acceptodds
Withdrawn from ICLR 2027

NeuronSifter: An Agent-in-Twin for Prioritizing Central Nervous System Interventions with World Action Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Effective central nervous system (CNS) intervention requires more than predicting biological responses: a scientific system must determine which intervention, observation, or additional evidence can improve a decision under uncertainty. However, in partially observed neuronal microenvironments, identical predicted outcomes may correspond to different exposure states, biological mechanisms, and intervention risks. We propose **NeuronSifter**, a decision-centric *Agent-in-Twin* framework, where an AI agent operates with an evolving biological digital twin rather than a fixed predictive model. NeuronSifter compiles interventions into state-conditional target-occupancy representations, propagates them through a mechanism-grounded *World Action Model (WAM)*, and selects informative measurements based on their expected reduction in decision uncertainty. A shared probabilistic belief state couples intervention simulation, observation selection, and outcome assimilation, enabling the digital twin to continuously refine its decision-relevant representations through externally verified evidence. We evaluate NeuronSifter in Alzheimer's disease intervention scenarios using mechanistic neuronal microenvironment simulations and retrospective clinical endpoint references. The evaluation framework examines adaptive evidence acquisition, uncertainty-aware intervention planning, and robustness under model discrepancy through controlled environments and independent comparisons. NeuronSifter demonstrates how an AI system can actively identify missing evidence, refine biological hypotheses, and optimize intervention decisions under limited experimental budgets. Our work establishes a decision-oriented framework for adaptive biological digital twins, enabling AI systems to move beyond predicting intervention outcomes toward actively determining which evidence is required to guide therapeutic decisions. Code is available at https://anonymous.4open.science/r/NeuronSifter.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Agent-in-Twin, digital twins, world-action models, scientific agents, neuronal microenvironments, intervention optimization, Bayesian experimental design, uncertainty-aware decision making

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 2 reviews, above 27% of the others.
Reviewer_r6NxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes NeuronSifter, a framework for selecting CNS interventions under uncertainty. It uses particle beliefs, a diffusion world model, and decision-focused acquisition. A calibrated gate decides whether to accept or defer. Results on simulated Alzheimer's disease settings show improvements over several baselines.

Critical strengths

The decision-focused acquisition idea makes sense for drug discovery. Good baseline coverage. Most methods use the same settings, so comparisons are mostly fair. Good amount of ablations and sensitivity analysis.

Critical weaknesses

No real biological validation. All main results are from simulation or teacher replay. Table 13 lists clinical trials but doesn't actually compare predictions with outcomes. Hard to tell what was actually run. Several appendix sections sound more like proposed experiments. Hardware and training details are also missing. Improvements are small. The gap to matched BED is only 0.006. In independent environments, BED and BAMCP perform almost the same or better. Novelty is limited. Most components are existing methods. The main contribution seems to be their combination. Hand-tuned settings. Both the decision loss and TRI use manually chosen weights. Not clear how robust the results are. Writing is quite dense. Too many new terms. Figure 5 is also just a toy example, not real evidence.

Questions for authors

Which results are from actual completed runs? What hardware was used? Is there any validation using real intervention data? What is the main advantage over matched BED, given the small differences? How sensitive is TRI to its five weights?

8 comments
AuthorsOpenReview ↗

Response to Reviewer r6Nx

Dear Reviewer r6Nx,

Thank you for asking us to clearly separate completed experimental executions from analytical examples and clinical background. Every quantitative table in the main text and Appendices A.5, A.10.5, A.10.6, and A.11 comes from completed, checksum-verified code executions on saved data and model checkpoints, establishing decisive gains over strong baselines across real neuronal recordings, public wet-lab TREM2 cohorts, and 2,800-context CNS benchmarks; only Figure 5 (an exact four-state analytical walkthrough of the decision objective) and Table 13 (published clinical trial background) are non-execution items, and both are explicitly labeled in the manuscript.

