The challenge of bitter peptides is particularly acute in the growing market for plant-based protein foods. As manufacturers turn to peas, soy, and other plant proteins to replace animal products, the enzymatic breakdown of those proteins during processing generates peptide mixtures whose sensory profiles are difficult to anticipate. Undesirable bitter notes are a frequent reason that plant-based alternatives fail to win over consumers. At the same time, bitter peptides are not simply a defect to be eliminated; many of them carry physiological properties and may, for example, play a role in regulating hunger and satiety. Understanding which sequences taste bitter, and why, therefore has value both for improving flavor and for designing foods with functional benefits.

Antonella Di Pizio, principal investigator of the new study and head of the Molecular Modeling research group at Leibniz-LSB@TUM, underscores the broader stakes of the work. To make plant-based protein sources more attractive for food production and to use them more sustainably, she explains, researchers need to understand which peptides taste bitter and what structural features characterize them. AI-based methods, she argues, can make an important contribution to exactly that understanding. Her team, which also included researchers from the Technical University of Munich and Pompeu Fabra University in Barcelona, set out to build a computational pipeline that could move the field beyond slow, trial-and-error sensory testing.

The technical core of the new approach lies in the combination of two complementary machine learning components. The first is a protein language model, a type of neural network trained to capture the statistical patterns of amino acid sequences in much the same way that large language models learn the structure of human text. The team trained this model using approximately 500 known bitter-tasting peptides, allowing it to internalize the sequence features associated with bitterness. The second component is BitterPep-GCN, a prediction model the group had developed in earlier work and described in 2024 in the Journal of Cheminformatics. BitterPep-GCN is a Graph Convolutional Network, a specialized form of artificial neural network designed to analyze structured data. In this case, the structured data are peptide molecules themselves, represented as graphs in which amino acid residues function as nodes and the chemical relationships between them as edges.

Graph-based representations give the model an advantage that purely sequence-based methods can lack. Because the network processes the peptide as a structured chemical object rather than a simple string of characters, it can learn how the arrangement, identity, and interactions of residues influence binding to bitter taste receptors. By fusing the knowledge encoded in the protein language model with the structural sensitivity of the graph convolutional network, the researchers created a system capable of both classifying existing peptides and generating novel candidate sequences. This dual capability, known in the field as de novo design, is what distinguishes the new method from earlier bitterness predictors that could only score peptides already in hand.

Putting the system to the test, the researchers first used it to generate 161 new peptide sequences that had never been experimentally characterized. The pipeline then filtered this set, identifying the candidates that, according to the model predictions, were highly likely to taste either strongly bitter or clearly non-bitter. Selecting the most promising of these designed molecules, the team had them chemically synthesized and submitted to a trained sensory panel for evaluation. Human tasters, rather than receptor assays alone, provided the ground truth, which is a demanding standard for any computational model of flavor.

The results were striking. Of the 31 designed peptides ultimately tasted, the trained panel confirmed the AI predictions in 25 cases, correctly classifying them as bitter or non-bitter. In the course of the experiments, the researchers also identified numerous previously unknown bitter-tasting and non-bitter-tasting peptides, expanding the experimental dataset available to the field. For a property as subtle and receptor-specific as bitterness, a prediction accuracy of roughly 80 percent in a blind de novo design setting represents a substantial step forward, and it demonstrates that generative models can produce chemically meaningful candidates rather than merely ranking known compounds.

Alexandra Steuer, first author of the study and a doctoral student in Di Pizio’s group, emphasizes what the results mean for the discipline. The findings show, she notes, that not only can the bitterness of peptides be predicted, but that the new AI-based method can also be used to specifically design new bitter-tasting peptides. Di Pizio adds that this brings researchers significantly closer to the goal of proactively controlling taste characteristics, rather than reacting to off-flavors after they appear in a finished product. The distinction between reactive quality control and proactive molecular design captures the practical promise of the approach.

