The researchers describe a modular Safety-by-Design framework built around two complementary analytical branches. The first branch is dedicated to candidate prioritization: given a large space of possible RNA-lipid nanoparticle designs, it identifies which formulations best satisfy competing objectives of efficacy and safety. The second branch addresses a subtler question, namely whether the biological picture produced by the framework remains consistent with what is known about human immune organization. By separating these two tasks, the authors argue, the framework avoids the common pitfall of blending mechanistic interpretation, formulation ranking, and biological validation into a single opaque pipeline.

At the heart of the prioritization branch sits a deliberately constructed synthetic design space of 500 RNA-LNP formulations. Each formulation is evaluated across a heterogeneous cohort of virtual patients, computational stand-ins whose simulated immune responses capture variability among real people. From these simulations the framework derives five synthetic endpoints related to efficacy and safety, which are then condensed into a composite Safety-by-Design score. The score is not a single magic number imposed from above; rather, it is a transparent aggregation that can be interrogated, reweighted, and stress-tested, which is precisely what the authors do in a series of sensitivity analyses.

To make the exploration of 100,000 possible patient-formulation pairs tractable, the team trained a Random Forest surrogate model to reproduce the composite score. The model performed well within the synthetic design space: in formulation-grouped five-fold cross-validation, it achieved a pooled out-of-fold coefficient of determination of 0.887, with a root mean square error of 0.055 and a mean absolute error of 0.044. Grouping the cross-validation by formulation is an important methodological choice, because it tests whether the model generalizes to formulations it has not seen, rather than merely memorizing patient-level variation within a familiar formulation.

With the surrogate in place, the researchers turned to Pareto-based multi-objective analysis, a technique borrowed from optimization theory in which a candidate is considered non-dominated if no alternative is better on every objective simultaneously. The primary analysis identified 21 such non-dominated formulations, representing the efficient frontier of the efficacy-safety trade-off. Among these, one candidate, designated LNP_0462, ranked first globally. Crucially, it also ranked first when the analysis was repeated separately within strata of low, intermediate, and high inflammatory risk, suggesting that its leading position is not an artifact of averaging across patient subgroups with very different immune profiles.

Robustness was examined from several angles. The authors tested 3,876 alternative weight combinations for the Safety-by-Design score and found that the ranking was generally stable across them, indicating that the conclusions do not hinge on one particular way of trading efficacy against safety. They also performed a complete balanced analysis of all 100,000 patient-formulation pairs, rather than relying on sampled subsets, and the resulting ranking closely matched the primary analysis, with a Spearman correlation of 0.998 and an overlap of 18 of the top 20 candidates. For a framework intended to guide experimental prioritization, this kind of internal consistency matters as much as raw predictive accuracy.

The second branch of the framework tackles biological plausibility in a way that is deliberately kept apart from the ranking machinery. The authors drew on trajectories generated by the Universal Immune System Simulator, a mechanistic model published in previous work, and condensed descriptors from those trajectories into five Mechanistically Informed Immune Signatures, or MIIS. These signatures were used to constrain synthetic multi-omics profiles, which were then compared at the level of immune programs with an independent human single-cell RNA-sequencing dataset, GSE171964. Importantly, the MIIS variables and the synthetic multi-omics profiles were not used as predictors, endpoint inputs, or Pareto objectives, so the consistency assessment cannot inflate the ranking itself.

The comparison was quantified with a metric the authors call BPCI, computed across seven post-vaccination days. The mean BPCI was 0.760, and it exceeded an exact shuffled-label null with a p-value of 0.025. The authors are careful, however, not to oversell this result. Day-specific evidence was heterogeneous, and donor-level uncertainty was substantial, so the finding supports the general idea that the framework’s synthetic outputs organize immune programs in a manner consistent with real human data, rather than validating any individual formulation’s clinical behavior. The MIIS branch, as the authors put it, provides an assessment of immune-program organization, not a validation of specific candidates.