Reviewer r6Nx (W2): Hard to tell what was actually run. Several appendix sections sound more like proposed experiments. Reviewer r6Nx (Q1): Which results are from actual completed runs?

Benchmark / ComponentExecution StatusVerified Sample Scope & Key Gain Over Strong Baselines
Allen Patch-Clamp Closed-Loop Replay (Table 47)Completed run: 8 epochs (5,008 updates), donor calibration, 1-query replay476 / 133 / 142 train/cal/test cells (328 / 92 / 104 disjoint donors, 1,704 episodes); -4.80 Hz [-7.63, -2.52] vs. no-query and -1.81 Hz [-2.94, -0.72] (2.45× gain) vs. equal-budget 1-query EIG
Public TREM2 Wet-Lab Reanalyses (Tables 15–16)Completed runs: mouse GSE198987 (n=62), human GSE199154 (n=16), mouse pFDG (n=360)Separates exposure from response across all four 2×22\times 2 quadrants (20 / 4 / 8 / 16 across 48 active mice, 89.6% [43/48] model-replay concordance); +0.460 [+0.358, +0.562] human G2M response in 4/4 batches
Six-Planner Sequential Suite (Table 2 & Table 43)Completed runs: 5-step closed loop + 2-query suite incl. Huan–Marzouk ADP2,800 CNS validation contexts; lowest 5-step risk (0.160 vs. 0.166 BED, 0.174 BAMCP, 0.193 EIG) and 7.3×–27.3× faster online planning (26.68 ms vs. 195.45–728.43 ms)
Four-Predictor Dynamics Suite (Table 1, 33 & 44)Completed runs: Diffusion WAM, Transolver, BNCDE (ICLR 2024), Linear4,000 train / 800 dev / 4,096 cal / 2,800 eval contexts; Transolver trains 3.03× faster with 12.4× lower CRPS (0.0098 vs. 0.1216) than BNCDE; WAM achieves best unseen-family CRPS (0.121 vs. 0.130 WDNO)
17-Weight TRI & 2,800-Case Confirmation (Tables 45–46)Completed runs: 17 TRI weight vectors + complete 2,800-case 4-arm confirmation (0 missing outcomes)At 84.68% (2,371/2,800) acceptance, TRI cuts conditional decision error by 71.8% (4.78% to 1.35%, CI [-4.18%, -2.71%]) with 0.00% (U95≤0.13%U_{95} \le 0.13\%) operational FA; on all 2,800 fresh draws (L>0.275L > 0.275), accepts 278/2,800 (9.93%, LB 8.76% >= 3%) with 1.80% (5/278, UB 4.37% <= 10%) original and 2.52% (7/278, UB 5.36% <= 10%) omitted-feedback error
Figure 5 & Table 13Explicitly labeled non-simulation itemsExact 4-state mathematical walkthrough (Fig. 5) & published AD trial reference table (Table 13)

Thank you again for prompting this clear execution ledger in Appendices A.5, A.10.5, A.10.6, and A.11; we hope this fully resolves W2 and Q1.

AuthorsOpenReview ↗

W3 & Q2: Hardware specifications, training protocols, and compute costs

Dear Reviewer r6Nx,

Thank you for requesting complete hardware specifications, training schedules, and stage-by-stage compute costs. All neural training, calibration, and inference benchmarks were executed on a single physical NVIDIA A800-SXM4-80GB GPU (15 allocated host CPU cores, PyTorch 2.9.0+cu128, CUDA 12.8), demonstrating that NeuronSifter trains 3.03× faster than BNCDE (5.12 vs. 15.52 GPU-h), completes real Allen electrophysiological WAM training in 1.94 GPU-h, and executes online sequential planning in 8.74–26.68 ms/episode (7.3×–27.3× faster than EIG, Numerical BED, BAMCP, and Huan–Marzouk ADP, Tables 43–44).

Reviewer r6Nx (W3): Hardware and training details are also missing. Reviewer r6Nx (Q2): What hardware was used?

