Biology's predictive models have learned to answer one direction of causation: given a gene perturbation, they can predict how the cell's transcriptome shifts. The clinically urgent question runs the other way: given the transcriptome of a cancer cell, what perturbation produced it, and what accessible intervention would revert it? This inversion is exactly where machine reasoning is known to break. Across mathematics, code execution, and chemistry tables, models that are highly accurate forward collapse when asked to infer causes from effects — GPT-4 drops from over 90% accuracy on forward math word problems to roughly 44% on the inverted "hiding" version of the same problems (Exploring Backward Reasoning in Large Language Models), and the same collapse appears when models must find code inputs that yield a given output (CRUXEval: A Benchmark for Code Reasoning, Understanding and Execution) or infer reaction conditions from an observed yield (Benchmarking Multimodal LLMs on Recognition and Understanding over Chemical Tables). Nobody has measured this inversion gap in regulatory biology, and nobody has asked whether the failure matters for the hardest targeting problem in oncology: transcription factors, the master switches of cell identity that lack the deep, rigid pockets a drug needs and are therefore called undruggable (A Small Molecule Selectively Targets N-Myc to Suppress Neuroblastoma Cancer Progression).
The proposal has two parts. First, build a benchmark that measures, for the first time, whether models that predict perturbation outcomes forward can abduct the perturbation backward from single-cell expression data. Second, build an abductive planner that sidesteps the reversal curse rather than fighting it: it never generates backward. It proposes candidate causes forward, verifies each candidate by running a forward simulator, and iterates — with a druggability mask that forbids proposing the undruggable transcription factor itself and forces the search toward its druggable flanking nodes: protein-interface partners, epigenetic co-factors, and chromatin-loop anchors.
Core Claim
The reversal curse is a generation problem, not a search problem. Ren et al. show that asking an autoregressive model to emit a backward plan directly conflicts with its left-to-right training, and that programmatically flipping the problem into a forward one improves planning success by 4% to 24% (Thinking Forward and Backward: Effective Backward Planning with Large Language Models). We claim the same flip works for cells: abduction over regulatory networks should be implemented as search over candidate interventions, each scored by a forward model, rather than as backward generation. Formally, given an observed differential expression profile $y$ and a forward simulator $f$ (a model mapping an intervention $I$ to a predicted profile), the planner searches
$$I^* \;=\; \arg\min_{I \,\in\, \mathcal{I}_{\text{druggable}}}\; d\big(f(I),\, y\big),$$
where $d$ is a distance that weights high-magnitude differentially expressed genes (following the Autofocus design in GEARS) and $\mathcal{I}_{\text{druggable}}$ is the intervention space with undruggable transcription factors masked out. The search is guided — a language model proposes candidates from the signature and a knowledge graph, in the spirit of ARGOS, which couples a SAT solver's entailed-literal "backbone" with LLM rule proposal and whose symbolically solved proofs were judged faithful in 100% of human review (A Balanced Neuro-Symbolic Approach for Commonsense Abductive Logic) — but every proposal must survive forward verification, so nothing enters the answer without a checkable causal trace.
Research Program
Part one is the benchmark, which we call transcriptomic abduction. It is constructed from two Perturb-seq datasets already standard in this literature: the Norman et al. CRISPRa screen in K562 cells, including its 131 two-gene combinations, and the Replogle et al. genome-scale CRISPRi screen of over 340,000 cells across K562 and RPE-1 lines, both used by GEARS (Predicting transcriptional outcomes of novel multigene perturbations with GEARS) and by the PertEval benchmark (PertEval-scFM: Benchmarking Single-Cell Foundation Models for Perturbation Effect Prediction). The task instances give the model only the post-perturbation differential expression profile and ask it to recover the perturbed gene or gene pair, scored by top-1, top-5, and top-20 recall, plus recovery of the genetic-interaction subtype for pairs. We add two hard splits that prior work says matter: a zero-shot split where the correct combination was never seen in training, and graph-sparsified distribution-shift splits in the style of PertEval's SPECTRA curves. A second tier of the benchmark tests flanking: given a signature known to be driven by an undruggable transcription factor, the model must retrieve an accessible indirect target rather than the factor itself. Ground truth comes from the curated biology in the file: the TP63-anchored chromatin loop roughly 340 kb upstream of MYC that raises MYC expression about five-fold in radiation-transformed bronchial cells (TP63 mediates the generation of tumour-specific chromatin loops that underlie MYC activation in radiation-induced tumorigenesis), the FOXM1–JUP axis in pancreatic cancer, where JUP overexpression shifts gemcitabine IC50 from about 8.5 μM to 15.7 μM (Transcriptional deregulation by FOXM1–JUP signaling confers dual oncogenic drivers for pancreatic tumorigenesis and therapeutic resistance), the PAX8–YAP1/TEAD interface in ovarian cancer precursors (Epigenetic remodeling regulates transcriptional changes between ovarian cancer and benign precursors), and the HDAC1/2/3 co-dependency discovered by screening 63,000 compounds (Chemical genomics reveals histone deacetylases are required for core regulatory transcription).
