Thomas R. Caulfield

Cellular morphology captures compound-induced changes in cell state, but the relationship between the magnitude of these changes and absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties remains dependent on the endpoint and assay context. We developed MorphADMET to examine these relationships and evaluate complementary prediction and reference-retrieval tasks across six imaging datasets, seven morphological representations, eleven binary endpoints, and aqueous solubility. Endpoint comparisons used differences in median morphological response with 95% bootstrap confidence intervals from 2,000 resamples. Positive compounds showed greater responses in selected contexts, including A549–Ames, HUVEC–CYP1A2, HUVEC–CYP3A4, and HRCE–CYP2D6, whereas U2OS hERG-negative compounds showed greater responses. Structure-based ensembles predicted strong morphological responses with average precision (AP) of 0.353–0.472 and outperformed a compact descriptor baseline in four of five contexts. Models using measured morphology showed endpoint-dependent discrimination, including AP of 0.601 and AUROC of 0.786 for hERG in A549 cells. Exploratory compartment-level analyses identified differences in nuclear texture, cytoplasmic organization, and cell shape between endpoint-label groups. In the structure-based ADMET benchmark, MorphADMET achieved the highest observed AP for four of eleven classification endpoints and the lowest solubility mean absolute error, 0.660 log10 mol/L. Morphology-selected reference compounds showed higher class-balanced hERG label agreement than random references in all six cellular contexts. A CYP2D6 neighborhood illustrated shared physicochemical properties alongside partial scaffold similarity. These findings support endpoint-specific use of chemical and morphological information for compound assessment, while distinguishing overall perturbation, predictive discrimination, and retrospective reference agreement.

Quantum gravity requires a common language in which microscopic quantum structure, dynamical geometry, and the maps between them carry compatible mathematical types. We develop one from an eleven-generator real quadratic carrier of adopted signature (4,7), whose Clifford algebra Cl(4,7) has 2048 real basis blades; we call the resulting structure Rolandion algebra (R11). We establish an exact ordered-blade multiplication law, identify the 1024-dimensional even subalgebra, and separate 110 directed blade-replacement channels from the 55 metric-skew generators of so(4,7). Scale dependence is carried by a Clifford–Jordan family interpolating between the associative Clifford product and its Jordan symmetrization; the complete associator is obtained in closed form and shown, at fixed scale and for the declared regular bimodule, to be a Hochschild coboundary. Scale-labelled subalgebras with canonical defects form one category, while typed transfer restrictions generate a separate path category and presheaf topos. In a direct-sum Hilbert realization, adjoint-paired transfer blocks yield a self-adjoint Hamiltonian, unitary evolution, and exact antisymmetric population currents. The demonstrated gravitational branch is a four-dimensional Lorentzian base with an internal R11 bundle and a real covariant effective action; an eleven-dimensional geometric interpretation remains conditional. A three-qubit information-transfer model and normalized gravitational-wave, CMB, and laboratory responses illustrate constraint interfaces; no ultraviolet completeness is claimed.

We present Liquid Adaptive AI as a theoretical framework and mathematical basis for artificial intelligence systems capable of continuous structural adaptation and autonomous capability development. This work explores the conceptual boundaries of adaptive AI by formalizing three interconnected mechanisms: (1) entropy-guided hyperdimensional knowledge graphs that could autonomously restructure based on information-theoretic criteria; (2) a self-development engine using hierarchical Bayesian optimization for runtime architecture modification; and (3) a federated multi-agent framework with emergent specialization through distributed reinforcement learning. We address fundamental limitations in current AI systems through mathematically formalized processes of dynamic parameter adjustment, structural self-modification, and cross-domain knowledge synthesis, while immediate implementation faces substantial computational challenges requiring infrastructure on the scale of current large language model training facilities, we provide architectural specifications, theoretical convergence bounds, and evaluation criteria as a foundation for future research. This theoretical exploration establishes mathematical foundations for a potential new paradigm in artificial intelligence that would transition from episodic training to persistent autonomous development, offering a long-term research direction for the field. A comprehensive ^{Supplementary Materials} document provides detailed technical analysis, computational requirements, and an incremental development roadmap spanning approximately a decade.

We propose Liquid AI, an architectural framework for artificial intelligence systems capable of continuous structural adaptation and autonomous capability development. Unlike existing approaches in continual learning and neural architecture search that operate within predetermined constraints, Liquid AI implements three novel mechanisms: (1) entropy-guided hyperdimensional knowledge graphs that autonomously restructure based on information-theoretic criteria; (2) a self-development engine using hierarchical Bayesian optimization for runtime architecture modification; and (3) a federated multi-agent framework with emergent specialization through distributed reinforcement learning. Our framework addresses fundamental limitations in current AI systems—static architectures, isolated knowledge domains, and human-dependent evolution—through mathematically formalized processes of dynamic parameter adjustment, structural self-modification, and cross-domain knowledge synthesis. We present architectural specifications, theoretical convergence bounds for self-modifying systems, and evaluation criteria for adaptive AI systems. Implementation considerations address computational complexity and distributed computing requirements. This work establishes theoretical and architectural foundations for a new paradigm in artificial intelligence that transitions from episodic training to persistent autonomous development. By enabling runtime structural adaptation, cross-domain knowledge integration, and collaborative intelligence emergence, Liquid AI provides a blueprint for AI systems that continuously evolve to address complex challenges without human intervention.