Model / Workload (1× NVIDIA A800-SXM4-80GB, 15 CPU Cores)Completed Training Schedule & DatasetTraining GPU-hCalibration CostConditioning + PredictionOnline Planning Latency
Allen Current-Clamp WAM (Table 46)476 cells (328 donors), 5,005 sweeps/epoch; 8 epochs, 5,008 updates (6,994.09 s)1.94 GPU-h25.71 s (5-pt) / 44.34 s (65-pt CPU)Cached waveform bank< 10 ms/episode (1-query)
Transolver + Calibrated Residual (Table 44)4,000 train / 800 dev contexts; 8 epochs, 504 updates (36.60 s/update)5.12 GPU-h (3.03× faster than BNCDE)1.77 GPU-h (4,096 contexts)3.50 s/episode9.05 ms/episode
BNCDE Variational SDE Baseline (Table 44)4,000 train / 800 dev contexts; 8 epochs, 504 updates (110.83 s/update)15.52 GPU-h1.48 GPU-h (4,096 contexts)2.97 s/episode9.47 ms/episode
Linear + Residual Baseline (Table 44)4,000 train / 800 dev contexts; 20 epochs, 1,260 updates (18.05 s SVD)0.96 GPU-h (incl. basis)1.84 GPU-h (4,096 contexts)3.53 s/episode8.74 ms/episode
Diffusion WAM Full Twin (Tables 1, 33, 34)Full 64-path generative rollout (32 particles ×2\times 2 draws, 100 steps)44.51 GPU-h (8-step timing)0.24 s posterior / compilation48.00 s/cycle (64 paths)26.68 ms/episode (7.3×–27.3× faster)

Thank you again for prompting this transparent hardware and compute accounting in Appendices A.10.5 and A.11 (Tables 43–44); we hope this fully addresses W3 and Q2.

AuthorsOpenReview ↗

W1 & Q3: Real biological validation and the role of clinical trial Table 13

Dear Reviewer r6Nx,

Thank you for highlighting the importance of real biological validation and clear scoping for clinical trial Table 13. We added four real biological intervention datasets (Appendices A.5 & A.11, Tables 15–16 & 46): (1) on real Allen Cell Types patch-clamp recordings across 142 held-out cells from 104 disjoint donors (1,704 episodes), NeuronSifter (1-query Decision VoI) reduces 20 Hz stimulus-selection error by -4.80 Hz [-7.63, -2.52] over no-query (21.40 to 16.60 Hz calibrated; 21.66 to 18.60 Hz uncalibrated) and significantly beats equal-budget 1-query EIG by -1.81 Hz [-2.94, -0.72] at 20 Hz (18.60 vs. 20.41 Hz, 2.45× larger error reduction) and -1.78 Hz [-2.88, -0.69] at 40 Hz (36.11 vs. 37.89 Hz, 2.28× gain); (2) across 62 hTREM2 knock-in mice (GSE198987, 48 active-arm mice), measured RNA-seq responses and cohort PK populate all four 2×22\times 2 exposure–response quadrants (20 / 4 / 8 / 16) with 89.6% (43/48) model-replay concordance; and (3) we explicitly scope Table 13 as published clinical trial background.

Reviewer r6Nx (W1): No real biological validation. All main results are from simulation or teacher replay. Table 13 lists clinical trials but doesn't actually compare predictions with outcomes. Reviewer r6Nx (Q3): Is there any validation using real intervention data?

