MorphoVerse: Shape-Space & Indian Poetry

• MorphoVerse is a unified framework that defines an infinite-dimensional shape space integrating cellular geometry, epigenetic markers, and quantitative morphogenesis.
• It operationalizes an epigenetic code through cell event graphs to simulate key processes like cell division, movement, and differentiation.
• MorphoVerse also encompasses a curated dataset of 1,570 Indian-language poems, supporting robust translation and image generation for cross-modal synthesis.

MorphoVerse denotes both a mathematical "shape-space" formalism for organismal morphogenesis and a morphologically rich, annotated corpus of Indian-language poetry designed for computational translation and multimodal generation. The term integrates theoretical frameworks from mathematical biology (Morozova et al., 2014), software models for epigenetic-driven development (Bessonov et al., 2019), and a large-scale dataset for poetry and image generation (Jamil et al., 17 Nov 2025). MorphoVerse serves as a nexus where concepts of shape, epigenetic code, cell event graphs, and linguistic structure coalesce for rigorous quantitative analysis and generative modeling.

1. Mathematical Foundation: Shape-Space Formalism

MorphoVerse, in its foundational sense, is defined as the infinite-dimensional space of all possible organismal shapes, where each "point" encodes both fine-grained cellular geometry and epigenetic state, as well as coarse-grained global metric properties (Morozova et al., 2014). At the cellular level, every eukaryotic cell $C$ is modeled as a star-convex subset of $\mathbb{R}^3$ with respect to its microtubule-organizing center (MTOC) $p_C$. The extent of $C$ is given by a radial function $r_C:S^2\rightarrow\mathbb{R}_{>0}$ and its surface parameterization $R_C(\xi)=p_C+r_C(\xi)\xi$ for $\xi\in S^2$.

Each cell also carries a set of real-valued marker fields $f^C_i:S^2\rightarrow\mathbb{R}$, representing local densities of epigenetic surface markers. The complete cellular state $F_C=(r_C,f^C_1,\ldots,f^C_N)\in [L^2(S^2)]^{N+1}$ allows quantification of differences via the natural $L^2$ metric.

Organismal assemblies %%%%10%%%% at time $t$ comprise finite unions of such cells, with centers $\{x_k=p_{C_{x_k}}\}$. The measured metric space $(\mathcal{O}_t,d_t,\mu_t)$ models the organism globally, where $\mu_t$ is a measure over cell centers. Over configuration space $\Gamma_n(\mathbb{R}^3)$, a vector bundle $E^{(n)}$ with Hilbert space fiber collects all cellular states.

An action principle on paths through this bundle governs morphogenetic dynamics:

$$S[\mathcal{O}_t]=\int_{t_0}^{t_1} L((\mathcal{O}_t,d_t,\mu_t),\dot{\mathcal{O}}_t)\,dt,$$

splitting into "kinetic" and "potential" terms, with the latter quantifying deviation from ideal forms via Gromov-Hausdorff-type distances. The resulting Euler-Lagrange equations model both cell motion and evolution of cellular marker fields.

2. Epigenetic Code Framework and Cell Event Graphs

Within the software implementation (Bessonov et al., 2019), MorphoVerse operationalizes the hypothesis that epigenetic spectra drive cell behavior. Each cell has a real-valued vector $E=(e_1,\ldots,e_n)$, summarizing surface marker concentrations. A mapping $f_j(E)$ computes propensities for five cell events: division (DIV), growth (GROW), death (DEATH), movement (MOVE), and differentiation (DIFF). The maximal propensity determines the instructive signal $S$ issued each step.

Spectrum transformation laws $T_S$ update $E$ according to the selected event, with division and growth operating along defined axes and inheritance constraints, movement affecting spatial position, and differentiation or apoptosis "locking in" or zeroing the spectrum.

Development is formalized as a directed graph $G=(V,\mathcal{E})$, where vertices are cell states $(E,\mathbf{x},\sigma)$ and edges are labeled by the cell events connecting states. In the absence of movement, this yields a rooted tree structure.

3. Corpus: MorphoVerse Dataset of Indian Poetry

MorphoVerse also designates a curated dataset of 1,570 morphologically rich Indian-language poems across 21 languages (Jamil et al., 17 Nov 2025). It covers Indo-Aryan, Dravidian, Tibeto-Burman, and Austro-asiatic families. Data acquisition utilized open-web sources, blogs, and archives, with annotation and verification by expert undergraduates, and human-vetted English translations for low-resource cases.

