Regenerative Islet Research¶

Purpose and boundary¶

IINTS-AF can become a useful computational workbench for studying stem-cell-derived islets, graft survival, and protein-level evidence. Its role must remain precise:

IINTS-AF can organize evidence, compare cell states, generate hypotheses, propagate uncertainty, and stress-test transparent graft models. It cannot discover or validate a treatment by itself.

This module is for pre-clinical research and education. It does not design a clinical therapy, prescribe treatment, predict transplant success for a person, or replace cell-biological, immunological, toxicological, animal, or clinical validation.

The research question is therefore not "which protein cures type 1 diabetes?" It is:

Which measurable molecular and cellular properties distinguish a mature, glucose-responsive, stress-resistant, immunologically manageable, and growth-controlled stem-cell-derived islet graft from an inadequate or unsafe graft?

What current evidence supports¶

The field has shown that cell replacement can restore measurable islet function, but the central engineering and safety problems are not solved.

A 2025 phase 1-2 study of the allogeneic stem-cell-derived product zimislecel reported restored islet function in a small, short-term cohort. Participants received immunosuppression, so this does not solve immune rejection or establish long-term safety and durability. See the NEJM report and NCT04786262.
A separate 2025 first-in-human report found short-term survival of gene-edited allogeneic donor islet cells without immunosuppression in one person. It was a single proof-of-concept recipient, the published follow-up was 12 weeks, and the cells were donor islets rather than a validated stem-cell-derived product. It must not be generalized into a cure claim. See the NEJM report.
Stem-cell-derived islets can achieve advanced glucose-responsive function, but metabolic and transcriptional differences from primary adult islets can remain. See the Nature Biotechnology maturation study.
Single-nucleus RNA and chromatin analysis has identified incomplete lineage specification and off-target enterochromaffin-like states in SC-islets. See the Nature Cell Biology multi-omics study.
Immune-evasion candidates such as the IFN/JAK/STAT-CXCL10 axis have pre-clinical evidence, but reduced immune visibility can create competing risks involving infection surveillance, NK-cell recognition, and abnormal cell growth. See the whole-genome CRISPR screen.

These findings support a multi-objective research platform. They do not support one protein, one structure, or one model score as a therapeutic answer.

Why AlphaFold is one layer, not the answer¶

AlphaFold is valuable for structural hypotheses, domain inspection, mutation localization, and identifying regions where a predicted model is uncertain. AlphaFold pLDDT and PAE represent prediction confidence. They do not measure:

expression or protein abundance;
post-translational modification or proteoforms;
secretion, trafficking, or membrane localization;
binding affinity or kinetic rate constants;
pathway causality;
beta-cell identity, maturation, or glucose-stimulated insulin secretion;
immune rejection, tumorigenicity, graft survival, or patient benefit.

IINTS-AF therefore applies this hard rule:

\[ \mathrm{pLDDT},\mathrm{PAE}\;\not\Rightarrow\; \text{physiological parameter, efficacy, or safety} \]

An AlphaFold observation may create a testable structural hypothesis. It may not automatically alter a patient or graft simulation.

Proposed evidence architecture¶

flowchart LR
    A[Stable identifiers<br/>UniProt / Ensembl] --> B[Protein evidence registry]
    C[Structure<br/>RCSB PDB / AlphaFold] --> B
    D[Proteomics<br/>PRIDE / ProteomeXchange] --> B
    E[Cell and tissue context<br/>CELLxGENE / HPAP / HPA / Bgee / GTEx] --> B
    F[Pathways and interactions<br/>Reactome / STRING] --> B
    G[Measured affinity<br/>BindingDB / ChEMBL] --> B
    H[Trials and publications<br/>ClinicalTrials.gov / primary literature] --> B

    B --> I[Cell-state comparison<br/>SC-islet vs primary adult islet]
    I --> J[Multi-objective evidence report]
    J --> K{Reviewed assay-to-parameter mapping?}
    K -- No --> L[Hypothesis only<br/>no simulation change]
    K -- Yes --> M[Bounded parameter distribution<br/>with provenance and uncertainty]
    M --> N[IINTS graft simulation]
    N --> O[Sensitivity, identifiability,<br/>validation and limitations report]

The review gate between the evidence report and simulation is deliberate. It prevents a database score, language model, or attractive protein structure from becoming an unsupported physiological constant.