Advances in genomic medicine accelerate the identification of mutations in disease-associated genes, but the pathogenicity of many mutations remains unknown, hindering their use in diagnostics and clinical decision-making. Predictive AI models are generated to combat this issue, but current tools display low accuracy when tested against functionally validated datasets. We show that integrating detailed conformational data extracted from molecular dynamics simulations (MDS) into advanced AI-based models increases their predictive power. We carry out an exhaustive mutational analysis of the disease gene PMM2 and subject structural models of each variant to MDS. AI models trained on this dataset outperform existing tools when predicting the known pathogenicity of mutations. Our best performing model, a neuronal networks model, also predicts the pathogenicity of several PMM2 mutations currently considered of unknown significance. We believe this model helps alleviate the burden of unknown variants in genomic medicine.

Selective inhibition of the more than 100 S1 family serine proteases is a long-standing challenge due to their active site similarity. Mesotrypsin, implicated in cancer progression, exemplifies these difficulties; no current inhibitors achieve selectivity over other human trypsins. We found an unexpected autoinhibited conformation of mesotrypsin via x-ray crystallography, revealing a cryptic pocket adjacent to the active site. Using high-throughput virtual screening targeting this cryptic pocket, we identified a conformationally selective small-molecule inhibitor that stabilizes the inactive state of mesotrypsin. This inhibitor demonstrates selectivity for mesotrypsin over other trypsins. Our findings challenge the accepted view of digestive trypsins as constitutively active enzymes lacking potential for allosteric regulation. Furthermore, analyses of other structures suggest that dynamic sampling of closed states with analogous allosteric cryptic pockets appears widespread among S1 serine proteases. These observations point to a potentially generalizable strategy to achieve selective inhibition, offering broad implications for drug development targeting serine proteases in cancer and other diseases.

Alzheimer’s disease (AD) is marked by the pathological accumulation of amyloid beta-42 (Aβ42), contributing to synaptic dysfunction and neurodegeneration. While extracellular amyloid plaques are well-studied, increasing evidence highlights intracellular Aβ42 as an early and toxic driver of disease progression. In this study, we present a novel, Generative AI–based drug design approach to promote targeted degradation of Aβ42 via the ubiquitin–proteasome system (UPS), using E3 ligase–directed molecular glues. We systematically evaluated the ternary complex formation potential of Aβ42 with three E3 ligases (CRBN, VHL, and MDM2) through structure-based modeling, ADMET screening, and docking. We then developed a Ligase-Conditioned Junction Tree Variational Autoencoder (LC-JT-VAE) to generate ligase-specific small molecules, incorporating protein sequence embeddings and torsional angle-aware molecular graphs. Our results demonstrate that this generative model can produce chemically valid, novel, and target-specific molecular glues capable of facilitating Aβ42 degradation. This integrated approach offers a promising framework for designing UPS-targeted therapies for neurodegenerative diseases.

Alzheimer’s disease (AD) is marked by the pathological accumulation of amyloid beta-42 (Aβ42), contributing to synaptic dysfunction and neurodegeneration. While extracellular amyloid plaques are well-studied, increasing evidence highlights intracellular Aβ42 as an early and toxic driver of disease progression. In this study, we present a novel, AI-assisted drug design approach to promote targeted degradation of Aβ42 via the ubiquitin-proteasome system (UPS), using E3 ligase–directed molecular glues. We systematically evaluated the ternary complex formation potential of Aβ42 with three E3 ligases—CRBN, VHL, and MDM2—through structure-based modeling, ADMET screening, and docking. We then developed a Ligase-Conditioned Junction Tree Variational Autoencoder (LC-JT-VAE) to generate ligase-specific small molecules, incorporating protein sequence embeddings and torsional angle-aware molecular graphs. Our results demonstrate that this generative model can produce chemically valid, novel, and target-specific molecular glues capable of facilitating Aβ42 degradation. This integrated approach offers a promising framework for designing UPS-targeted therapies for neurodegenerative diseases.

Advances in genomic medicine have accelerated the identification of mutations in disease- associated genes, but the pathogenicity of many mutations remains unknown, hindering their use in diagnostics and clinical decision-making. Predictive AI models have been generated to combat this issue, but current tools display low accuracy when tested against functionally validated datasets. We show that integrating detailed conformational data extracted from molecular dynamics simulations (MDS) into advanced AI-based models can increase their predictive power. We carried out an exhaustive mutational analysis of the disease gene PMM2 and subjected structural models of each variant to MDS. AI models trained on this dataset outperformed existing tools when predicting the known pathogenicity of mutations. Our best performing model, a neuronal networks model, was also able to predict the pathogenicity of several PMM2 mutations currently considered of unknown significance. We believe this new model will help alleviate the burden of unknown variants in genomic medicine.

Mutations in the PINK1 and PRKN genes are the most frequent genetic cause of early-onset Parkinson disease. The pathogenic p.R275W substitution in PRKN is the most frequent substitution observed in patients, and thus far has been characterized mostly through overexpression models that suggest a possible gain of toxic misfunction. However, its effects under endogenous conditions are largely unknown. We used patient fibroblasts, isogenic neurons, and post-mortem human brain samples from carriers with and without PRKN p.R275W to assess functional impact. Immunoblot analysis and immunofluorescence were used to study mitophagy activation, and mitophagy execution was analyzed by flow cytometry of the reporter mitoKeima. The functional analysis was accompanied by structural investigation of PRKN p.R275W. We observed lower PRKN protein in fibroblasts with compound heterozygous p.R275W mutations. Isogenic neurons showed an allele-dose dependent decrease in PRKN protein. Lower PRKN protein levels were accompanied by diminished phosphorylated ubiquitin and decreased MFN2 modification. Mitochondrial degradation was also allele-dose dependently impaired. Consistently, PRKN protein levels were drastically reduced in human brain samples from p.R275W carriers. Finally, structural simulations showed significant changes in the closed form of PRKN p.R275W. Our data suggest that under endogenous conditions the p.R275W mutation results in a loss-of-function by destabilizing PRKN.