Applying the neuro-symbolic ODE discovery framework (Latent Grammar Flow) to single-cell perturbation time-series from HarmonyCell-harmonised datasets will recover interpretable differential-equation motifs that capture causal regulatory logic missed by attention-based foundation models, quantifiable by comparing LGF-discovered symbolic ODEs against SAE causal circuit maps for predictive accuracy on held-out CRISPR perturbation outcomes.

Computer ScienceAug 17, 2026Evaluation Score: 68%

Adversarial Debate Score

63% survival rate under critique

Expert panel critique

Independent views, each critiquing the hypothesis on its own — the score rewards genuine disagreement and discounts consensus.

Mistral: The hypothesis is well-grounded in recent neuro-symbolic and single-cell interpretability literature, with clear falsifiability and a strong comparative framework (LGF vs. SAE/attention models). However, it lacks direct empirical validation from the owner’s experiments, which focus on precision/o...

ChatGPT: The hypothesis is falsifiable and proposes a concrete held-out CRISPR benchmark, while the cited literature supports the motivation that attention and SAE features may lack regulatory logic. However, no validated owner experiments directly support it, and HarmonyCell harmonisation may distort dyn...

Claude: The hypothesis is well-structured and falsifiable — LGF-discovered ODEs versus SAE causal circuits on held-out CRISPR outcomes is a concrete benchmark — and draws coherently on real published methods (LGF, HarmonyCell, SAE circuit tracing) that collectively support the proposed pipeline; however,...

Supporting Research Papers

Neuro-Symbolic ODE Discovery with Latent Grammar Flow
Understanding natural and engineered systems often relies on symbolic formulations, such as differential equations, which provide interpretability and transferability beyond black-box models. We int...
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...
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...
Symbolic Discovery of Stochastic Differential Equations with Genetic Programming
Automated scientific discovery aims to improve scientific understanding through machine learning. A central approach in this field is symbolic regression, which uses genetic programming or sparse regr...
HarmonyCell: Automating Single-Cell Perturbation Modeling under Semantic and Distribution Shifts
Single-cell perturbation studies face dual heterogeneity bottlenecks: (i) semantic heterogeneity--identical biological concepts encoded under incompatible metadata schemas across datasets; and (ii) st...

Computational Result

📖 Literature-assessed (LLM)· literature_meta

An LLM's reading of the literature — not computational verification.

Potential for LGF to uncover regulatory motifs remains to be validated.

Method: literature_meta · Result: inconclusive · Confidence: 60%

Formal Verification

Z3 logical consistency:✅ Consistent

Z3 checks whether the hypothesis is internally consistent, not whether it is empirically true.

Experimental Validation Package

This discovery has a Claude-generated validation package with a full experimental design.

Precise Hypothesis

On HarmonyCell-harmonised single-cell CRISPR perturbation time-series (e.g., Perturb-seq/CRISPRi-screen datasets with ≥3 timepoints post-perturbation), a Latent Grammar Flow (LGF) neuro-symbolic ODE discovery pipeline will recover a set of symbolic differential-equation motifs (≤10 terms each, interpretable as gene-gene or gene-pathway interaction rules) that, when used to predict held-out CRISPR perturbation outcomes (log-fold-change in target/downstream gene expression at 24-72h), achieve equal or higher predictive accuracy (Pearson r or R² on held-out perturbations) than causal circuit maps extracted via Sparse Autoencoders (SAEs) from an attention-based single-cell foundation model (e.g., scGPT/Geneformer-class), specifically on the subset of perturbations involving indirect/multi-hop regulatory effects (≥2 edges from perturbed gene to measured readout).

