Agentic AI-driven modulation of single-cell foundation model attention (ELISA/scGPT) will reveal coalition-based regulatory deviations in confluent tissue dynamics (active foam model), where persistent Brownian motions synchronize with co-expression clusters rather than individual gene signals.
Agentic AI-driven modulation of single-cell foundation model attention (ELISA/scGPT) will reveal coalition-based regulatory deviations in confluent tissue dynamics (active foam model), where persistent Brownian motions synchronize with co-expression clusters rather than individual gene signals.
Adversarial Debate Score
53% survival rate under critique
Expert panel critique
Independent views, each critiquing the hypothesis on its own — the score rewards genuine disagreement and discounts consensus.
Supporting Research Papers
- Systematic Evaluation of Single-Cell Foundation Model Interpretability Reveals Attention Captures Co-Expression Rather Than Unique Regulatory Signal
We present a systematic evaluation framework - thirty-seven analyses, 153 statistical tests, four cell types, two perturbation modalities - for assessing mechanistic interpretability in single-cell fo...
- ELISA: An Interpretable Hybrid Generative AI Agent for Expression-Grounded Discovery in Single-Cell Genomics
Translating single-cell RNA sequencing (scRNA-seq) data into mechanistic biological hypotheses remains a critical bottleneck, as agentic AI systems lack direct access to transcriptomic representations...
- Causal Circuit Tracing Reveals Distinct Computational Architectures in Single-Cell Foundation Models: Inhibitory Dominance, Biological Coherence, and Cross-Model Convergence
Motivation: Sparse autoencoders (SAEs) decompose foundation model activations into interpretable features, but causal feature-to-feature interactions across network depth remain unknown for biological...
- Sparse autoencoders reveal organized biological knowledge but minimal regulatory logic in single-cell foundation models: a comparative atlas of Geneformer and scGPT
Background: Single-cell foundation models such as Geneformer and scGPT encode rich biological information, but whether this includes causal regulatory logic rather than statistical co-expression remai...
Literature Assessment
An LLM's reading of the literature — not computational verification.
AI models reveal complex gene interactions but face challenges in individual signal relevance.
Method: literature_meta · Result: inconclusive
Formal Verification
Z3 checks whether the hypothesis is internally consistent, not whether it is empirically true.
This discovery has a Claude-generated validation package with a full experimental design.
Precise Hypothesis
When a single-cell foundation model (scGPT or equivalent transformer trained on scRNA-seq) is subjected to agentic, iterative attention-head perturbation across a confluent tissue-derived single-cell atlas, the resulting attention-deviation maps will cluster into gene coalitions (co-expression modules of ≥5 genes, pairwise co-attention correlation r≥0.5) that (a) are statistically distinguishable from single-gene attention saliency (permutation p<0.01), (b) are reproducible across ≥3 independent random seeds/agent runs (Jaccard overlap ≥0.6 on module gene membership), and (c) recover the empirically-derived CA-RIM CD8+ T-cell axis (DNMT1–ZNF740–BRD3–CTSS–STAT3–IFNG network) as a top-decile coalition ranked by coalition-attention-deviation score, at a rate exceeding that of coalitions recovered from randomly permuted cell-type labels (ΔAUC ≥0.15).
- Attention-deviation-derived coalitions show <0.6 Jaccard overlap across seeds (i.e., non-reproducible artifacts of stochastic training/inference).
- Coalition membership is statistically indistinguishable (permutation p>0.05) from coalitions derived from label-shuffled or randomly-initialized-model controls.
- The known CA-RIM axis (DNMT1/ZNF740/BRD3/CTSS/STAT3/IFNG) does not appear in the top quartile of ranked coalitions in ≥2 of 3 independent atlases (GSE193770, GSE138614, CELLxGENE Census subset).
- Coalition-based attention signal provides no incremental predictive value over simple co-expression correlation (Spearman) on held-out cells — i.e., ΔAUC <0.05 vs. classical WGCNA/Leiden co-expression baseline.
