The scale of the underlying literature is striking. The team systematically searched PubMed and Web of Science for studies reporting food-derived or dietary compounds with documented effects on epigenetic pathways, filtering out reviews, meta-analyses, non-English texts, and articles lacking essential metadata. The final corpus comprises 1,963 primary research articles spanning from 1991 through August 2026, with publication counts climbing steeply in recent years. The year 2025 alone contributed 211 articles, the highest annual total in the dataset, followed by 2022 with 153 and 2024 with 150. This temporal pattern mirrors the broader surge of interest in nutriepigenomics, the study of how nutrition modulates the chemical tags and chromatin structures that govern gene expression.

Building such a database by hand would have taken years of painstaking curation, so the Mexican team, led by Aylin Del Moral-Morales and José L. Medina-Franco of the DIFACQUIM Research Group, turned to artificial intelligence. They used a large language model, the Claude API accessed through the LangChain framework, to read each article and extract structured pairs of compound names and food sources. The model was instructed to adopt the analytical stance of a medicinal chemist, annotating every dietary, beverage, or food-derived compound explicitly described as a modulator of DNA methylation, histone modification, or non-coding RNA pathways. Crucially, the temperature parameter was set to zero to suppress output variability and minimize hallucinations, and the prompts explicitly excluded vague group designations such as vitamins or polyphenols in favor of identifiable molecular entities.

Raw extraction from scientific text is only the beginning; the real challenge is making thousands of heterogeneous terms speak the same language. Food source names varied wildly across the literature, so the researchers embedded every unique dietary source string as a dense numerical vector using the all-MiniLM-L6-v2 sentence-transformer model. Agglomerative hierarchical clustering with a cosine distance threshold of 0.3 then grouped lexically and semantically similar terms, and the shortest member of each cluster became the canonical label. This semantic harmonization collapsed the messy vocabulary of the literature into 585 standardized food and beverage source terms, from tea and grapes to turmeric, soy, citrus, olive, broccoli, beef, milk, egg, and fish oil.

Compound identity received equally rigorous treatment. Every extracted compound name was mapped to the PubChem database through the PubChemPy interface, retrieving compound identifiers, IUPAC names, InChIKeys, SMILES structure strings, and curated synonyms. When a query returned multiple matches, the first valid PubChem CID was adopted as the reference. The structures then passed through RDKit, a cheminformatics toolkit that stripped salts down to their primary fragments, neutralized formal charges, canonicalized tautomers, and converted everything into a single canonical SMILES form. Records lacking a valid CID, an annotated mechanism, or a compound name were systematically excluded, ensuring that every entry in the final database is chemically unambiguous and interoperable with other resources.

To add mechanistic depth, the team integrated quantitative binding data from BindingDB, a curated knowledgebase of protein-small molecule interactions. Only records involving human proteins were retained, and each interaction had to include a numeric affinity measure in nanomolar units, such as IC50, Ki, Kd, or EC50. The target proteins were then cross-referenced against the EpiFactors database, which classifies epigenetic regulators by molecular function, so that only proteins with validated roles in epigenetic regulation survived the filter. Each binding record was annotated with functional categories such as writer, reader, or eraser of epigenetic marks, allowing researchers to trace a dietary molecule all the way from the dinner table to a specific enzymatic pocket on a chromatin-modifying enzyme.

The resulting numbers reveal where the field’s attention has concentrated. Among the 653 compounds, folate, choline, L-methionine, resveratrol, and vitamin B12 top the citation rankings, each linked to more than 100 unique literature references. Curcumin, butyrate, bisphenol A, betaine, and sulforaphane round out the most-studied list. The distribution spans micronutrients, bioactive phytochemicals such as quercetin, genistein, and epigallocatechin gallate, and even synthetic or environmental agents like valproic acid. The strong emphasis on methyl donors, B vitamins, and polyphenols reflects their well-established mechanistic ties to DNA methylation, histone modification, and histone deacetylase inhibition. On the food side, tea leads the diversity ranking with 46 unique epigenetically active compounds, ahead of grapes, turmeric, and soy, and plant-derived sources generally outpace animal-derived ones in chemical variety.