Di Pizio also stresses that the research is ready to be implemented in application frameworks. In the long term, the new findings could help to specifically control the formation of bitter-tasting peptides during food production, a capability she describes as particularly relevant for plant-based, protein-rich foods, whose acceptance often suffers because of undesirable flavor notes. If manufacturers can predict, early in product development, which peptides will emerge from a given protein source and processing regime, they could adjust fermentation starters, enzyme choices, or formulation strategies to steer the flavor outcome. Conversely, the ability to design bitter peptides on demand could support research into appetite regulation and satiety, where bitterness may play a functional physiological role.

The study, titled De novo design and experimental characterization of bitter peptides, was published in npj Science of Food on June 25, 2026, with a author team including Steuer, Ferri, Eckrich, Heidenkampf, Mittermeier-Kleßinger, Schaefer, Behrens, Ferruz, Dawid, and Di Pizio. Training data for the language model came from the Bitter Peptide Space (BPS)-1000 database maintained by the Leibniz Institute, a curated resource that underpins much of the group’s computational work. The research was performed computationally at its core, with experimental validation through synthesis and human sensory testing, and the authors declare no competing interests. As machine learning continues to move from analyzing existing molecules to creating new ones, the Munich-led study offers a concrete demonstration that the taste of tomorrow’s foods, down to the individual peptide, can increasingly be designed rather than discovered.

Subject of Research: AI as a tool in flavor research

Article Title: AI as a tool in flavor research

Article References: AI as a tool in flavor research. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: tool, flavor, research, scientific research, peer-reviewed research, research findings

The appeal of AI in this setting rests on a simple asymmetry. Clinical trials generate enormous volumes of structured and unstructured data—imaging, laboratory values, genomic profiles, clinician notes, wearable-device streams and patient-reported outcomes—yet historically only a small fraction of that information has been used systematically in trial design and conduct. Machine-learning models, particularly modern deep-learning and foundation-model architectures, excel at extracting signal from exactly these high-dimensional, heterogeneous datasets. When trained on curated clinical data, they can identify which patient subpopulations are most likely to respond to a therapeutic candidate, predict which sites will struggle with recruitment, flag data inconsistencies long before database lock, and estimate the probability that an ongoing trial will meet its primary endpoint. The result is a shift from retrospective, intuition-driven trial management toward a prospectively optimised, continuously adaptive process.

Patient recruitment illustrates the potential most vividly. Failed recruitment and poor participant retention are among the leading reasons that trials miss their timelines or terminate early. Traditional approaches rely on broad eligibility criteria, manual chart review and outreach from a limited number of academic centres, which systematically under-enrols patients from rural areas, lower-income communities and historically marginalised groups. AI-driven approaches invert this model. Natural-language processing systems can scan millions of de-identified electronic health records to locate every patient matching a complex eligibility profile, including patients whose relevant conditions are described only in free-text notes rather than coded diagnoses. Matched patients can then be connected with trial sites through their treating physicians or through decentralised trial infrastructure that brings the study to the participant rather than the reverse. When designed carefully, such tools can also be tuned to improve diversity in enrolment, addressing a long-standing scientific weakness as well as an ethical one: a trial population that does not resemble the eventual treated population limits the generalisability of the results.

Trial design itself is being recomputed. Classical phase II and phase III trials typically use fixed designs conceived months before enrolment begins, and any mid-course revision requires protocol amendments that can delay readouts by many months. AI-assisted design tools draw on large repositories of historical trial outcomes, natural-history data and simulation frameworks to model, before a single patient is enrolled, how a trial will behave under different assumptions about effect size, dropout, endpoint variability and site performance. Bayesian adaptive designs, which allow the randomisation ratio and dose allocation to shift as interim data accumulate, become far more practical when machine-learning models supply reliable predictive components. Digital twins—computational replicas of individual patients built from rich baseline data—offer an emerging complement, allowing a portion of the control information in a trial to be estimated rather than observed, and thereby reducing the number of real participants who must be randomised to achieve a given statistical power. Regulators have begun engaging with these methods, and several external control arms constructed from historical or registry data have already supported regulatory submissions in rare-disease oncology.

