Although our proposed approach shows improvements over previous models, the prediction of environmental impacts solely based on molecular information remains a rough approximation.

Increasing calls for safer and more sustainable approaches to innovation in the chemical sector necessitate adapted methods for the environmental assessment of emerging chemical technologies. While these technologies are still in the research and development phase, gaining an early understanding of their potential implications is crucial for their eventual introduction into markets worldwide. Life Cycle Assessment (LCA) is a core tool which has been recently adapted for such purpose. Prospective LCA approaches aim to develop plausible future-oriented models which account for the evolution of factors both intrinsic and extrinsic to the technologies assessed. Such future-oriented models introduce many indeterminacies, which could, to some extent, be addressed by Machine Learning techniques. Recent demonstrations of such techniques in the context of prospective LCA, as well as promising avenues for further research, are critically discussed.

ChatGPT is a powerful language model from OpenAI that is arguably able to comprehend and generate text. ChatGPT is expected to greatly impact society, research, and education. An essential step to understand ChatGPT’s expected impact is to study its domain-specific answering capabilities. Here, we perform a systematic empirical assessment of its abilities to answer questions across the natural science and engineering domains. We collected 594 questions on natural science and engineering topics from 198 faculty members across five faculties at Delft University of Technology. After collecting the answers from ChatGPT, the participants assessed the quality of the answers using a systematic scheme. Our results show that the answers from ChatGPT are, on average, perceived as “mostly correct”. Two major trends are that the rating of the ChatGPT answers significantly decreases (i) as the educational level of the question increases and (ii) as we evaluate skills beyond scientific knowledge, e.g., critical attitude.

The estimation of polymer properties is of crucial importance in many domains such as energy, healthcare, and packaging. Recently, graph neural networks (GNNs) have shown promising results for the prediction of polymer properties based on supervised learning. However, the training of GNNs in a supervised learning task demands a huge amount of polymer property data that is time-consuming and computationally/experimentally expensive to obtain. Self-supervised learning offers great potential to reduce this data demand through pre-training the GNNs on polymer structure data only. These pre-trained GNNs can then be fine-tuned on the supervised property prediction task using a much smaller labeled dataset. We propose to leverage self-supervised learning techniques in GNNs for the prediction of polymer properties. We employ a recent polymer graph representation that includes essential features of polymers, such as monomer combinations, stochastic chain architecture, and monomer stoichiometry, and process the polymer graphs through a tailored GNN architecture. We investigate three self-supervised learning setups: (i) node- and edge-level pre-training, (ii) graph-level pre-training, and (iii) ensembled node-, edge- & graph-level pre-training. We additionally explore three different transfer strategies of fully connected layers with the GNN architecture. Our results indicate that the ensemble node-, edge- & graph-level self-supervised learning with all layers transferred depicts the best performance across dataset size. In scarce data scenarios, it decreases the root mean square errors by 28.39% and 19.09% for the prediction of electron affinity and ionization potential compared to supervised learning without the pre-training task.

Large chemical reaction data sets often suffer from incompleteness, such as missing molecules or stoichiometric information. Incomplete chemical reaction equations currently hinder us to perform automated mass balances across large sets of chemical reactions. In this work, we integrate two approaches for computational completion of partial reaction equations. Specifically, we combine a rule-based method and a machine learning model, a tailored version of the pre-trained Molecular Transformer, to complete reactions. The rule-based method takes sets of helper species into a linear solver and therewith balances some incomplete reactions. The machine learning model is trained to take partial reactions as inputs and predicts missing molecules and stoichiometries. We apply our methodology to the USPTO STEREO chemical reaction data set. The rule-based method completes about 50 % of the reactions. The language model shows a top 1 accuracy of 88.3 % on our test set and high validity (> 99 % of outputs are valid SMILES).

The ability to quantify structural changes of the endoplasmic reticulum (ER) is crucial for understanding the structure and function of this organelle. However, the rapid movement and complex topology of ER networks make this challenging. Here, we construct a state-of-the-art semantic segmentation method that we call ERnet for the automatic classification of sheet and tubular ER domains inside individual cells. Data are skeletonized and represented by connectivity graphs, enabling precise and efficient quantification of network connectivity. ERnet generates metrics on topology and integrity of ER structures and quantifies structural change in response to genetic or metabolic manipulation. We validate ERnet using data obtained by various ER-imaging methods from different cell types as well as ground truth images of synthetic ER structures. ERnet can be deployed in an automatic high-throughput and unbiased fashion and identifies subtle changes in ER phenotypes that may inform on disease progression and response to therapy.

Graph neural networks (GNNs) are emerging in chemical engineering for the end-to-end learning of physicochemical properties based on molecular graphs. A key element of GNNs is the pooling function which combines atom feature vectors into molecular fingerprints. Most previous works use a standard pooling function to predict a variety of properties. However, unsuitable pooling functions can lead to unphysical GNNs that poorly generalize. We compare and select meaningful GNN pooling methods based on physical knowledge about the learned properties. The impact of physical pooling functions is demonstrated with molecular properties calculated from quantum mechanical computations. We also compare our results to the recent set2set pooling approach. We recommend using sum pooling for the prediction of properties that depend on molecular size and compare pooling functions for properties that are molecular size-independent. Overall, we show that the use of physical pooling functions significantly enhances generalization.

Designing a simple, yet representative reaction network for subsequent micro-kinetic analysis is important for limiting the cost of evaluation and ensuring model solvability. This is currently achieved by employing sensitivity analysis over a comprehensive reaction network (CRN) to screen reaction species. However, as a reaction network is being simplified for a particular catalyst composition, it loses its transferability to other compositions. Therefore, in this study, a two-way approach is presented to circumvent this problem. Firstly, a generalizable model outcome is identified, i.e. minimum reactant conversions (x_R), based on a mass-flow analysis. Then, a stepwise workflow is developed for constructing a partial reaction network (PRN) to insure transferability of min (x_R) for a range of varying catalyst energetics, in the absence of experimental data for validation. Lastly, the transferability of this approach is demonstrated for CH₄ dry reforming by developing a PRN using Ni(1 1 1) as the initial catalyst and testing it over Ru(0 0 1).

Fuels with high-knock resistance enable modern spark-ignition engines to achieve high efficiency and thus low CO₂ emissions. Identification of molecules with desired autoignition properties indicated by a high research octane number and a high octane sensitivity is therefore of great practical relevance and can be supported by computer-aided molecular design (CAMD). Recent developments in the field of graph machine learning (graph-ML) provide novel, promising tools for CAMD. We propose a modular graph-ML CAMD framework that integrates generative graph-ML models with graph neural networks and optimization, enabling the design of molecules with desired ignition properties in a continuous molecular space. In particular, we explore the potential of Bayesian optimization and genetic algorithms in combination with generative graph-ML models. The graph-ML CAMD framework successfully identifies well-established high-octane components. It also suggests new candidates, one of which we experimentally investigate and use to illustrate the need for further autoignition training data.