Disproof criteria:

LGF symbolic ODEs achieve statistically indistinguishable or lower predictive accuracy than SAE circuit maps on held-out CRISPR perturbations (difference in R²/Pearson r not significant at p<0.05, paired bootstrap, n≥30 held-out perturbations).
Recovered symbolic motifs fail to match known regulatory relationships above chance rate when benchmarked against curated ground-truth networks (e.g., TRRUST, DoRothEA) — precision/recall not significantly above random edge permutation baseline.
LGF discovery fails to converge or produces degenerate/overfit expressions (>50 terms, unstable across bootstrap resamples of cells) on ≥2 of 3 benchmark datasets.
Performance advantage, if present, is fully explained by dataset-specific artifacts (e.g., disappears when repeated on a second independent HarmonyCell-harmonised dataset).

Spine & Adversarial ReadReady for validation

“This hypothesis tests whether symbolic ODE motifs discovered by a neuro-symbolic framework (LGF) from single-cell perturbation time-series predict held-out CRISPR perturbation outcomes more accurately than causal circuit maps extracted from attention-based single-cell foundation models via sparse autoencoders.”

highThe comparison is not apples-to-apples: SAE causal circuit extraction from single-cell foundation models is a nascent, unstandardized technique, while LGF is purpose-built for this exact task — any observed LGF advantage may reflect baseline immaturity rather than a genuine methodological insight about symbolic vs. attention-based causal discovery.
Partially addressed via the ablation in methodology step 11 (SAE without causal tracing, raw attention baseline) and by requiring replication on a second dataset, but the EVP does not fully resolve this — a stronger design would include multiple SAE extraction methods (e.g., 2-3 published causal-tracing variants) to show the result is robust to baseline choice, which is not currently specified.
mediumWhy HarmonyCell specifically, and why LGF specifically, rather than established alternatives (e.g., Harmony/scVI for integration, SINDy or PySR for symbolic ODE discovery)? The methodology does not justify these particular tool choices over well-validated alternatives, which is a common rejection point.
Not resolved in this EVP — the protocol assumes HarmonyCell and LGF as given (per the discovery's framing) without a comparative justification against SINDy, PySR, or standard Harmony/scVI. A revision should add a baseline arm using established symbolic regression (PySR/SINDy) and standard batch correction to isolate whether gains come from the specific tools or the general approach.
mediumGround-truth regulatory networks (DoRothEA/TRRUST) are themselves derived largely from bulk, steady-state, or literature-curated data with known incompleteness and context-dependence; using them as the evaluation gold standard for single-cell, perturbation-specific, dynamic regulatory logic may systematically penalize methods that discover context-specific true edges absent from these databases.
Acknowledged as a boundary condition and failure mode but not solved — mitigated only partially by also scoring against held-out perturbation outcomes (an orthogonal, functional ground truth) rather than relying solely on curated network overlap.

Experimental Protocol

Minimum viable test: single well-powered Perturb-seq time-series dataset (e.g., a published CRISPRi time-course with ≥50 perturbations, ≥3 timepoints, ≥5000 cells/condition) run through both pipelines (LGF and SAE-on-foundation-model), with a fixed 80/20 perturbation-level train/test split, repeated across 5 random splits, comparing predictive accuracy and edge-recovery precision/recall against a curated ground-truth regulatory network subset.

Required datasets:

A published time-resolved Perturb-seq/CRISPRi dataset (e.g., Replogle et al. 2022-style genome-scale Perturb-seq, or a smaller time-course dataset such as Papalexi et al. ECCITE-seq if temporal resolution available) — target: ≥50 perturbations × ≥3 timepoints.
HarmonyCell harmonisation pipeline (or equivalent Harmony/scVI-based batch correction) applied to integrate the dataset with at least one independent single-cell atlas for cross-batch validation.
Pre-trained attention-based single-cell foundation model (scGPT or Geneformer) with SAE causal circuit extraction pipeline (e.g., using existing sparse-autoencoder-for-interpretability toolkits adapted to single-cell embeddings).
Curated ground-truth regulatory network (DoRothEA, TRRUST, or a curated ChIP-seq-derived network) restricted to genes present in the perturbation panel.
Held-out CRISPR perturbation outcomes (log-fold-change matrices) not used in either model's training/fitting.