Once a trial is running, AI changes the economics of monitoring. Risk-based monitoring, in which scrutiny is concentrated on the sites and data elements most likely to harbour errors, depends on recognising patterns across thousands of simultaneous data streams—precisely the task that anomaly-detection algorithms perform well. Centralised statistical monitoring can flag sites whose data distributions deviate from the norm, whether because of fraud, systematic measurement error or simple process breakdown, without the cost of sending monitors to every site on a fixed schedule. Automated adverse-event signal detection can surface safety trends earlier in the data stream, shortening the interval between an emerging risk and a protocol response. Speech-recognition and summarisation models are increasingly used to draft clinical notes, assist with adjudication of endpoints that require expert review, and reduce the administrative burden that currently consumes a large share of investigator time. Because the marginal cost of applying a trained model to new data is close to zero, these efficiencies can scale across an entire portfolio rather than applying to a single study.

The pharmaceutical industry’s interest follows directly from the arithmetic. Industry analyses have repeatedly estimated that bringing a new drug to market costs on the order of one to two billion dollars, with clinical development accounting for the majority of that expenditure and with most of the cost attributable to failed trials. Even a modest improvement in the probability of technical success, achieved through better target-patient matching or earlier detection of futility, translates into hundreds of millions of dollars in expected savings per programme and, more importantly, into faster access for patients to therapies that work. Every month shaved from a development timeline extends the effective patent-protected market life of a medicine, which strengthens the commercial case, but the public-health case is at least as strong: pipelines that iterate faster can respond more quickly to emerging pathogens, to rare diseases that currently have no treatment at all, and to the individualisation of therapy in fields such as oncology where one-size-fits-all efficacy is the exception rather than the rule.

None of this means the transformation is automatic. Machine-learning models inherit the biases of the data on which they are trained, and clinical datasets under-represent precisely the populations in whom trial evidence is weakest. An algorithm trained mostly on data from large academic hospitals may perform poorly for patients managed in community settings, and a recruitment tool optimised narrowly for speed could worsen enrolment diversity if fairness constraints are not built in explicitly. The opacity of complex models also sits awkwardly with regulatory expectations: a sponsor that uses an AI-derived covariate in a statistical analysis plan, or an AI-curated external control arm, must be able to explain to reviewers how the model was built, validated and monitored. Regulators including the United States Food and Drug Administration and the European Medicines Agency have signalled that they will evaluate such tools under existing frameworks for risk-based software validation, but the guidance landscape is still maturing, and sponsors who treat AI components as unexamined black boxes do so at their own regulatory peril.

Data governance presents a second structural challenge. The models that promise the greatest gains in recruitment and design are those trained on the largest and most diverse clinical datasets, yet health data are fragmented across institutions, jurisdictions and incompatible record systems, and privacy law constrains how they can be pooled. Federated learning, in which models are trained across multiple sites without moving the underlying patient data, offers a technically elegant partial solution, but it introduces its own questions about model ownership, auditability and the equitable distribution of the value created. Standard-setting efforts, from common data models to documented provenance for training sets, will determine whether AI-enabled trials become a broadly shared capability or a competitive advantage concentrated in a handful of organisations with the largest proprietary data estates.

The most realistic near-term picture is therefore not one of autonomous AI running trials, but of a human-machine division of labour in which algorithms perform the enumeration, matching, monitoring and simulation that humans cannot do at scale, while clinicians, statisticians and regulators retain judgment over what counts as evidence. Under that division of labour, the measurable signs of change are already visible: screening times measured in days rather than months at leading sponsors, growing numbers of adaptive and model-informed designs entering regulatory review, and decentralised, data-rich trial formats that were logistically implausible a decade ago. If those trends continue, the defining feature of the next generation of clinical trials will not be any single algorithm but the integration of computation into every stage of the evidentiary pipeline—from the first patient matching query to the final submission—producing trials that are faster, smaller where possible, larger where necessary, and ultimately more representative of the patients the resulting medicines are meant to serve.