This restraint is characteristic of the paper’s overall posture. The work is explicitly labeled a proof of concept, and the authors state that experimental testing and evaluation in additional human datasets are required before any conclusions can be drawn about biological or clinical performance. The study involved no new human participants or animals; it relied on computational simulations, synthetic data, and secondary analysis of a publicly available, de-identified transcriptomic dataset, so no additional ethics approval was required. The research received no specific grant funding, and the authors declare no competing interests. The article was received on 28 July 2026, accepted on 27 August 2026, and published open access on 7 September 2026.

Even as a proof of concept, the framework points toward a shift in how RNA therapeutics might be developed. Instead of treating formulation design, safety assessment, and biological plausibility as separate silos handled by different teams at different stages, the Catania group’s approach integrates them into one transparent, modular pipeline in which every assumption can be varied and every ranking can be stress-tested. The surrogate model makes exhaustive exploration affordable, the Pareto analysis makes trade-offs explicit rather than hidden in a weighted average, and the MIIS branch provides an independent sanity check against real human immunology. If subsequent experimental work confirms that computationally prioritized candidates such as LNP_0462 translate into safe and effective formulations, frameworks of this kind could become a standard first filter in the design of RNA-LNP therapeutics, narrowing the laboratory search space and, ultimately, shortening the path from molecular design to clinical candidate.

Subject of Research: A computational Safety-by-Design framework for prioritizing RNA-lipid nanoparticle formulations and assessing biological consistency

Article Title: A modular Safety-by-Design framework for RNA-LNP prioritization and mechanistically informed biological consistency assessment

Article References: Di Salvatore, V., Russo, G., & Pappalardo, F. (2026). A modular Safety-by-Design framework for RNA-LNP prioritization and mechanistically informed biological consistency assessment. BMC Bioinformatics. https://doi.org/10.1186/s12859-026-06630-w

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06630-w

Keywords: RNA therapeutics, lipid nanoparticles, Safety-by-Design, virtual patients, Random Forest, Pareto optimization, mechanistic modeling, systems immunology, synthetic multi-omics, single-cell RNA sequencing, explainable artificial intelligence, drug safety

The research, led by Vinícius Vicente Soares and colleagues at the Federal University of Health Sciences of Porto Alegre in Brazil, describes the technical development and a pilot expert-based content validation of an interactive voice prototype built on GPT-4o, OpenAI’s multimodal model. Rather than testing educational outcomes in a cohort of students—a step the authors explicitly defer to future work—the study documents how such a system is actually engineered, and what a seasoned psychiatrist found when she sat down to interview its virtual patients. The result is both a proof of concept and an unusually candid inventory of the technology’s shortcomings.

The team built four psychiatric personas covering conditions chosen for their epidemiological weight and their semiological richness: major depressive disorder, bipolar disorder in a manic episode, schizophrenia, and attention-deficit/hyperactivity disorder. Together these span the core domains a trainee must probe—mood, thought, and attention—and demand very different interviewing styles, from drawing out a psychomotorically slowed, withdrawn depressed patient to keeping pace with the flight of ideas of mania. Initial profiles were drafted by the researchers from DSM-5-TR and ICD-11 criteria and established psychopathology literature, with the GPT-o3 model used strictly as a writing assistant to polish cohesion, a deliberate design choice meant to keep the clinical content under human control.

Each persona then went through a structured human-in-the-loop refinement process. A board-certified psychiatrist with more than fifteen years of experience in medical education—the study’s subject matter expert—conducted test interviews with each virtual patient, and her qualitative feedback drove successive revisions of the system prompts. Three refinement cycles per persona were completed, adjusting symptomatological nuances such as increasing response latency for the depressed patient and toning down caricature-like behaviors, until the expert judged each profile clinically plausible for a training environment.