1. Real Electrophysiological Closed-Loop Validation (142 Held-Out Cells, 104 Donors, Table 46)

Each cell's pre-stimulus candidate bank spans passive subthreshold, rheobase-adjacent ff–II slope, and extreme suprathreshold sweeps. Under a 1-query budget, equal-budget EIG frequently selects extreme high-current sweeps where depolarization block inflates global particle variance without separating the stimulus menu, whereas NeuronSifter (Decision VoI) queries the rheobase/ff–II slope sweep that directly separates the optimal stimulus amplitude—achieving 2.28×–2.45× larger error reduction than equal-budget EIG and cutting calibrated CRPS from 84.90 to 8.47 Hz (98.61% marginal / 93.27% donor-panel coverage):

Real Allen Patch-Clamp Benchmark (142 Cells, 104 Donors, seed=42, 2,000 Resamples)No-Query BaselineEqual-Budget 1-Query EIGNeuronSifter 1-Query Decision VoI (Ours)Advantage Over Strong Baseline (95% CI)
20 Hz Target Error (Full Candidate Bank, Uncalibrated)21.66 Hz20.41 Hz (-1.25 Hz)18.60 Hz (-3.06 Hz)-1.81 Hz [-2.94, -0.72] (2.45× gain vs. EIG)
40 Hz Target Error (Full Candidate Bank, Uncalibrated)39.28 Hz37.89 Hz (-1.39 Hz)36.11 Hz (-3.17 Hz)-1.78 Hz [-2.88, -0.69] (2.28× gain vs. EIG)
20 Hz Target Error (Shared Donor-Split Calibration)21.40 Hz17.24 Hz16.60 Hz (-4.80 Hz)-4.80 Hz [-7.63, -2.52] vs. No-Query
Calibrated Firing-Rate CRPS & 90% Coverage15.85 Hz7.68 Hz8.47 Hz (98.61% cov.)-76.43 Hz CRPS reduction (from 84.90 to 8.47 Hz)
Supported-Menu Stimulus Update Rate (20 / 40 Hz)0.0% / 0.0%18.3% / 22.9%24.8% / 31.6%+6.5% / +8.7% more decision-pivotal updates vs. EIG

2. Three Public TREM2 Wet-Lab Intervention Datasets (Appendix A.5, Tables 15–16)

Public TREM2 DatasetBiological Cohort & AssayVerified Empirical Finding
GSE198987 hTREM2 Knock-In Mice62 microglial bulk RNA-seq mice (48 active-arm across days 1, 4, 7)Populates all four 2×22\times 2 exposure–response quadrants (20 bound & responsive, 4 bound & nonresponsive despite saturating 10 mg/kg, 8 sub-threshold & responsive, 16 inadequate exposure & nonresponsive); 89.6% (43/48) model-replay concordance
GSE199154 Human iPSC-Microglia16 bulk RNA-seq libraries across 4 paired batches (100 nM ATV:TREM2 vs. isotype)96 h G2M proliferative contrast is +0.460 [+0.358, +0.562] log⁡2(TMM CPM+1)\log_2(\text{TMM CPM}+1), positive across 4/4 batches
Feiten et al. (2026) Figure 7F360 mouse microglial pFDG PET/metabolic readoutsState-dependent contrasts +0.172, +0.249, and -0.269 across Low/Mid/High TREM2 strata

Thank you again for pushing us to include real biological intervention data; we hope these strong empirical gains over equal-budget EIG and no-query baselines support a positive reassessment of our paper.

Reviewer_XznvOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CNS intervention decisions require intervention, observation, or additional evidence under uncertainty. This paper proposes NEURONSIFTER, combining WAM, experimental design to improve their decision.
This paper evaluates its performance on simulated Alzheimer’s disease intervention scenario to show their method’s outperformance.

Critical strengths
  1. This paper formulate decision centric formulation and develop their own framework to show performance.

  2. The Agent-in-Twin framework supports posterior distribution update and refinement in the computational loop.

  3. This paper outperform on several evaluation metrics.
Critical weaknesses
  1. EIG based method purely focus on uncertainty. Beyond BAMCP, would comparison with sequential Bayesian design via approximate dynamic programming (e.g., Huan and Marzouk [1]), using the same terminal loss and budget, provide a fairer assessment? [1] Huan, Xun, and Youssef M. Marzouk. "Sequential Bayesian optimal experimental design via approximate dynamic programming." arXiv preprint arXiv:1604.08320 (2016).

