BioByte 115: plant-permeable T6H increases wheat yields, AI in cognitive modeling, steps toward a universal snake antivenom, and advancements in conversational diagnostic AI
Welcome to Decoding Bio’s BioByte: each week our writing collective highlight notable news—from the latest scientific papers to the latest funding rounds—and everything in between. All in one place.
What we read
Membrane-permeable trehalose 6-phosphate precursor spray increases wheat yields in field trials [Griffiths et al., Nature Biotechnology, April 2025]
With the human population estimated to increase to 9 billion by 2050 it is essential to improve the yield of crops. Wheat is a major staple food source providing around 20% of calories and protein in a daily human diet. In this paper, the authors use a trehalose-6-phosphate (T6P) signalling precursor (DMNB-T6P) to improve the yields of wheat plants under a variety of different conditions. T6P is a plant sugar signal that promotes anabolic pathways, including starch biosynthesis (the world’s most important food carbohydrate staple). Genetic modification to increase T6P synthesis has proven difficult in wheat. For this reason, the authors use a plant-permeable DMNB-T6P analog.
The authors tested the effect of DMNB-T6P timed microdoses on wheat yield over four years across a variety of agricultural conditions including under well-watered and water-stressed conditions. These experiments showed on average an increase in wheat yield ranging from 5-17% compared to controls. Interestingly, these yield increases were observed under both good and sup-optimal conditions although to a lower extent under sup-optimal conditions. Importantly, the authors also established that DMNB-T6P could be synthesised in sufficiently large batches by contract research organisations.
To understand why DMNB-T6P increased yield in wheat the authors analysed gene expression in wheat grains. As expected, this showed an upregulation of genes associated with key, flux-associated steps in starch biosynthesis. Analysis of carbon dioxide uptake and linear electron flow in leaves revealed an increase in carbon dioxide fixation and an increase in linear electron flow. Linear electron flow generates the required energy and reducing power for the Calvin cycle to fix carbon dioxide. Together these finds show multiple pathways contributing to an increase in starch biosynthesis.
Overall, this study reveals that DMNB-T6P enables wheat yield increases above existing selective breeding strategies. This technology may provide a step change in the development of new technology to enhance wheat yields.
Cognitive modeling using artificial intelligence [Goodman & Frank, PsyArXiv Preprints, April 2025]
Frank and Goodman of Stanford have written a review on using model AI models to study human cognition. Why bring AI into cognitive science? Traditional cognitive models often rely on simplified inputs and handcrafted rules which limits how accurately they reflect real human experience. Current AI systems such as LLMs, are trained on vast naturalistic data (text, images, video) just as humans are, making them truly “stimulus computable”. This means researchers can feed the same inputs to both people and machines and compare the outputs directly. This begs the question as to how far could such models go as models of the human mind?
The authors argue that to treat an AI system as a model of the mind, three alignments are essential. First, there must be clear mapping between network components and psychological constructs such as beliefs and goals. Second, the model’s training regimen must mirror human learning in modality, order and scale. Third, experiments must go beyond overall benchmarks and examine how small changes in inputs affect both model activations and human responses under matched conditions.
Finally, they spotlight three urgent challenges. Reproducibility suffers when top models remain closed-source black boxes, so science needs open weights, code, and data. Most comparisons use narrow convenience samples, so bridging the gap between model universes and diverse human populations is critical. And although interpretability tools have advanced, networks with billions of parameters still defy simple explanation, making causal intervention techniques and alignment with neural data vital for truly understanding these machines of mind.
Snake venom protection by a cocktail of varespladib and broadly neutralizing human antibodies [Glanville et al., Cell, April 2025]
A new discovery of a three-component antivenom cocktail has the potential to revolutionize the way snake bites are treated. Each year, incidents with snakes result in around 81,000 to 138,000 deaths, with an additional 300,000 to 400,000 permanent disabilities. Much of the difficulty in treating these bites arises from the remarkable heterogeneity in venom composition both within a population and across species, meaning that most antivenoms will only work in a small percentage of cases. Difficulties identifying the offending snake further compound this problem, and since antivenoms are largely derived from large mammals, they often cause adverse immune responses upon administration. In seeking to address these limitations, Glanville et al. use the antibodies in the blood of Tim Friede—a hyperimmune individual who has been dosing himself with snake venom for 18 years—and an “iterative addition” approach to find a three-component antivenom that can treat bites from at least 19 snakes in the Elapidae family.
Individual snake venoms can comprise 7-50 unique protein toxins, with the vast majority of them in Elapidae being three-finger neurotoxins (3FTXs) and phospholipase A2s (PLA2s). Since small molecule inhibitors are already for PLA2, the focus of this study was on 3FTXs, specifically long-chain neurotoxins (LNX) and short-chain neurotoxins (SNX). After selecting a panel of 19 medically-relevant snakes with diverse genomes and geographical origins, the authors used phage display to isolate those antibodies which were broadly neutralizing when introduced to recombinant LNX from four of these snakes. Finding that LNX-D09 was in particularly high abundance, ELISA and SPR experiments were run to validate reactivity and elucidate kinetics respectively. Crystal structures were collected to elucidate the atomic underpinning for the inhibition, revealing that LNX-D09 bids in across an interface that many LNXs use.
When screened in mice against the selected panel of Elapidae, LNX-D09 was found to protect against venoms from six species. Adding a known PLA2 inhibitor—varespladib—increased this number by three. Seeking to extend the applicability of the cocktail, the authors underwent the same screening process used for the LNXs with SNXs, finding that SNX-B03 also exhibited broad inhibition. When SNX-B03 was added to the cocktail, protection increased to a total of all nineteen snake venoms in the panel.