Success:

LGF achieves ≥10% relative improvement in R² (or equivalent Pearson r improvement of ≥0.05) over SAE circuit maps on indirect (≥2-edge) held-out perturbations, significant at p<0.05 across 5 splits.
LGF-recovered ODE motifs achieve edge-recovery F1 ≥0.4 against ground-truth network (vs. SAE circuit F1), representing a ≥15% relative F1 improvement.
Results replicate directionally (same sign of advantage) on a second independent dataset.
LGF motifs remain stable across bootstrap resamples (≥70% term overlap across 5 splits), demonstrating genuine discovery rather than overfitting.

Failure:

No significant difference (or SAE outperforms) in predictive accuracy on indirect-effect held-out perturbations across ≥3 of 5 splits.
LGF motifs show <20% overlap with ground-truth regulatory edges (indistinguishable from random baseline).
LGF fails to converge or produces unstable/non-reproducible expressions across resamples on the primary dataset.
Advantage, if observed, vanishes or reverses on the second independent dataset.

ROI Projection

Commercial:

Directly applicable to pharma target discovery pipelines (mechanism-of-action elucidation, combination therapy prediction via ODE simulation of multi-gene perturbations), precision medicine biomarker discovery, and as a licensable interpretability toolkit for any lab running Perturb-seq/CRISPR screens. Addressable market includes computational biology teams at biotech/pharma (target discovery, ~$500M+ annual spend on perturbation screening platforms industry-wide) and academic core facilities adopting single-cell perturbation methods.

TIME_TO_RESULT_DAYS: 120

Implementation Sketch

# Stage 1: Data prep
data = load_perturbseq_timeseries(dataset_id)
data_harmonized = HarmonyCell.integrate(data, reference_atlas)
validate_effect_size_retention(data_harmonized, positive_controls, threshold=0.8)
trajectories = estimate_pseudotime_trajectories(data_harmonized)
derivatives = compute_derivatives(trajectories, method="GP_smoothing")

# Stage 2: LGF pipeline
for split in range(5):
    train, test = split_by_perturbation(data_harmonized, ratio=0.8, seed=split)
    latent_grammar = LGF.encode(train.trajectories)
    symbolic_odes = LGF.search(
        latent_grammar, max_terms=10, grammar_prior="biological_motifs"
    )
    fitted_odes = LGF.fit_parameters(symbolic_odes, train.derivatives)
    lgf_preds = simulate_forward(fitted_odes, test.initial_conditions, test.timepoints)

    # Stage 3: SAE baseline
    fm = load_pretrained("scGPT")
    fm_finetuned = finetune(fm, train)
    activations = extract_activations(fm_finetuned, test.perturbations)
    sae = SparseAutoencoder.fit(activations, sparsity=0.05)
    causal_circuits = trace_causal_paths(sae, method="activation_patching")
    sae_preds = predict_via_circuit(causal_circuits, test.perturbations)

    # Stage 4: Scoring
    score_predictive(lgf_preds, sae_preds, test.ground_truth, stratify_by="edge_distance")
    score_edge_recovery(symbolic_odes, causal_circuits, ground_truth_network)

aggregate_and_test_significance(all_splits, method="wilcoxon_signed_rank")

Abort checkpoints:

After Step 2 (harmonisation validation): if effect-size retention <70%, abort and re-select dataset/harmonisation parameters before proceeding.
After Step 4 initial LGF run: if symbolic search fails to converge to stable low-term expressions on training data within compute budget (48 GPU-hours per split), abort and revisit grammar priors/search constraints.
After Step 6 (first 2 splits scored): if LGF shows no directional advantage even on indirect-effect subset, halt full 5-split run and perform diagnostic ablation before committing to remaining compute.
After second dataset replication attempt: if primary-dataset advantage fails to replicate directionally, abort further scale-up and report as inconclusive/negative.

NAMED_EXPERTS: []

CLOSEST_EXISTING_WORK: []

NOVELTY_NARROWING_REQUIRED: false

Source

AegisMind Research

Need AI to work rigorously on your problems? AegisMind uses the same multi-model engine for personal and professional use. Get started