  2. When predicting intervention-conditioned outcome distributions, do we need WAM? Could a simpler probabilistic dynamics model achieve comparable accuracy at lower computational cost?

Questions for authors
  1. WAM is computationally expensive, and the performance gain over Transolver in Table 31 appears modest relative to the computational cost. Does this gain justify the use of WAM in this setting? 
 

  2. What advantage does the neural-operator architecture offer over alternative probabilistic dynamics models (e.g., BNCDE [2]) in this setting? [2] Hess, Konstantin, et al. "Bayesian neural controlled differential equations for treatment effect estimation." International Conference on Learning Representations. Vol. 2024. 2024.
5 comments
AuthorsOpenReview ↗

Response to W1: Sequential Bayesian design and approximate dynamic programming

Dear Reviewer Xznv,

Thank you for your constructive feedback and for suggesting a matched sequential Bayesian design comparison via approximate dynamic programming (ADP [1]). Under the exact same terminal decision loss, assay menu, and measurement budget, NeuronSifter achieves the lowest terminal risk on multi-step CNS planning (0.160 vs. 0.166 for sequential BED and 0.193 for EIG, Table 2) and the lowest terminal loss on two-step sequential design (0.3326, Table 43), while eliminating Huan–Marzouk ADP's 35.69 s offline value-function fitting stage and accelerating online sequential planning by 27.3× over ADP (26.68 ms vs. 728.43 ms/episode) and 7.3×–21.6× over EIG, numerical BED, BAMCP, and adaptive BED (Appendix A.10.5, Table 43).

Reviewer Xznv (W1): EIG based method purely focus on uncertainty. Beyond BAMCP, would comparison with sequential Bayesian design via approximate dynamic programming (e.g., Huan and Marzouk [1]), using the same terminal loss and budget, provide a fairer assessment?

Sequential Planner (seed=42, 2,000 Bootstrap Resamples)5-Step CNS Risk (Budget 5, Table 2)2-Query Loss Including Queries [95% CI] (Table 43)Menu RegretOffline Value Fit (s)Online Planning Latency (ms/episode)Speedup vs. Baseline
NeuronSifter Complete-Mixture (Ours)0.1600.3326 [0.3292, 0.3363]0.000250.0026.681.0× (Reference)
Huan–Marzouk ADP [1]0.1650.3329 [0.3294, 0.3365]0.0002535.69728.4327.3× faster
Matched Numerical BED0.1660.3330 [0.3295, 0.3367]0.000250.00200.097.5× faster
BAMCP (Bayes-Adaptive MCTS)0.1740.3330 [0.3296, 0.3367]0.000250.25247.639.3× faster
Sequential EIG0.1930.3331 [0.3297, 0.3368]0.000250.00195.457.3× faster
Adaptive BED0.1680.3333 [0.3298, 0.3370]0.000250.00575.6921.6× faster
  1. Decision-Directed Advantage Over Uncertainty-Only EIG: On the 5-step closed-loop CNS benchmark (Table 2), replacing global entropy reduction (Sequential EIG, risk 0.193) with our decision-directed Value of Information (VoI) reduces terminal risk to 0.160 (-17.1% risk reduction, and $\Delta = -0.006\ [-0.009, -0.003]$ over matched numerical BED at 0.166) and reaches the 0.22 risk target at 33.0% lower measurement cost (2.286 vs. 3.410). Under independent generator shift (Table 42), decision-directed acquisition reduces external decision risk by -0.022 [-0.028, -0.016] compared with equal-budget EIG. On real Allen patch-clamp recordings (Table 46), 1-query Decision VoI outperforms equal-budget 1-query EIG by -1.81 Hz [-2.94, -0.72] (2.45× larger error reduction) at 20 Hz and -1.78 Hz [-2.88, -0.69] (2.28× gain) at 40 Hz.
  2. Computational Advantage Over Huan–Marzouk ADP [1]: Because NeuronSifter maintains a structured 32-particle joint exposure–response posterior, our complete-mixture planner evaluates expected terminal decision loss directly over the represented biological hypotheses in vectorized closed form (26.68 ms/episode), avoiding both backward-induction polynomial value regression (35.69 s offline fit + 728.43 ms/episode in Huan–Marzouk ADP [1]) and 4,096-sample nested Monte Carlo rollouts (195.45–575.69 ms/episode).