Although impressive, this study is not without its limitations. The in vivo experiments were largely carried out in mice, and the administered venom was kept constant, a luxury not granted in nature. It was also limited to only Elapidae, which constitute a portion of venomous snakes. Authors highlight that studying other genera such as Viperidae will be necessary to elucidate similar generalizable antibodies.
The advances put forth in this study bring a universal antivenom one step closer to reality. The three-component cocktail the authors discovered works to nullify the need for species identification prior to receiving antivenom and address for intra- and inter- species genetic variation. Being human-derived, it also simultaneously reduces the immune-related side effects of non-human antivenoms and increases the half-life, enabling facile stockpiling of a single antivenom. With subsequent advances, it is possible that snake bites will readily become a problem of the past.
Advancing Conversational Diagnostic AI with Multimodal Reasoning [Saab et al., Google Research, May 2025]
LLMs are quickly showing incredible utility in clinical contexts. Following their previous work on using LLMs to improve differential diagnoses (McDuff et al., Nature, April 2025), researchers from Google extend the diagnostic capabilities of their model AMIE from just text-based information to utilizing multi-modal information.
Saab et al. adapt on their previous AMIE model through the integration of multimodal medical data, such as images and clinical documents, into diagnostic conversations. The system employs a novel state-aware diagnostic framework that adaptively updates the differential diagnosis, strategically asks for more information (including requesting and interpreting images), delivers diagnoses, and continuously tracks and updates its understanding of the patient's state like a real clinician would. In ablation studies they find that the state tracking and dialogue modules both significantly improve the top-1 differential diagnosis rate (rate that AMIE’s first diagnosis is the correct one). In randomized, double-blind tests adapted from objective structured clinical examinations (OSCEs), evaluated by an automated rater and 18 clinician specialists, AMIE demonstrated superior performance compared to real primary care physicians on 7 out of 9 axes related to multimodal handling and 29 out of 32 non-multimodal axes, including diagnostic accuracy.
Scaling healthcare delivery is inherently hard and increasingly expensive, due to factors such as the demanding nature of medical training, the need for specialization, and the challenges of centralizing care. Demonstrating the potential to outperform PCPs in certain diagnostic tasks, the development of multimodal AMIE represents a significant milestone in adapting AI systems for effective healthcare delivery.
Structuring Scientific Innovation: A Framework for Modeling and Discovering Impactful Knowledge Combinations [Chen et al., arXiv, April 2025]
Chen et al. propose a structured view of scientific discovery that emphasizes the role of method combinations in shaping disruptive insights, as the most influential scientific discoveries stem from the combination of traditional ideas from prior work. LLMs have been applied to generate research ideas and synthesize existing knowledge. However, these methods are 1) unable to systematically identify and integrate knowledge that leads to precise matching of research problems and methods, 2) handicapped due to hallucinations, and 3) due to the absence of objective metrics are unable to assess the transformative impact of new proposed discoveries.
In order to take the use of LLMs further in scientific discovery, beyond idea generation, the authors introduce the Disruptive Index (DI) to quantify whether a scientific discovery drives a paradigm shift. In order to generate potentially disruptive problem-method combinations, the team built a platform that uses a framework to identify and integrate problem-method combinations, evaluates using the DI and validates the platform effectiveness by identifying high-disruptiveness method combinations across scientific domains. The approach achieves a higher hit rate of high disruptiveness method combinations than the best LLM by 9%.
Notable deals
FutureHouse launched the FutureHouse Platform, a sleek all in one web and API interface that brings together its suite of AI driven scientific agents to streamline everyday research workflows. By combining four specialized agentic tools: Crow for rapid Q&A literature searches, Falcon for deep review and synthesis, Owl for gap finding (“Has anyone done X?”) and Phoenix for chemistry planning, scientists can accelerate literature reviews, data mining and basic bioinformatic analyses. Try it free here.
Stately Bio, founded by Frank Li, has closed a $12 million seed round led by AIX Ventures to advance its non-destructive cell‐profiling platform. Unlike conventional assays that compromise cell health, Stately’s technology tracks identity, behavior and maturation in living cells without killing or significantly perturbing them. The new capital will fuel the development of next-generation stem cell therapies for regenerative medicine.
NewLimit raises $130 million Series B led by Kleiner Perkins. The biotech has a bold mission: to extend human healthspan using epigenetic reprogramming. Using AI-driven discovery of combinatorial TFs and RNA delivery, the company restores youthful gene expression and function in key cell types, such as hepatocytes, T cells and endothelial cells to tackle age-related diseases. With proprietary human-cell screening platforms, humanized animal models and clinically proven LNP-RNA delivery, NewLimit is advancing differentiated programs in metabolism, immunology and vascular health toward first-in-class reprogramming therapies.
Deerfield raises a $600M third fund to invest further in therapeutics, medical devices and healthcare services. This fund also pursues the aim of investing in emerging tech, including machine learning and artificial intelligence as they seek to advance healthcare and the life sciences. Deerfield boasts an impressive in-house ecosystem and history supporting the healthcare industry since 2005, collaborating with 29 leading research institutions and multiple non-profits—a major component of Fund III’s focus as well.
What we liked on socials channels
Field Trip
Did we miss anything? Would you like to contribute to Decoding Bio by writing a guest post? Drop us a note here or chat with us on Twitter: @decodingbio.