Preprocessing involved normalization, cleaning, and annotator agreement checks (Cohen's $\kappa>0.80$). Tokenization was performed using LLM-matched schemes (Byte-Level BPE or SentencePiece), preserving morphological diversity and linguistic richness.

MorphoVerse Dataset Key Statistics

Feature	Value/Description	Source/Method
Number of poems	1,570	Scraped/verified
Number of languages	21	Family coverage
Annotation agreement	Cohen's $\kappa>0.80$	Annotator voting
Morphological richness	Flagged inflections, compounds	Manual validation

4. Translation and Multimodal Generation: TAI Framework

The TAI (Translation and Image Generation) pipeline leverages the MorphoVerse dataset to test LLM-based translation and multimodal (text-to-image) synthesis (Jamil et al., 17 Nov 2025). The central module for translation is Odds-Ratio Preference Alignment (ORPO), which fine-tunes an LLM $\theta$ to maximize the odds of generating preferred translations $y_w$ over less preferred $y_l$:

$$\text{odds}_\theta(y|x) = \frac{P_\theta(y|x)}{1-P_\theta(y|x)}, \qquad OR_\theta(y_w, y_l) = \frac{\text{odds}_\theta(y_w|x)}{\text{odds}_\theta(y_l|x)}$$

The loss combines standard SFT and OR penalty:

$$L_{ORPO} = \mathbb{E}_{(x,y_w,y_l)}[L_{SFT}(\theta;x,y_w) + \lambda\,L_{OR}(\theta;x,y_w,y_l)]$$

For image generation, translated poems are mapped into semantic graphs, where nodes are tokens annotated with lemmas and WordNet synsets, and edges encode dependencies and hypernym relations. Clusters (thematic/metaphorical) are identified via modularity optimization. Prompts for diffusion models are synthesized by LLMs from these semantic graphs.

Diffusion conditioning follows standard procedures, with prompt embedding $e=E_{prompt}(I)$ steering the denoising trajectory in stable diffusion backbones.

5. Empirical Evaluation and Morphological Impact

Quantitative experiments (Jamil et al., 17 Nov 2025) on the MorphoVerse dataset yield the following translation and image alignment results for the best-performing pipeline (Gemma-2 + ORPO):

Metric	Score
ROUGE-1	0.6922
ROUGE-2	0.4451
ROUGE-L	0.6340
BLEU-4	0.2864
METEOR	0.5693
COMET	0.4034
Long-CLIP	0.2436
BLIP	0.4613
ImageReward	0.5342

Human evaluation involved expert rating of generated images on semantic fidelity, visual completeness, and cultural authenticity, supporting that iterative prompt refinement enhances scores (initial $\approx2.7$, stabilized $\approx4.1$ out of 5).

Qualitative analyses show that ablation of ORPO or semantic graph prompts leads to literal translations, loss of metre/metaphor, and culturally impoverished images, while TAI captures poetic motifs and cultural nuances effectively.

6. Theoretical and Practical Significance

MorphoVerse provides a quantitative language for morphogenesis, regeneration, and comparative morphology (Morozova et al., 2014, Bessonov et al., 2019), allowing analysis of developmental trajectories as geodesics in shape-space, encoding minimal morphogenetic "work" by Gromov-Hausdorff distance. In computational linguistics, it facilitates robust cross-modal translation and generation (Jamil et al., 17 Nov 2025), supporting both SDG 4 (Quality Education) and SDG 10 (Reduced Inequalities).

A plausible implication is the convergence of mathematical biology, computational modeling, and cross-cultural linguistics within a unified formalism, enabling hypothesis testing, comparative studies, and synthetic generation from cell shape to poetic imagery.

7. Open Challenges and Future Directions

Current limitations include integrating fully end-to-end multimodal models that ingest raw semantic graphs for image generation, deeper evaluation of prosody and rhyme, expansion to additional scripts and oral traditions, and refinement of artist style transfer techniques with cultural preservation (Jamil et al., 17 Nov 2025). In morphogenetic modeling, extending the quantitative framework to incorporate stochasticity, multi-scale tissue architecture, and real-world validation remain open research directions. This suggests the potential for MorphoVerse to underpin next-generation generative, comparative, and analytical tools across both the biological and computational humanistic sciences.