The prompt engineering, described in unusual technical detail, is arguably the study’s most useful contribution. Each persona was operationalized through a master instruction with standardized components: a contextualization block defining the patient’s identity and framing the encounter as a first psychiatric assessment; a core identity section containing a biographical summary, central symptomatology with concrete verbal and behavioral examples, the predominant affective state that directly informs the text-to-speech prosody, and the patient’s level of insight into their condition; and dynamic interaction guidelines specifying vocal style, speech rhythm, and reactivity—so that the virtual patient might open up to an empathic interviewer but grow defensive when delusional beliefs were directly challenged.

Equally important were the constraints. So-called negative prompts were written to suppress known failure modes of large language models: hallucination of facts, the reflex to be overly helpful, drift into generic AI-assistant behavior, and the temptation to use technical psychiatric jargon—which would let a trainee off the hook of translating a patient’s own words into semiological terms. The prompt also embedded a formative feedback mechanism: on the command “END,” the model generates a structured critique of the interviewer’s performance, identifying the semiological phenomena displayed by the persona, evaluating the interviewer’s approach, and suggesting improvements, such as exploring suicide risk more thoroughly.

Evaluation followed a heuristic evaluation protocol adapted from usability engineering. For each persona, the expert received only a brief case summary—a fictitious name, age, and chief complaint—without access to the underlying prompt, then conducted a free, unstructured thirty-minute interview that was fully recorded. She rated each simulation on Likert scales covering clinical fidelity, persona consistency, affective expression, interaction quality, and pedagogical value, and her open-ended comments were subjected to thematic content analysis. The authors acknowledge a key limitation here: the evaluator was involved in the iterative development of the personas, so the assessment was not fully independent, and the study rests on a single expert rather than a panel.

The quantitative picture was strikingly consistent. The pedagogical value dimension received the maximum score of five across all four simulations, suggesting the expert saw genuine educational potential regardless of clinical imperfections. The ADHD persona earned the best overall rating at 4.4, while the schizophrenia persona scored lowest at 3.8. Fidelity-related dimensions, particularly interaction with the interviewer, lagged behind, and the item assessing whether the simulated patient avoided stereotyping received the minimum score for three of the four personas—a red flag that shaped much of the qualitative analysis.

Three themes dominated the expert’s critique. First, clinical stereotyping: the virtual patients tended to deliver textbook presentations, announcing their symptoms with a clarity real patients rarely muster. Of the schizophrenia persona, the evaluator observed that although the patient described typical symptoms in a detailed and coherent manner, she was very stereotyped, since patients with schizophrenia generally do not reveal their symptoms so readily. Second, vocal naturalness fell short: the manic patient talked a great deal but lacked the characteristic pressure of speech of true mania, stopping in ways the evaluator found unnatural, and the artificial prosody undermined immersion. Third, platform barriers intruded: content moderation on the commercial ChatGPT platform censored discussion of hypersexuality in the mania persona—a core symptom—producing an artificially aseptic scenario, and the model could not reproduce regional linguistic traits, a limitation scored at the minimum for every persona.

The feedback mechanism, by contrast, performed well, generating structured and pertinent analyses that correctly identified phenomena such as flight of ideas and expansive mood. The authors’ conclusion is measured: voice-based LLM simulations are technically feasible and pedagogically promising, but current limits in multimodal realism, the risk of teaching students to recognize caricatures rather than the heterogeneous face of mental illness, and the opacity of these models mean the tool is best suited for supervised formative practice, not high-stakes assessment. As a proof of concept, the study maps the terrain ahead—better prompt engineering, dedicated platforms, multi-center trials with real trainees, and rigorous faculty oversight—before AI patients can take a lasting seat in the psychiatry classroom.