Thank you again for suggesting Huan and Marzouk [1]; adding this head-to-head comparison in Appendix A.10.5 (Table 43) directly demonstrates NeuronSifter's superiority in both multi-step decision risk and online planning efficiency, and we hope this addresses W1 and supports a positive reassessment.

References

  1. Xun Huan and Youssef M. Marzouk. Sequential Bayesian optimal experimental design via approximate dynamic programming. arXiv:1604.08320, 2016.
AuthorsOpenReview ↗

Response to W2: Necessity of WAM vs. simpler probabilistic dynamics models

Dear Reviewer Xznv,

Thank you for raising this important question on the role of the World Action Model (WAM) versus simpler probabilistic dynamics models. Expressive neural dynamics models are essential for accurate intervention-conditioned trajectory forecasting—our neural-operator backbone (Transolver) achieves 1,396× lower trajectory CRPS (0.0098 vs. 13.6846) than linear dynamics and 12.4× lower CRPS (0.0098 vs. 0.1216) than continuous-time BNCDE at 3.03× lower training cost (5.12 vs. 15.52 GPU-h, Table 44), while Diffusion WAM achieves the highest multi-step distributional fidelity on unseen intervention regimens (CRPS 0.110 vs. 0.119) and held-out intervention families (CRPS 0.121 vs. 0.134, ordering accuracy 0.880 vs. 0.865, Table 1 & Table 33).

Reviewer Xznv (W2): When predicting intervention-conditioned outcome distributions, do we need WAM? Could a simpler probabilistic dynamics model achieve comparable accuracy at lower computational cost?

Dynamics Model Tier (seed=42, 2,000 Bootstrap Resamples)Held-Regimen / 2-Query CRPS ↓\downarrowHeld-Family CRPS ↓\downarrowTrajectory NRMSE ↓\downarrowRegimen Ordering Accuracy ↑\uparrowTerminal Decision Risk / Loss ↓\downarrowOffline Training Cost
Diffusion WAM (Full Generative Twin, Table 1 & 33)0.1100.121 [-0.009 vs. WDNO]0.1620.8800.160 (5-step risk)44.51 GPU-h (8-step timing)
Transolver + Calibrated Residual (Fast Neural Tier, Table 44)0.0098 [0.0097, 0.0099]0.134 (Table 33)0.01630.8650.3326 [0.3292, 0.3360]5.12 GPU-h (3.03× faster than BNCDE)
BNCDE Variational SDE Adaptation (Table 44)0.1216 [0.1215, 0.1217]—0.2107—0.3327 [0.3292, 0.3360]15.52 GPU-h
Linear + Residual Baseline (Table 44)13.6846 [13.6615, 13.7083]—12.3313—0.3329 [0.3294, 0.3362]0.96 GPU-h (18.05 s SVD)
  1. Why Simple Parametric Models Fail on Trajectory Distributions: Simple linear dynamics (Linear + residual) cannot resolve coupled spatial-compartmental PDE transport and saturating receptor kinetics, resulting in severe trajectory distortion (CRPS 13.6846, NRMSE 12.3313—over 1,396× worse CRPS than Transolver).
  2. Two Complementary Neural Operating Tiers: NeuronSifter's modular interface supports both (i) a fast neural-operator tier (Transolver + residual, 5.12 GPU-h) that delivers sharp trajectory distributions (CRPS 0.0098, NRMSE 0.0163) and the lowest two-query terminal loss (0.3326) at low compute cost, and (ii) the full Diffusion WAM tier, which achieves the best multi-step rollout fidelity across unseen intervention families (CRPS 0.121, 90% coverage error 0.006 vs. 0.025 for WDNO) and highest regimen ranking accuracy (0.880, Table 33). Thank you again for prompting this accuracy-versus-compute comparison in Appendix A.10.5 (Table 44); we hope this directly resolves W2.
AuthorsOpenReview ↗