- Agentic modulation loop fails to converge (attention perturbation objective non-monotonic / oscillating) in >30% of runs.
Spine & Adversarial Read
- highAttention weights in transformer models are widely criticized in NLP/ML literature as unreliable proxies for feature importance ('attention is not explanation'); building a biological discovery claim on attention-deviation clustering risks mistaking architecture-specific artifacts for coalition biology.Partially addressed by the WGCNA co-expression control and cross-architecture (scGPT vs Geneformer) replication step, which tests whether the signal is architecture-independent. However, the EVP does not include a gradient-based or perturbation-based alternative attribution method (e.g., integrated gradients, SHAP) as a third independent check — this gap should be closed before external submission.
- highWhy scGPT/Geneformer and agentic head-masking specifically, rather than simpler, cheaper, more established methods (e.g., direct co-expression network analysis, which the underlying Goodman 2026 analysis already used via STRING proximity)? The methodology choice needs an explicit justification for why attention-based coalition detection adds information beyond the existing composite score/network proximity pipeline already run in Phase 3/4.The protocol does include a direct head-to-head comparison (ΔAUC vs. WGCNA baseline, success threshold ≥0.10) which operationalizes this justification empirically rather than asserting it a priori — if the ΔAUC gate is not met, the method should be judged to add no value over existing simpler pipelines, which is an honest built-in falsification path rather than an unstated assumption. This is the correct design but means the justification is contingent on the outcome, not established in advance.
- mediumThe single-cell datasets used (GSE193770, GSE108000, GSE138614) are modest in size (thousands to tens of thousands of cells) for training/probing a foundation model's attention meaningfully at the fine-tuning stage; overfitting during the ≤5-epoch fine-tune could itself manufacture spurious, non-generalizable attention patterns that then appear reproducible only because they overfit the same small sample across seeds.Not fully resolved — the protocol specifies light fine-tuning and cross-dataset validation (step 10) which mitigates but does not eliminate this risk. A held-out patient-level (not cell-level) split for fine-tuning versus attention-probing was not explicitly specified and should be added as a methodological correction before execution.
Experimental Protocol
Minimum viable test (MVT): single dataset (GSE193770, CD8+ T cells, ~8,000–12,000 cells after QC), single foundation model (scGPT human checkpoint), 3 random seeds, comparing agentic-attention coalitions vs. (i) WGCNA co-expression modules, (ii) label-permuted null, on recovery of the DNMT1-ZNF740-CTSS-STAT3-IFNG seed network. Full validation extends to 3 datasets, 2 foundation models (scGPT + Geneformer as cross-architecture check), and adds CSF-derived samples if available (GSE108000).
- GSE193770 (CD8+ T cell scRNA-seq, MS) — primary
- GSE138614 (bulk, used for replication cross-check of DEG direction, not coalition detection)
- GSE108000 (CSF-derived cells, secondary validation)
- CELLxGENE Census (immune reference atlas, for label transfer / annotation confidence)
- GTEx v10 (blood TPM baseline for CTSS liquid-biopsy plausibility check, not core to attention analysis)
- scVI atlas at gs://aegismind-tpu-results/ms_phase2/results/ (32,239 cells — use this figure, not the superseded 36,966)
- Pretrained foundation models: scGPT (human whole-body checkpoint), Geneformer (cross-check)
- STRING v12 network (50-gene MS seed set) for proximity scoring baseline
- Code repo: github.com/tradingjohn/ms-transcriptomics-carrim (extend, do not fork blind)
- Jaccard overlap of coalition membership across 3 seeds ≥0.6 (primary reproducibility gate).
- CA-RIM seed coalition (DNMT1/ZNF740/BRD3/CTSS/STAT3/IFNG) ranks in top decile (≥90th percentile) of all detected coalitions in ≥2 of 3 datasets.
- Permutation test vs. null (label-shuffled) significant at p<0.01, with ΔAUC ≥0.15 over null.