Subject of Research: Development and expert-based content validation of a GPT-4o voice prototype for simulating psychiatric patient interviews in medical education

Article Title: LLM-Based Psychiatric Interview Simulation: Technical Development and Pilot Expert-Based Content Validation of a Voice Prototype

Article References: Soares, V. V., Passos, F. F. D. C., Shansis, F. M., & Herbert, J. S. (2026). LLM-Based Psychiatric Interview Simulation: Technical Development and Pilot Expert-Based Content Validation of a Voice Prototype. Academic Psychiatry. https://doi.org/10.1007/s40596-026-02422-9

Image Credits: AI Generated

DOI: 10.1007/s40596-026-02422-9

Keywords: large language models, GPT-4o, psychiatric education, virtual patients, clinical simulation, psychiatric semiology, prompt engineering, medical education, voice interface, mental status examination, content validation, artificial intelligence

In-silico trials have already proven their utility across a remarkable range of cardiovascular applications, from planning Fontan surgeries for congenital heart defects to optimizing drug delivery, stent placement, and transcatheter aortic valve implantation, widely known as TAVI. Yet the field suffers from persistent methodological weaknesses. Clinical datasets used to seed these simulations are often small, demographics are inconsistently represented, uncertainty is poorly quantified, and data transparency remains limited. A virtual trial is only as trustworthy as the population it simulates, and building a convincing virtual population requires access to hundreds or thousands of detailed anatomical models that real hospitals simply cannot provide at scale. Synthetic data generation has emerged as the most practical remedy, and machine learning methods have proven the most effective among the available techniques.

Until now, the dominant tool for statistical shape analysis in this domain has been the statistical shape model, or SSM, typically built on principal component analysis. These classical models rest on two restrictive assumptions: that shape variation is essentially linear, and that every shape can be described through point-to-point correspondence with every other shape in the dataset. Both assumptions break down when confronted with the complex, nonlinear deformations of real cardiovascular anatomy, and establishing point correspondence across patient meshes is labor-intensive and error-prone. Nonlinear extensions using kernel methods and manifold learning have partially relaxed these constraints, but they introduce heavy computational overhead and generalize poorly.

The Romanian team’s framework sidesteps these limitations entirely by adopting implicit neural representations, a family of techniques that encode three-dimensional shapes as continuous mathematical functions rather than discrete collections of points or voxels. At the heart of the approach lies a signed distance field, or SDF, a function that assigns to every point in space its distance to the object’s surface, with the sign indicating whether the point lies inside or outside. The surface itself is implicitly defined as the zero-level set of this field. Because a neural SDF is continuous and resolution-independent, it avoids the discretization biases of voxel grids, the memory overhead of volumetric representations, and the topological information loss that plagues point-cloud decoders. Unlike classical mesh-based models, it is also topology-agnostic, meaning the same architecture can represent wildly different anatomical structures.

A key technical innovation concerns how the network is trained. The seminal DeepSDF method of Park and colleagues required ground truth signed distance values, which are only well-defined for watertight surfaces with unambiguous interiors. Real aortic geometries, being open tubular structures reconstructed from cardiac CT scans, are anything but watertight. The researchers instead employed implicit geometric regularization, an approach introduced by Gropp and colleagues that exploits the fact that a true signed distance function satisfies the Eikonal partial differential equation, meaning the gradient of the field has unit magnitude everywhere. By penalizing deviations from this condition at randomly sampled points in space, and by combining it with zero-level constraints on surface points and alignment with surface normal vectors, the network learns a faithful distance field from surface samples alone, with no ground truth distance values required.

Handling non-watertight geometry demanded one further component. When the network’s zero-level set is extracted from an open surface, it produces a closed shape that includes extraneous regions beyond the anatomy of interest. To solve this, the team trained an auxiliary boundary classifier that labels each point as belonging to the inlet, the outlet, or the remaining surface. Fitted planes through the classified inlet and outlet points allow the reconstructed mesh to be precisely truncated, isolating the aortic root. The complete pipeline operates as an autodecoder: each training shape receives a trainable latent vector of 128 dimensions, and a network of roughly 1.7 million parameters, organized in eight fully connected layers with softplus activations and a skip connection, maps these latent codes and spatial coordinates to distance values. At inference, new shapes are decoded through the Marching Cubes algorithm on a 256-cubed grid, then truncated and rescaled to physical dimensions.