Response to Q1: Whether the performance gain over Transolver in Table 31/33 justifies WAM's computational cost

Dear Reviewer Xznv,

Thank you for asking us to quantify how diffusion WAM's predictive gains compare with its computational cost relative to Transolver. NeuronSifter provides a clear Pareto frontier across two operating regimes: (1) for high-stakes regimen discovery across unseen intervention mechanisms, Diffusion WAM achieves the best held-regimen CRPS (0.110 vs. 0.119), best held-family CRPS (0.121 vs. 0.134, -9.7% error reduction), highest pairwise regimen ordering accuracy (88.0% vs. 86.5%), and lowest biophysical violation rate (0.012 vs. 0.015, -20.0%, Table 33), while (2) for rapid high-throughput screening, our Transolver backbone delivers 0.340× rollout cost, 5.12 GPU-h full training (3.03× faster than BNCDE), sharp trajectory accuracy (CRPS 0.0098), and 0.3326 terminal decision loss (Table 44).

Reviewer Xznv (Q1): WAM is computationally expensive, and the performance gain over Transolver in Table 31 appears modest relative to the computational cost. Does this gain justify the use of WAM in this setting?

Predictor (Matched 64-Sample Rollout, Table 33)Held-Regimen CRPS ↓\downarrowHeld-Family CRPS ↓\downarrowRegimen Ordering Accuracy ↑\uparrowBiophysical Violation Rate ↓\downarrowRelative Rollout Cost
Diffusion WAM (Full Generative Twin)0.1100.121 (-9.7%)0.880 (+1.5 pp)0.012 (-20.0%)1.000×
Transolver Neural-Operator Backbone0.1190.1340.8650.0150.340×
Compute & Decision Hierarchy (NVIDIA A800-80GB, Table 44)Completed Epochs / UpdatesTraining GPU-hCalibration GPU-hConditioning + Prediction (s/episode)Online Planning (ms/episode)
Transolver + Calibrated Residual8 / 5045.121.773.509.05
BNCDE Variational SDE Adaptation8 / 50415.521.482.979.47
Linear + Residual Baseline20 / 12600.96 (18.05 s SVD)1.843.538.74
Diffusion WAM (High-Fidelity Tier)Frozen checkpoint + 8-step timing44.51 (8 steps)—48.00 (64 paths, Table 34)26.68
  1. Why Diffusion WAM Is Justified for Final Regimen Prioritization: In CNS therapeutics, advancing a suboptimal intervention regimen to wet-lab or in-vivo validation incurs substantial experimental cost. Improving pairwise regimen ordering accuracy from 86.5% to 88.0%, reducing unseen intervention-family CRPS by 9.7% (0.121 vs. 0.134), and reducing biophysical constraint violations by 20.0% (0.012 vs. 0.015) directly prevents costly downstream misrankings, yielding our lowest 5-step closed-loop risk (0.160, Table 2).
  2. Fast Neural Screening Tier (Transolver): For large-scale candidate pre-screening where compute budget is primary, Transolver (5.12 GPU-h training, 0.340× rollout cost) integrates seamlessly with NeuronSifter's shared exposure–response posterior and occupancy compiler. Thank you again for highlighting this practical trade-off in Appendix A.10.5 (Tables 33 & 44); we hope this supports a favorable reassessment.

Comments on the paper

20 comments
AuthorsOpenReview ↗

Response to Reviewer PCGF

Thank you for this question. Below we (1) confirm both methods share the exact same candidate bank, posterior update, and scoring protocol in every run, (2) explain why Round 1 pre-filtered the candidate bank to the 3 decision-relevant action channels and what changed between the two runs, and (3) give the 2×22 \times 2 reconciliation table.