The dataset also exposes gaps that could shape future research. Of the 43 distinct epigenetic mechanisms cataloged, DNA methylation and HDAC inhibition dominate the record counts, while RNA-related modifications such as RNA methylation, demethylation, and alternative splicing remain thinly represented. Among the 80 protein targets, histone-modifying writers and erasers are the most prevalent functional classes, with 29 and 20 factors respectively, whereas Polycomb group proteins and DNA modification cofactors count only one or two members each. Binding affinity data cover 101 unique compound-target pairs, with IC50 measurements accounting for over 90 percent of the 2,877 binding records, while Ki and Kd values are comparatively rare. These imbalances offer a roadmap for underexplored territory, particularly non-histone targets and less prevalent modification types.

What sets NutriEpiDB apart from earlier compilations is its commitment to machine readability and responsible artificial intelligence use. The dataset is FAIR2-certified, meeting findability, accessibility, interoperability, and reusability standards extended with AI-readiness and responsible AI criteria. All tables are provided in open formats including CSV, Parquet, and JSON, structured according to the MLCommons Croissant standard, and annotated with schema.org and W3C PROV-O provenance metadata that trace every variable back to its data generation step. The resource is released under the Open Data Commons Attribution License, permitting unrestricted reuse with attribution, and is accessible through an interactive data portal alongside a complete data package containing raw and processed data, transformation scripts, and methodological documentation.

The authors are candid about the limitations. Because the database was built primarily from article titles and abstracts, population descriptors such as ethnicity, genetic background, and country of recruitment are largely unavailable, and the English-language restriction plus reliance on PubMed and Web of Science may underrepresent certain geographic and cultural perspectives. Selection bias toward well-studied targets like histone deacetylases is evident, and compounds lacking well-defined PubChem structures, including some metabolites and obscure phytonutrients, are necessarily absent. Most binding data derive from in vitro assays, and the authors explicitly caution that the compounds and food sources listed should not be interpreted as dietary or therapeutic recommendations, since safety, tolerability, and dose-dependent effects must be assessed individually. Still, with semi-automated update pipelines planned and interoperability with PubChem, EpiFactors, UniProt, and BindingDB maintained as those repositories evolve, NutriEpiDB is positioned to become foundational infrastructure for hypothesis generation, machine learning model development, and evidence synthesis at the intersection of nutrition, epigenetics, and computational biology.

Subject of Research: A curated database of dietary compounds with epigenetic targets built from systematic literature mining and AI-assisted extraction

Article Title: NutriEpiDB: a database for dietary compounds with epigenetic targets

Article References: NutriEpiDB: a database for dietary compounds with epigenetic targets. (n.d.). https://doi.org/10.3389/fddsv.2026.1946880

Image Credits: AI Generated

DOI: 10.3389/fddsv.2026.1946880

Keywords: nutriepigenomics, epigenetics, dietary compounds, database, large language models, DNA methylation, histone modification, binding affinity, PubChem, nutrigenomics, FAIR data, cheminformatics

The challenge stems from the nature of TCR recognition itself. Naturally occurring TCRs often bind tumor antigens with suboptimal affinity, and simply boosting binding strength can backfire, reducing specificity and raising the risk of off-target attacks on healthy tissue. Affinity enhancement strategies have historically produced receptors that cross-react with unrelated peptides, a potentially lethal side effect. Writing in Frontiers in Chemical Biology, Mario Frezzini and Daniele Narzi argue that the field needs more than raw computational power; it needs mechanistic understanding, a way of knowing not just which receptor variants bind well, but why they bind well and what molecular consequences each mutation carries.

Artificial intelligence has transformed the search for better TCRs in recent years. Deep learning models can explore vast stretches of TCR sequence space, predicting which amino acid sequences are compatible with a desired structural framework and generating libraries of candidate variants in silico. Structure-based deep learning approaches have shown genuine promise in proposing receptor variants with improved binding properties. But these tools, the authors contend, are largely data-driven and descriptive: they rely on static structural representations and scoring functions that fail to capture the dynamic, ever-fluctuating nature of protein-protein interactions. The result is an embarrassment of riches—thousands of candidates that are difficult to prioritize, and little insight into the physical determinants of recognition.