Part two is the planner. It combines three components. A language model reads the differential expression signature together with neighborhood context from the same co-expression and Gene Ontology graphs GEARS uses, and proposes a shortlist of candidate interventions with natural-language mechanistic rationales — always stated forward ("perturbing X would repress these genes"), never backward. A solver-style filter, analogous to ARGOS's backbone computation, prunes candidates that are entailed or contradicted by already-verified facts in the search trace. Each surviving candidate is verified by running GEARS forward and comparing its predicted profile to the observed one under $d$; the loop continues, beam-style, until the match is within tolerance or the budget is spent. The druggability mask is a curated annotation layer marking transcription factors with flat, dynamic, or disordered surfaces — SOX2, N-Myc, TP63, FOXM1, PAX8, HoxB13 — as non-proposable, so the planner can only return accessible flanking nodes: PPI partners, epigenetic enzymes, RNA modifiers such as the ALKBH5 axis upstream of YY1, or degradation-delivery targets of the SCORT–Cas13d type (SCORT–Cas13d Nanotherapy Precisely Targets the 'Undruggable' Transcription Factor HoxB13 in Metastatic Prostate Cancer In Vivo). The point of the mask is not to encode new biology but to convert the undruggability constraint, which today is handled case-by-case by human reasoning, into a search prior a machine can honor.
A specific analysis will test whether the planner can recover the counterintuitive dependency that linear pathway intuition misses: the chemical-genomics finding that HDACs, classically repressors, are required for active core-regulatory transcription, with inhibition causing hyper-accessibility yet collapsing the loops that super-enhancers need (Chemical genomics reveals histone deacetylases are required for core regulatory transcription). A planner that retrieves HDAC1/2/3 from a PAX3-FOXO1-driven signature demonstrates that forward-verified search captures mechanisms that graph connectivity cannot.
Proposed Experiments
The main experiment runs all baselines and the planner on the transcriptomic abduction benchmark under identical data splits, compute, and tuning budgets, reporting recall and interaction-subtype precision on the in-distribution, zero-shot-combination, and SPECTRA-shifted splits. On the flanking tier, we report whether each method's top retrieval lies in the curated accessible set for each of the four ground-truth case studies, and whether the planner's mechanistic trace is faithful, using the human-review protocol from the ARGOS paper as the template. Because the planner emits an explicit chain of proposals and forward checks, every answer is auditable; for the strongest baseline we state the numbers a cited paper reports so the bar is concrete before we run anything.
Baselines
The naive reference point is the Mean Baseline from the PertEval benchmark, which predicts the average perturbation response and which, remarkably, the best single-cell foundation models barely beat (PertEval-scFM: Benchmarking Single-Cell Foundation Models for Perturbation Effect Prediction); we also include direct LLM prompting, where a general model is handed the signature and asked to name the perturbed gene in one shot. No reported result exists on Lacuna for direct LLM perturbation recovery from expression vectors, so we measure it ourselves; the SVAMP collapse from over 90% forward to roughly 44% on the inverted hiding task (Exploring Backward Reasoning in Large Language Models) motivates expecting failure, but expression data is a different modality and we treat the number as unknown. The strongest competitor is an exhaustive cycle-consistency scan built on GEARS itself: enumerate every candidate perturbation, run GEARS forward on each, and rank by the same distance $d$. GEARS is the right anchor because it is the dominant forward model — on the Norman dataset under SPECTRA distribution shifts it achieved an AUSPC of roughly 0.008 while the best foundation model, scGPT, trailed at roughly 0.045, barely beating the Mean Baseline (PertEval-scFM: Benchmarking Single-Cell Foundation Models for Perturbation Effect Prediction); it also reported 40% higher precision than existing approaches on genetic-interaction subtypes and a 53% improvement on zero-shot two-gene combinations (Predicting transcriptional outcomes of novel multigene perturbations with GEARS). The comparison is fair because both methods use the same GEARS checkpoint and the same distance, and we cap the planner to no more forward-simulation calls than the exhaustive scan uses; beating it means higher top-k recall and interaction-subtype precision at equal or lower simulation count, with the decisive margin expected on pair recovery, where the combinatorial space is large, and on the shifted and flanking tiers. Two ablations isolate the ingredients: removing the druggability mask (the planner should then retrieve TP63 or FOXM1 itself, demonstrating the mask's causal role) and removing the LLM proposer (replacing proposals with random beam candidates, isolating whether guided proposal matters once forward verification is in place).
Why not just…?