Core protein panels¶

The bundled registry is stored in src/iints/data/regenerative_protein_panels.json. UniProt accessions are used as stable protein identifiers; gene symbols are labels, not primary keys.

Panel	Initial proteins	Question
Beta-cell identity and function	INS, PDX1, NKX6-1, MAFA, GCK, ABCC8, KCNJ11, PCSK1, PCSK2, SLC30A8	Are cells correctly specified, mature, and able to sense glucose, process proinsulin, and secrete insulin?
Immune visibility and evasion	B2M, CIITA, CXCL10, CD274, CD47, HLA-E	How are T cells and NK cells engaged, and what surveillance risks accompany reduced immune visibility?
Stress survival and graft support	HIF1A, VEGFA, HSPA5, TXNIP, NFE2L2	Can cells tolerate hypoxia, ER stress, oxidative stress, inflammation, and delayed vascularization?
Residual pluripotency and growth safety	POU5F1, NANOG, SOX2, MKI67	Are undifferentiated or abnormally proliferating cells detectable and controlled over time?

These are evidence panels, not intervention recommendations. A marker may be useful for measurement without being a safe or effective engineering target.

Python callers can inspect a panel and its required evidence sources without performing any automatic efficacy scoring:

from iints.research.regenerative_islet import build_regenerative_evidence_plan

plan = build_regenerative_evidence_plan("beta_cell_identity_and_function")
print(plan.to_dict())

Proteomics dataset ingestion (PRIDE / MaxQuant / DIA-NN)¶

To transform raw or processed mass spectrometry matrices into the strict comparator contract, use iints research regenerative import-proteomics:

# 1. Ingest MaxQuant proteinGroups.txt
iints research regenerative import-proteomics \
  --input-file data/maxquant/proteinGroups.txt \
  --sample-metadata data/sample_annotations.csv \
  --source-id PXD001539 \
  --format maxquant \
  --output-csv data/standardized_islet_proteomics.csv

# 2. Ingest DIA-NN / Spectronaut report.tsv
iints research regenerative import-proteomics \
  --input-file data/diann/report.tsv \
  --sample-metadata data/sample_annotations.csv \
  --source-id PXD064528 \
  --format diann \
  --output-csv data/standardized_islet_proteomics.csv

# 3. Ingest wide matrix (TSV/CSV with sample columns)
iints research regenerative import-proteomics \
  --input-file data/matrix.tsv \
  --sample-metadata data/sample_annotations.csv \
  --source-id PXD001539 \
  --format wide_matrix \
  --output-csv data/standardized_islet_proteomics.csv

Sample metadata files (sample_annotations.csv / .tsv / .json) define sample-level experimental groups:

sample_id,group,batch_id,source_id
SC_Sample_1,sc_islet,batch_A,PXD001539
SC_Sample_2,sc_islet,batch_A,PXD001539
SC_Sample_3,sc_islet,batch_B,PXD001539
Primary_Sample_1,primary_islet,batch_A,PXD001539
Primary_Sample_2,primary_islet,batch_A,PXD001539
Primary_Sample_3,primary_islet,batch_B,PXD001539

Current comparator workflow¶

The SDK includes a descriptive SC-islet versus reference-islet protein comparator:

iints research regenerative panels

iints research regenerative compare \
  --dataset data/standardized_islet_proteomics.csv \
  --panel beta_cell_identity_and_function \
  --normalization-note "Joint median normalization within one experiment" \
  --output-dir results/regenerative/protein_comparison

The input is a protein-level CSV or Parquet table. It requires these columns:

Column	Meaning
`gene_symbol`	Protein-coding gene symbol matching a bundled panel
`group`	Experimental group; defaults are `sc_islet` and `primary_islet`
`sample_id`	Independent biological sample identifier
`value`	Finite protein-level abundance after documented preprocessing
`unit`	Measurement unit or normalized-intensity label
`scale`	Exactly `linear` or `log2`
`source_id`	Dataset or study provenance identifier
`batch_id`	Strongly recommended batch identifier

Peptide rows must be aggregated to one protein-level value per gene and sample before comparison. Units and scales must match between groups. The comparator does not normalize unrelated studies automatically; different source_id sets force a review_required result.