- Incremental value over WGCNA/Leiden co-expression baseline: ΔAUC ≥0.10 in recovering known seed-network genes.
- Cross-architecture consistency (scGPT vs. Geneformer): coalition composition overlap ≥0.5 Jaccard on the CA-RIM-relevant gene set.
- Jaccard overlap <0.4 across seeds → attention-derived coalitions deemed noise-dominated.
- CA-RIM coalition rank falls below median (50th percentile) in ≥2 datasets.
- No significant difference from label-permuted null (p>0.05).
- ΔAUC vs. WGCNA baseline <0.05 (i.e., foundation-model attention adds no interpretive value over classical co-expression).
- Cross-architecture overlap <0.2 (result is model-specific artifact, not a biological signal).
ROI Projection
Implementation Sketch
# Phase A: Data prep atlas = load_and_qc(["GSE193770","GSE108000"], exclude_mispooled_covid=True, n_cells_target=32239) labels = label_transfer(atlas, reference=CELLxGENE_Census, min_confidence=0.8) # Phase B: Foundation model fine-tune model = load_pretrained("scGPT-human-whole-body") model = fine_tune(model, atlas, task="CA-RIM_vs_control", epochs<=5) # Phase C: Agentic attention perturbation loop for seed in [1,2,3]: agent = AttentionPerturbationAgent(model, seed=seed) deviation_matrix = [] for cycle in range(200): heads_to_mask = agent.select_heads(strategy="uncertainty_guided") output_delta = model.forward_with_mask(atlas, heads_to_mask) gene_attention_delta = compute_gene_level_attention_shift(output_delta) deviation_matrix.append(gene_attention_delta) agent.update_policy(reward=stability_score(output_delta)) coalitions[seed] = cluster(deviation_matrix, method="leiden", metric="cosine", min_size=5) # Phase D: Controls wgcna_modules = run_wgcna(atlas.expression_matrix) null_coalitions = repeat_pipeline(atlas, labels=permute(labels)) # Phase E: Scoring & statistics score = lambda c: 0.3*mean_abs_delta(c) + 0.3*seed_reproducibility(c) + 0.2*string_proximity(c, seed_genes=MS_50GENE_SET) + 0.2*normalized_size(c) ranked = rank(coalitions, score) jaccard = pairwise_jaccard(coalitions[1], coalitions[2], coalitions[3]) p_value = permutation_test(ranked_CA_RIM_percentile, null_coalitions, n=1000) delta_auc_vs_wgcna = auc(coalitions_top, seed_gene_recovery) - auc(wgcna_modules, seed_gene_recovery) # Phase F: Cross-architecture replication model2 = load_pretrained("Geneformer") repeat(Phase B-E, model=model2, datasets=["GSE193770"])
- Day 10 (post-QC): if usable CD8+ T-cell count <10,000 after confidence filtering, abort/redesign before fine-tuning.
- Day 25 (post fine-tune): if fine-tuned model classification AUC for CA-RIM vs. control <0.65, abort — model has not learned a usable task-relevant representation for attention probing.
- Day 40 (post agentic loop, seed 1): if perturbation policy fails to converge (reward oscillating, no monotonic stability improvement) in >30% of cycles, abort and revise agent strategy before running seeds 2–3.
- Day 55 (post 3-seed run): if Jaccard overlap <0.4 across seeds, halt before running full 3-dataset + cross-architecture extension (do not spend the remaining ~70% of compute budget).
NAMED_EXPERTS: []
CLOSEST_EXISTING_WORK: []
NOVELTY_NARROWING_REQUIRED: false
SPINE_STATEMENT: This hypothesis is tested by whether agentic perturbation of single-cell foundation model attention reproducibly recovers the empirically-derived DNMT1-ZNF740-BRD3-CTSS-STAT3-IFNG CD8+ T-cell coalition as a top-ranked gene module, at a rate significantly exceeding label-permuted and classical co-expression baselines.