Why not just fine-tune a single-cell foundation model in reverse? This is the most serious alternative, and we run it rather than argue it away: an MLP or linear probe on scGPT or Geneformer embeddings, trained to classify the perturbed gene from the expression vector, on the same splits. The evidence against it is direct — PertEval probed five such models and found them an order of magnitude behind GEARS on Norman, barely beating the Mean Baseline, with rapid decay under SPECTRA shifts (PertEval-scFM: Benchmarking Single-Cell Foundation Models for Perturbation Effect Prediction) — and the broader reversal literature shows autoregressive backbones pay a token-probability penalty when forced to infer causes from outputs (Thinking Forward and Backward: Effective Backward Planning with Large Language Models). We are honest that a probe may do fine on in-distribution single-gene recovery; our claim is scoped to the zero-shot-combination, shifted, and flanking splits, where the probe's own benchmark says it collapses. If the probe matches the planner there too, the reversal curse does not transfer to this modality — a finding worth reporting on its own.
Why not just use master regulator analysis or graph traversal? Classical network biology navigates from downstream gene lists to upstream regulators via centrality or shortest paths over interaction graphs. The HDAC result is the counterexample: repressors discovered to be required for active transcription, with inhibition producing hyper-accessibility that nevertheless collapses core-regulatory loops — a mechanism invisible to connectivity reasoning (Chemical genomics reveals histone deacetylases are required for core regulatory transcription). GEARS likewise beat network-style baselines by 40% precision on non-additive interaction subtypes precisely because interactions are not paths (Predicting transcriptional outcomes of novel multigene perturbations with GEARS). We include shortest-path and centrality baselines on the same co-expression and GO graphs; they may well be strong on simple single-TF recovery, and if they win there we report it — but they have no semantics for chromatin-loop anchors or druggability, which is where the flanking tier lives.
Why not just use BiAAE end-to-end? The bidirectional adversarial autoencoder already maps target transcriptomic profiles directly to SMILES structures via a disentangled latent space (Molecular Generation for Desired Transcriptome Changes With Adversarial Autoencoders). But it is a correlational, bulk-level mapper trained on LINCS L1000 with no mechanism to condition on whether the target has a binding pocket, no single-cell resolution, and no verifiable intermediate steps; given an undruggable-TF signature it can only emit compounds that correlate with the signature, not compounds that act through a specified flanking mechanism. Our object is intervention selection with an auditable causal trace, not molecule generation, and the druggability mask is exactly the ingredient BiAAE lacks. If a reviewer wants molecules, our retrieved flanking node plus its known chemistry (for example, existing HDAC probes from the 63,000-compound screen) is the output to compare against BiAAE — but that comparison is outside this proposal's scope, and we say so.
Decisive Checks
Three outcomes would settle the proposal's fate, and we commit to reporting whichever occurs. If the exhaustive GEARS scan matches the planner on every tier, the benchmark stands as the contribution and the LLM proposer is unnecessary — the search problem is easy enough to brute-force at this scale, and we would say so. If the reverse-fine-tuned probe matches the planner on the shifted and combinatorial splits, the reversal curse is falsified for transcriptomic inversion and the field learns that the asymmetry is modality-specific. If the planner with the mask retrieves accessible flanking nodes — HDACs for the super-enhancer signature, JUP or the FOXM1–JUP interface for the PDAC signature, loop components for the MYC-high radiation signature — where the unmasked planner retrieves the undruggable factor itself, then the core claim holds: undruggability becomes a search constraint, and abduction over regulatory networks is best done by machines that only ever reason forward.
References
- Exploring Backward Reasoning in Large Language Models
- CRUXEval: A Benchmark for Code Reasoning, Understanding and Execution
- Benchmarking Multimodal LLMs on Recognition and Understanding over Chemical Tables
- A Small Molecule Selectively Targets N-Myc to Suppress Neuroblastoma Cancer Progression
- Thinking Forward and Backward: Effective Backward Planning with Large Language Models
- A Balanced Neuro-Symbolic Approach for Commonsense Abductive Logic
- Predicting transcriptional outcomes of novel multigene perturbations with GEARS
- PertEval-scFM: Benchmarking Single-Cell Foundation Models for Perturbation Effect Prediction
- TP63 mediates the generation of tumour-specific chromatin loops that underlie MYC activation in radiation-induced tumorigenesis
- Transcriptional deregulation by FOXM1–JUP signaling confers dual oncogenic drivers for pancreatic tumorigenesis and therapeutic resistance
- Epigenetic remodeling regulates transcriptional changes between ovarian cancer and benign precursors
- Chemical genomics reveals histone deacetylases are required for core regulatory transcription
- SCORT–Cas13d Nanotherapy Precisely Targets the ‘Undruggable’ Transcription Factor HoxB13 in Metastatic Prostate Cancer In Vivo
- Molecular Generation for Desired Transcriptome Changes With Adversarial Autoencoders