Subject of Research: AI-enabled clinical trials

Article Title: AI-enabled clinical trials

Article References: Raynaud, M., Trayanova, N., Mannon, R. B., André, F., Doraiswamy, P. M., & Loupy, A. (2026). AI-enabled clinical trials. Nature Reviews Bioengineering. https://doi.org/10.1038/s44222-026-00487-7

Image Credits: AI Generated

DOI: 10.1038/s44222-026-00487-7

Keywords: AI-enabled, clinical, trials, scientific research, peer-reviewed research, research findings

To understand why surface wrinkles matter so much, it helps to revisit the basic physics of multi-junction solar cells. In a single-junction device, photons above the absorber’s bandgap deposit their excess energy as heat, while photons below the gap pass through unabsorbed — a fundamental trade-off that caps efficiency around the Shockley–Queisser limit. Stacking multiple absorbers with different bandgaps circumvents this constraint: a wide-bandgap top cell harvests high-energy blue photons, while narrower-gap layers beneath capture the red and infrared remainder. All-perovskite triple-junction designs are particularly attractive because perovskites offer tunable bandgaps, solution processability and strong defect tolerance. But the topmost layer in these stacks must be an ultrawide-bandgap perovskite — approaching roughly 2 electron volts — and that is precisely where the chemistry turns hostile.

The hostility stems from bromide. To push a perovskite’s bandgap into the ultrawide regime, formulators must replace much of the iodide in the crystal lattice with the smaller, more electronegative bromide anion. High bromide content, however, destabilizes the film during deposition and annealing. As the solvent evaporates and the crystallites grow, differential stresses and uneven halide distribution produce a surface texture the researchers describe as wrinkling — nanoscale and microscale corrugations that ripple across the top interface of the film. These wrinkles are not merely cosmetic. They increase surface area where recombination can occur, disrupt the uniformity of subsequent charge-transport layers, and scatter light in uncontrolled ways. Worse still, the same high-bromide chemistry promotes halide disorder: iodide and bromide ions segregate into bromide-rich and iodide-rich domains, creating local bandgap fluctuations that act as recombination hotspots and undermine the very wide bandgap the design depends on.

The new work, discussed in the News & Views analysis, tackles both problems simultaneously through a combined surface reconstruction and halide homogenization strategy. Surface reconstruction, in essence, reprocesses the topmost region of the perovskite film after it forms, relaxing the strained, defect-rich outer layer and allowing it to recrystallize into a flatter, more orderly termination. Halide homogenization addresses the compositional side of the problem, encouraging iodide and bromide to distribute evenly through the lattice rather than pooling into segregated phases. Together, the two interventions produce an ultrawide-bandgap perovskite surface that is morphologically smooth, compositionally uniform and electronically cleaner — three qualities that wide-bandgap perovskites have historically failed to deliver at once.

The consequences for device performance are dramatic. In the reported all-perovskite triple-junction devices, the smoothed 2-electron-volt top absorber transmits its full complement of sub-bandgap photons to the middle and bottom cells while extracting high-energy charge carriers with far fewer losses. The certified efficiency of 29.3 percent represents a benchmark for all-perovskite triple-junction technology, a figure that places these fully thin-film stacks in the same conversation as the best silicon-based tandems while retaining the advantages of low-temperature, scalable fabrication. Certification matters here: independent verification distinguishes genuine reproducible performance from laboratory optimism, and a certified number signals to the field that the surface-engineering approach survives the scrutiny of recognized testing protocols.