The framework was trained on 73 peak-systole aortic surface meshes reconstructed from cardiac CT images of patients scheduled for TAVI procedures, each clinically validated by expert cardiologists. Calcifications were represented as separate meshes, and a second, conditioned implicit network was trained to generate calcium deposits attached to a given aortic shape by receiving the aortic latent vector as an additional input. Reconstruction fidelity was measured with symmetric Chamfer distance, achieving a mean of just 0.135 square millimeters for aortic roots, evidence of high-fidelity geometric representation. New synthetic patients were produced by linearly interpolating between pairs of learned latent vectors, and from this process the team generated a validation cohort of 2,500 synthetic patients stratified by gender.

The critical question is whether these synthetic hearts pass clinical muster, and the answer appears to be yes. The researchers extracted diagnostically relevant features from both real and synthetic cohorts by slicing each mesh at half-millimeter intervals along the blood flow direction, identifying four anatomically significant cross-sections: the aortic annulus, the left ventricular outflow tract, the sino-tubular junction, and the aortic sinus. They also computed aortic valve area using a custom iterative ray-casting algorithm that converged in roughly five iterations and matched ground truth CT measurements with a mean absolute error of 0.092 square centimeters. Synthetic annulus diameters averaged 24.31 plus or minus 1.65 millimeters, closely matching the 24.6 plus or minus 2.4 millimeters reported in clinical literature for aortic stenosis populations. Statistical tests, including the Mann-Whitney U test and Welch’s t-test, found no significant differences in distribution or mean for any feature, and the distribution of aortic stenosis severity, classified as mild, moderate, or severe according to valve area thresholds, aligned well between real and synthetic cohorts.

The calcification results are perhaps the most striking. Generated calcium deposits localized to regions of mechanical stress and geometric irregularity within the conditioned aortic shape, exactly where aortic stenosis pathology develops in real patients. The deposits exhibited realistic size, morphology, and density variations that correlated with stenosis severity, and histogram analysis of 500 synthetic calcifications per shape showed spatial distributions closely matching real ground truth patterns. One acknowledged trade-off is a modest reduction in variance across most anatomical measurements, an expected consequence of latent vector interpolation when generating far more shapes than the original dataset contained. The aortic valve area, notably, preserved its variability, suggesting the primary clinical indicator remains statistically intact. The authors anticipate that clinicians will eventually generate shapes from curated samples representing specific pathological conditions rather than through random interpolation, which should further attenuate the smoothing effect.

The implications extend well beyond the aortic root. The modular architecture generalizes to other cardiovascular structures, including watertight geometries such as heart chambers, where boundary classification becomes unnecessary. The learned latent vectors can serve as compact low-dimensional representations of patient anatomy, conditioning physics-informed surrogate models that predict hemodynamic quantities like transvalvular pressure gradients, paravalvular leakage flow, and wall shear stress across entire synthetic cohorts without running full-scale fluid simulations for each geometry. This could accelerate TAVI in-silico trials, where device performance is acutely sensitive to annulus diameter, sino-tubular junction dimensions, and calcification distribution. Funded by the European Union’s Horizon 2020 program under the SIMCor project, the work points toward a future where virtual patient populations, assembled on demand and tailored to specific study designs, become a standard pillar of cardiovascular device development, reducing the cost, duration, and ethical burden of bringing life-saving technologies to patients.

Subject of Research: Generative deep learning modeling of cardiovascular geometries for in-silico clinical trials

Article Title: Generative modelling of cardiovascular geometries for in-silico trials

Article References: Gasparovici, A., & Serban, A. (2026). Generative modelling of cardiovascular geometries for in-silico trials. Neural Computing and Applications, 38(18), Article 745. https://doi.org/10.1007/s00521-026-12463-8

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12463-8

Keywords: implicit neural representations, signed distance field, in-silico trials, cardiovascular modeling, aortic root, synthetic data generation, TAVI, aortic stenosis, statistical shape models, calcifications, deep learning, virtual patients