Molecular dynamics simulations sit at the opposite end of the spectrum. By explicitly modeling every atom of a TCR-peptide-MHC complex as it jiggles, flexes, and interacts in a simulated watery environment, MD reveals conformational plasticity, residue-specific energetic contributions, and the cooperative behavior of the binding interface. Decades of simulation work have shown that TCR recognition is anything but a rigid lock-and-key event; the complementarity-determining regions, the receptor’s antigen-grasping loops, fluctuate significantly, contributing to both adaptability and specificity. The catch is computational cost. Reliable MD characterization demands all-atom simulations under physiological conditions, roughly ten independent replicas per system, and sampling times of at least hundreds of nanoseconds per replica to reach statistical convergence. That makes MD suitable for analyzing only a few tens of closely related variants, not for screening hundreds of candidates.

Rather than viewing these limitations as weaknesses, Frezzini and Narzi propose exploiting them deliberately. In their framework, MD is positioned as a low-throughput, high-information filter—a step that examines a carefully chosen handful of candidates in exquisite detail. The pipeline begins with biologically grounded receptor selection: repertoire sequencing of tumor-infiltrating lymphocytes, peripheral blood, or antigen-specific T cell populations, combined with bioinformatic analysis of clonal expansion, persistence, and CDR3 sequence similarity, identifies candidate receptors with known or putative tumor reactivity. AI-driven design then explores the local sequence space around these validated scaffolds, generating variant libraries. Fast pre-screening tools—empirical scoring functions, rapid energetic estimates, or machine learning predictors—trim the candidate pool to a manageable size before the expensive physics-based refinement begins.

At the heart of the workflow, MD serves as an integrative mechanistic engine rather than a brute-force screener. Trajectory analyses quantify how flexibly the CDR loops move, which interaction patterns stabilize the complex over time, and, through residue-specific energy decomposition schemes, which individual amino acids contribute favorably to binding. Crucially, the simulations also map allosteric communication pathways—networks of dynamically coupled residues that connect the binding interface to distal regions of the receptor. Local perturbations can propagate through the structure and modulate the behavior of residues tens of angstroms away, meaning that mutations in one part of the receptor can reshape binding chemistry at another. This dual description of energetics and allostery gives the pipeline a mechanistic criterion for prioritizing variants that goes far beyond static sequence-based metrics.

The power of this approach is illustrated by the authors’ own recent computational work on a pair of closely related engineered TCRs, c796 and its parent c728. The two receptors differ by a single conservative mutation, E31D, in the CDR1 loop, yet experiments show the variant binds its target peptide-MHC complex roughly six-fold more tightly. Counterintuitively, detailed energetic analysis revealed that the mutated residue itself contributes nothing directly to the improved affinity. Instead, the single substitution triggers a redistribution of interaction energies across multiple residues, including positions in distal CDR loops that now stabilize the interface more favorably. Network analyses of simulation trajectories traced allosteric pathways linking the mutation site to these distant regions. In other words, binding affinity in TCR systems emerges from a cooperative web of interactions, and even a minimal sequence change can produce substantial functional consequences through dynamic, long-range mechanisms—precisely the kind of hidden logic that only MD-based mechanistic analysis can expose.

The framework’s final stage returns to the laboratory, where two complementary experimental technologies validate the computationally refined candidates. Surface plasmon resonance, the gold standard for biophysical characterization, directly measures the association and dissociation rates of TCR-peptide-MHC binding, providing quantitative affinity and kinetic parameters for a limited set of high-confidence variants. But SPR requires recombinant protein expression and purification, offers relatively low throughput, and cannot capture the conformational and energetic complexity that governs T cell activation. That gap is filled by cellular reporter assays such as RAPTR, which evaluate TCR activation in response to specific peptide-MHC complexes within a living cellular environment, capturing downstream signaling and functional output at higher throughput. Importantly, binding affinity alone does not necessarily correlate with functional efficacy, so these orthogonal layers of validation—molecular binding on one side, cellular function on the other—are essential for identifying receptors that are both strong and clinically useful.