Outputs include:

protein_comparison.csv with medians, descriptive log2 differences, sample counts, and bootstrap intervals when both groups have at least three samples;
comparison_report.json with input hash, provenance, warnings, panel coverage, and explicit non-clinical boundaries;
comparison_report.md for scientific review;
comparison_forest.html when Plotly is installed.

A forest plot is used instead of a radar chart. A radar chart can conceal missing proteins, mix unrelated scales, and visually imply that the axes form one validated biological score.

The comparator deliberately does not calculate proposed shortcuts such as MAFA / (CHGA + TPH1): CHGA is a broad neuroendocrine marker and can be present in desired endocrine cells, so placing it in an "immaturity" denominator is not defensible. It also does not collapse PD-L1, CD47, HLA, and NK-cell risk into one immune score. Those observations remain separate until validated functional immune assays support a specific model.

Evidence tiers¶

Every observation should retain its source, organism, cell line or donor, assay, units, batch, processing method, and uncertainty. IINTS-AF should not collapse unlike evidence into one confidence number.

Tier	Evidence	Permitted conclusion
1	Peer-reviewed human clinical evidence	Translational observation within the exact studied population, intervention, dose, site, follow-up, and co-treatment
2	Primary human islet or human ex-vivo evidence	Human biological relevance under the tested assay conditions
3	Human stem-cell-derived islet in-vitro evidence	Cell-product behavior for the tested line, protocol, stage, batch, and challenge
4	In-vivo animal or humanized-model evidence	Pre-clinical graft behavior within that model; not direct patient efficacy
5	Association, network, predicted structure, or in-silico model	Hypothesis generation only

Contradictory findings remain visible. Missing evidence remains missing; it is not imputed by a local AI model.

Multi-objective evaluation¶

A single "cure score" would hide dangerous trade-offs. The workbench should retain a vector of outcomes:

\[ \mathbf{y}=\left[ I_{identity}, F_{GSIS}, R_{stress}, V_{vascular}, E_{immune}, S_{growth} \right] \]

where each component has its own assay definition, uncertainty, and evidence tier. Candidate hypotheses are compared as a Pareto set rather than ranked by an arbitrary weighted sum. For example, improved immune evasion cannot cancel evidence of uncontrolled proliferation.

Rules for connecting assays to simulation¶

Only a measurable assay with a reviewed mapping may affect the transplant model. The mapping must be bounded, unit-aware, versioned, and uncertainty preserving:

\[ \theta_{model}\sim p\left(\theta\mid y_{assay},\;context,\;provenance,\;uncertainty\right) \]

Examples of potentially defensible future mappings are:

Measured evidence	Possible model state or parameter	Required validation
Dynamic GSIS or perifusion curve, C-peptide, proinsulin/insulin ratio	secretion capacity, glucose threshold, secretion delay	Independent primary-islet benchmark and repeated lines/batches
Viability under quantified hypoxia	hypoxia death-rate distribution	Oxygen measurement, exposure duration, cell density, and external validation
Cytokine or immune-cell challenge survival	inflammatory or adaptive death-rate distribution	Defined effector cells, donor replication, cytokine concentrations, and in-vivo comparison
Perfusion, vascular density, or oxygen recovery	vascularization and oxygen time constants	Placement-specific longitudinal graft data
Longitudinal residual-pluripotency and proliferation assays	safety gate only	It must never be traded against glucose performance

The following mappings remain forbidden:

pLDDT or PAE to insulin sensitivity, graft function, or therapy efficacy;
STRING confidence to a causal effect size;
one protein abundance measurement to a patient outcome;
ClinVar classification to an effect size without functional evidence;
language-model text to a numerical model parameter;
an immune-evasion marker to a safety conclusion.

Data stack¶

The first reproducible evidence bundle should combine complementary sources:

Canonical protein identity: UniProt sequence, isoform, domain, location, and reviewed annotation.
Structure: experimental complexes from RCSB PDB first; AlphaFold models and confidence where experimental structures are absent.
Proteomics: PRIDE/ProteomeXchange abundance and subcellular evidence. Useful starting records are PXD001539, a quantitative human beta-cell proteome, and PXD064528, a cross-species subcellular islet atlas. The latter concerns type 2 diabetes remodeling and must not be mislabeled as direct type 1 diabetes evidence.
Cell state: CELLxGENE/HPAP and published SC-islet versus primary-islet single-cell and multi-omic references.
Pathway context: Reactome and STRING. STRING edges can represent several evidence channels and are not automatically physical binding or causality.
Measured interactions: BindingDB and ChEMBL, preserving assay type and units. Kd, Ki, IC50, and EC50 must not be silently interchanged.
Translational state: ClinicalTrials.gov and peer-reviewed clinical publications, with intervention, immunosuppression, sample size, follow-up, adverse events, and funding retained.