The path to this point has been built on a sequence of steady advances that the Nature Energy analysis situates in context. Earlier work on wide-bandgap perovskites documented the efficiency ceilings imposed by halide segregation and interfacial recombination, establishing the diagnosis that the new treatment directly addresses. Subsequent studies demonstrated high-performance perovskite-based multi-junction architectures and clarified how the constituent subcells could be optically and electrically matched. The underlying research article by Zhang and colleagues, which the analysis accompanies, now closes the loop by showing that the materials problem — the wrinkled, halide-disordered surface of ultrawide-bandgap absorbers — could be engineered away rather than merely tolerated. In the ecosystem of photovoltaics research, this is how fields advance: not by a single breakthrough, but by each study removing the specific obstacle the previous one identified.

From a manufacturing standpoint, the significance of the result extends beyond the headline efficiency. Solution-processed perovskites can in principle be deposited by printing, blade coating or roll-to-roll methods at temperatures far below those required for high-quality silicon, promising low embodied energy and low cost. But every additional processing step must be compatible with the layers beneath it, and triple-junction stacks compound this challenge: the top cell’s surface must accept a charge-transport layer, then a recombination layer, then the next absorber, each without degrading what came before. A wrinkled, compositionally inhomogeneous surface jeopardizes every one of those interfaces. By flattening and homogenizing the top absorber before the stack is completed, the reconstruction strategy improves not just a single interface but the integrity of the entire device architecture — a systems-level benefit that efficiency figures alone understate.

There are also deeper scientific questions that the work reframes. Halide segregation in mixed-halide perovskites has been studied extensively under illumination, where it manifests as photoinduced phase separation; the wrinkling phenomenon tied to high bromide content during film formation is a related but distinct failure mode rooted in crystallization mechanics. Understanding how strain, surface energy and halide mobility interact during deposition — and how a post-deposition reconstruction can erase the damage — offers a template that may generalize beyond ultrawide-bandgap compositions. If the same principles can be applied to tune the surfaces of other wide-bandgap perovskites, or to heal defects in the buried interfaces of the lower subcells, the entire tandem design space opens up. The Kyoto authors, who have long studied the chemistry of perovskite film formation from both academic and industrial perspectives — Wakamiya is co-founder and chief science and technology advisor of the perovskite startup EneCoat Technologies — are well placed to translate that understanding into manufacturable processes.

Challenges remain on the road to commercialization, and the analysis is careful to keep them in view. Triple-junction perovskite devices must prove their operational stability over thousands of hours of illumination, thermal cycling and moisture exposure — tests that mixed-halide compositions, with their susceptibility to ion migration, have historically found difficult. Scaling the surface reconstruction from small-area champion cells to large-area modules without sacrificing uniformity is another hurdle, as is demonstrating that the certified 29.3 percent can be reproduced with the reproducibility that industrial partners require. And the perovskite field as a whole still faces questions of lead content and end-of-life recycling that no efficiency record can resolve. Yet the trajectory is unambiguous: a technology that could not break 20 percent a few years ago is now approaching the practical ceiling of single-junction silicon, with theoretical headroom that silicon simply does not have.

For now, the image that captures the advance is the one its title invokes: wrinkles, ironed out. Nanoscale corrugations and halide disorder that once seemed like unavoidable consequences of wide-bandgap chemistry have been shown to be engineering problems with engineering solutions. With a certified 29.3 percent efficiency in an all-perovskite triple-junction cell, the field has demonstrated that the highest-performance perovskite architectures need not be compromised by their most challenging component. As the analysis by Nakamura, Truong and Wakamiya makes clear, the recipe for next-generation photovoltaics may hinge not on exotic new materials, but on learning to make the surfaces of the ones we already have perfectly flat, perfectly mixed and perfectly quiet — at least at the level of the electron.

Subject of Research: Ironing out the wrinkles

Article Title: Ironing out the wrinkles

Article References: Nakamura, T., Truong, M. A., & Wakamiya, A. (2026). Ironing out the wrinkles. Nature Energy. https://doi.org/10.1038/s41560-026-02127-1

Image Credits: AI Generated

DOI: 10.1038/s41560-026-02127-1

Keywords: Ironing, wrinkles, scientific research, peer-reviewed research, research findings, science news