What distinguishes the proposed strategy from previous pipelines is its iterative, bidirectional character. Energetic fingerprints and allosteric maps derived from MD are fed back into the AI design models, narrowing and sharpening the sequence space they explore in the next round. Experimental results, in turn, benchmark and improve the computational predictions. The authors describe this as a closed-loop, multi-scale design paradigm, building on recent demonstrations that lab-in-the-loop AI-guided optimization can work for therapeutic antibody engineering. Emerging generative tools such as AlphaFlow and BioEmu may eventually accelerate conformational sampling upstream, and the authors envision hybrid strategies in which AI-generated structural ensembles initialize diverse conformations for explicit-solvent MD refinement—though they caution that these newer methods have not yet been validated for multichain TCR-peptide-MHC complexes or mutation-induced energy changes.

The approach is not without caveats. MD-based mechanistic insight is only as reliable as the structural model underlying the simulations, and accurate models of TCR-peptide-MHC complexes remain challenging to obtain because of the intrinsic flexibility of the CDR loops and the complexity of protein-protein recognition. AI tools like AlphaFold have improved structural prediction enormously, but performance for receptor-ligand complexes is still limited without experimental structural data. The authors therefore stress that every mechanistic hypothesis generated in silico must be experimentally validated before guiding the next round of optimization. If the framework delivers on its promise, however, the payoff could be substantial: a scalable, adaptive design engine that unifies statistical learning with physical principles, capable of producing the high-affinity, exquisitely specific, and safe T cell receptors that next-generation cancer immunotherapy demands. The convergence of AI, molecular dynamics, and experimental validation may finally turn TCR engineering from an empirical art into a predictive science.

Subject of Research: Integrating artificial intelligence, molecular dynamics simulations, and experimental assays for mechanism-aware engineering of T cell receptors in cancer immunotherapy

Article Title: Bridging AI, molecular dynamics, and experiments for mechanism-aware TCR engineering in cancer immunotherapy

Article References: Frezzini, M., & Narzi, D. (2026). Bridging AI, molecular dynamics, and experiments for mechanism-aware TCR engineering in cancer immunotherapy. Frontiers in Chemical Biology, 5, Article 1909618. https://doi.org/10.3389/fchbi.2026.1909618

Image Credits: AI Generated

DOI: 10.3389/fchbi.2026.1909618

Keywords: T cell receptor engineering, cancer immunotherapy, artificial intelligence, molecular dynamics simulations, allosteric communication, peptide-MHC complex, surface plasmon resonance, RAPTR assay, protein design, adoptive T cell therapy, binding affinity, computational biology

The work, led by Ashwin Srinivasan of BITS Pilani in Goa, together with Tirtharaj Dash, A. Baskar, Michael Bain, Sanjay Kumar Dey, and Mainak Banerjee, addresses a class of hybrid neurosymbolic models that remains relatively under-explored. Rather than letting a neural generator run free and filtering its output afterward, SNG inverts the relationship. A symbolic learner first examines logical specifications of what counts as feasible data, sometimes from as few as a single example. That specification then constrains the conditional information supplied to a neural generator, which is required to reject any candidate instance that violates the symbolic description. The result is a pair consisting of a symbolic description of feasible instances and a set of newly generated instances that provably satisfy it.

Technically, the framework rests on a formal semantics built from partially ordered sets. The researchers construct what they describe as base and fibre posets, combined into an overall partial order that characterizes the space of hybrid systems. The symbolic component in their implementation draws on a restricted form of Inductive Logic Programming, a family of techniques dating back decades that learns logical rules from examples and background knowledge. The neural component is a large language model prompted to produce candidate molecules as SMILES strings, the standard text encoding of molecular structure. Each candidate is checked against the learned logical specification before it is accepted, ensuring that every generated molecule conforms to the constraints the symbolic learner has inferred.

To score candidate hypotheses, the system employs a Bayesian heuristic originally developed in the ILP literature, which balances the prior probability of a hypothesis against how well it explains positive and negative examples. The search over hypotheses proceeds through nested hyper-rectangles defined over molecular factors such as predicted binding affinity, molecular weight, and synthetic accessibility score. Sampling within these rectangles uses a Latin Hypercube approach, and each successive experiment narrows the constraints, producing hypotheses that are logically entailed by their predecessors. The authors prove formal properties about this procedure, including guarantees that the generated sets remain within the feasible extension of the learned hypothesis.