Raw patient-level, donor-level, or restricted data must not be committed to the public GitHub repository. The MDMP layer should record licenses, consent and use constraints, checksums, transformations, and source versions.

First experiments for IINTS-AF¶

1. Maturity gap benchmark¶

Compare primary adult human beta cells with several SC-islet stages and batches using identity, glucose-sensing, proinsulin-processing, secretion, and off-target cell-state evidence. Hold out entire cell lines and differentiation batches; randomly splitting cells from the same batch would create leakage.

2. Stress-resilience benchmark¶

Compare baseline, hypoxia, ER-stress, oxidative-stress, and inflammatory conditions. Report function and viability together. A condition that preserves insulin output briefly while increasing stress or death markers is not an improvement.

3. Immune trade-off map¶

Build separate evidence axes for T-cell activation, NK-cell activation, chemokine signaling, complement, graft function, and long-term growth safety. Do not optimize for "invisibility" alone.

4. Assay-to-graft sensitivity study¶

After independent review of a mapping, propagate a distribution rather than a point estimate through iints.research.stem_cell_transplant. Report which graft conclusions are robust and which are dominated by uncertain parameters.

5. Safety-panel gate¶

Residual pluripotency, unexpected cell identity, abnormal proliferation, and longitudinal growth must be hard gates. Better glucose metrics cannot override a failed safety gate.

Implementation roadmap¶

Phase 0 - now present¶

machine-readable, validated protein panels;
stable UniProt identifiers;
explicit evidence-source requirements;
code-level rejection of direct protein-to-physiology mappings;
tests that preserve AlphaFold and causality boundaries;
unit-aware SC-islet versus reference protein comparison with provenance, missingness, replication, and cross-source review gates;
machine-readable and Markdown comparison reports plus an optional forest plot;
the existing transparent multi-compartment graft simulator.

Phase 1 - evidence connectors¶

read-only provider adapters for UniProt, RCSB PDB, AlphaFold, PRIDE, ProteomeXchange, CELLxGENE, Reactome, STRING, BindingDB, ChEMBL, and ClinicalTrials.gov;
immutable raw response, normalized record, request parameters, timestamp, source version, license, and checksum;
cache and offline replay so an analysis is reproducible.

Phase 2 - cell-state comparison¶

extend the current protein comparator beyond descriptive abundance;
primary-human-islet reference profiles;
SC-islet stage, line, batch, and challenge metadata;
single-cell quality control, pseudobulk summaries, donor-aware statistics, and off-target cell detection;
proteomic and transcriptomic evidence kept as separate modalities.

Phase 3 - reviewed model bridge¶

an assay-to-parameter registry with units, equations, parameter bounds, source evidence, reviewer, and version;
uncertainty propagation, global sensitivity, and parameter-identifiability analysis;
no mapping when evidence is unknown, conflicting, or context-mismatched.

Phase 4 - external validation¶

preregistered endpoints and locked analysis plans;
leave-one-line, leave-one-batch, and leave-one-dataset-out validation;
independent wet-lab and transplant-science collaborators;
comparison against primary human islets and external graft studies;
versioned negative results and failure cases, not only successful figures.

What IINTS-AF should not build automatically¶

The SDK should not generate actionable gene-editing recipes, wet-lab protocols, clinical dose recommendations, patient selection, or autonomous treatment plans. Local AI may summarize cited evidence and identify contradictions, but it may not invent measurements, assign physiological constants, or approve a hypothesis.

Success criteria¶

This direction is successful when an external researcher can answer:

Which exact data and evidence support each claim?
Is an observation structural, molecular, cellular, pre-clinical, or clinical?
Which cell line, batch, donor, assay, and context produced it?
What contradictory or missing evidence remains?
Which assumptions connect an assay to the IINTS graft model?
How sensitive is the conclusion to those assumptions?
Did every identity, immune, function, stress, and growth-safety gate pass?

That would make IINTS-AF a transparent hypothesis and validation workbench for regenerative diabetes research, rather than a system that overclaims a cure.