The application that gives the work its urgency is early-stage drug design, specifically the discovery of lead compounds, the validated starting points from which medicinal chemists develop actual drug candidates. The team evaluated their SNG implementation on benchmark protein targets whose biology is well understood, including JAK2 and DRD2, alongside dopamine beta-hydroxylase, or DBH, a comparatively poorly studied enzyme implicated in cardiovascular disease. Binding affinities of generated molecules were estimated computationally using the GNINA molecular docking program, which predicts how strongly a small molecule will lodge itself in the binding pocket of a protein target.

The results are striking. On the benchmark problems, where drug targets are well characterized, SNG performance was statistically comparable to state-of-the-art methods for molecular generation. More intriguing still were the exploratory experiments on DBH, a target for which little domain knowledge exists in the literature. Molecules generated by the system for this target exhibited predicted binding affinities on par with leading clinical candidates. Because the symbolic learner had to infer target-specific constraints from just five known inhibitors, the achievement suggests the framework can operate productively even in the data-sparse regimes that characterize most real drug discovery projects, where high-throughput screening data is expensive and incomplete.

The choice of neural generator turned out to matter in unexpected ways. The researchers compared OpenAI’s GPT-4o, Anthropic’s Claude 3.5 Sonnet, and a small 87-million-parameter GPT-2 model trained on roughly 480 million molecules from the ZINC chemical database. The general-purpose models performed well on heavily studied targets like JAK2 and DRD2, presumably because their vast training corpora include relevant biochemical literature. But the chemistry-specialized Molecule-GPT2 outperformed them on DBH, the lesser-studied target, where knowledge had to be extracted from the inhibitor molecules themselves rather than recalled from pretraining. Notably, the entire study cost only around ten US dollars in API fees per commercial model, underscoring that the framework is not wedded to any particular language model and could accommodate open-source alternatives.

Perhaps the most persuasive endorsement came from human experts. A structural biologist and a synthetic organic chemist reviewed the system’s outputs and found the symbolic specifications genuinely useful as preliminary filters for triaging candidates. Several of the generated molecules were judged viable for synthesis and wet-laboratory testing, a meaningful threshold for any computational design method. The experts’ assessments also highlighted a practical advantage of the neurosymbolic approach: because the system produces an explicit logical description of what makes a molecule feasible, chemists can inspect, critique, and refine the criteria rather than treating the generator as an inscrutable black box.

The authors are careful to situate their contribution within the broader neurosymbolic movement, sometimes called the third wave of neurosymbolic AI, which seeks to combine learning and reasoning rather than choosing between them. They acknowledge that the categorization of systems into purely symbolic or purely neural roles is necessarily a simplification, since induction typically requires deduction and complex reasoning involves trial-and-error learning. They also conjecture that the symbolic learner at the heart of SNG may yield more reliable hypotheses and reduce the number of experiments a laboratory robot would need to conduct in automated discovery loops. Extensions to wider classes of hybrid systems, including those where the neural component plays the primary role, are outlined in the paper’s appendices.

The implications extend well beyond one enzyme or one pharmaceutical project. By demonstrating that logical specifications learned from tiny datasets can steer a generative language model toward chemically meaningful and synthetically plausible outputs, the study offers a template for constrained generation in any scientific domain where correctness matters, from materials design to protein engineering. The researchers have released their code and data openly through a public repository, and the framework’s modest computational demands mean that laboratories without access to frontier-scale models could adopt the approach immediately. As generative AI moves from producing text to proposing tangible scientific artifacts, frameworks like SNG suggest that the path forward lies not in bigger models alone, but in models that know the rules they must obey.

Subject of Research: Hybrid neurosymbolic AI systems that integrate symbolic learning with neural generation for discovering lead molecules in early-stage drug design

Article Title: Symbolic Neural Generation with Applications to Lead Discovery in Drug Design

Article References: Srinivasan, A., Dash, T., Baskar, A., Bain, M., Dey, S. K., & Banerjee, M. (2026). Symbolic Neural Generation with Applications to Lead Discovery in Drug Design. Machine Learning, 115(9), Article 210. https://doi.org/10.1007/s10994-026-07136-5

Image Credits: AI Generated

DOI: 10.1007/s10994-026-07136-5

Keywords: neurosymbolic AI, drug discovery, large language models, inductive logic programming, molecular generation, lead discovery, machine learning, binding affinity, symbolic reasoning, drug design, GPT-4o, dopamine beta-hydroxylase