1. Identical Candidate Bank & Scoring Protocol Within Every Run

Within the new run—and within each row of the 2×22 \times 2 table—1-Query Decision VoI and 1-Query EIG share strictly identical conditions:

  1. Cohort & episodes: Same 142 completed-record Allen cells across 104 donors (1,704 episodes; 2,000 paired donor-bootstrap resamples, seed = 42).
  2. Candidate query bank (QQ): At step 1, both score the exact same candidate bank QQ for that cell.
  3. Likelihood & update: Same Gaussian query model (σq=8 Hz\sigma_q = 8\,\text{Hz}), 65-pt Gauss–Hermite quadrature, and closed-form affine Gaussian update on y(q1)y(q_1).
  4. Action set (AA) & metric: After 1 query, both pick a^∈A={low, mid, high}\hat{a} \in A = \{\text{low, mid, high}\} to minimize posterior expected target error E[∣y(a)−r⋆∣∣D1]\mathbb{E}[\lvert y(a) - r^\star \rvert \mid D_1], scored by true held-out error ∣ytrue(a^)−r⋆∣\lvert y_{\text{true}}(\hat{a}) - r^\star \rvert.

2. Why Round 1 Pre-Filtered QQ to the 3 Action Channels—and What Changed (16.63 →\to 20.41 Hz for EIG; 18.60 Hz Invariant for VoI)

Two orthogonal factors separate the Round 1 table from the expanded evaluation: (A) Candidate Bank Scope (QactQ_{\text{act}} vs. QfullQ_{\text{full}}) and (B) Shared 46-Donor Offset Calibration.

A. Why Round 1 Pre-Filtered Qact=AQ_{\text{act}} = A, and What Changed on QfullQ_{\text{full}}:

  • Why Round 1 had a manual pre-filter: In Round 1, candidate queries were indexed directly from the 3-level action table Qact=A={low, mid, high}Q_{\text{act}} = A = \{\text{low, mid, high}\} (426 programs) to enforce a 1-to-1 probe-to-action match. However, as you noted on Nov 14, restricting queries to only those 3 decision-relevant channels artificially shields global EIG from nuisance/off-target channels: when every allowed query is already a decision-relevant action channel, global entropy maximization cannot waste its query on informative-but-decision-irrelevant sweeps.
  • Why we removed the pre-filter in the new run (Qfull=A∪QnuisQ_{\text{full}} = A \cup Q_{\text{nuis}}): Removing that manual pre-filter exposes the full completed Allen sweep bank QfullQ_{\text{full}} (adding high-variance subthreshold/off-target nuisance sweeps QnuisQ_{\text{nuis}} alongside AA):
    1. Uncalibrated VoI stays at 18.60 Hz: VoI weights queries solely by how much their update can flip a⋆∈Aa^\star \in A around 20 Hz. Nuisance sweeps QnuisQ_{\text{nuis}} have near-zero covariance with the {low, mid, high}\{\text{low, mid, high}\} boundary, so VoI gives them near-zero value and keeps selecting from AA (18.60 Hz in both runs).
    2. Uncalibrated EIG rises from 16.63 to 20.41 Hz: Global EIG maximizes total mutual information across all latent dimensions. With QnuisQ_{\text{nuis}} present, EIG picks high-spread nuisance sweeps with large global variance reduction but weak resolution near the 20 Hz boundary—raising EIG error from 16.63 Hz (when pre-filtered to AA) to 20.41 Hz on QfullQ_{\text{full}} (2.45× larger error reduction for VoI over 0-Query = 21.66 Hz; paired VoI−EIG=−1.81 Hz\text{VoI} - \text{EIG} = -1.81\,\text{Hz}, [-2.94, -0.72]).

B. Shared 46-Donor Calibration (18.60 →\to 16.60 Hz for VoI): Even on QactQ_{\text{act}}, the uncalibrated gap (18.60 vs. 16.63 Hz) came from raw cross-donor firing-rate offset bias (+3–5 Hz), making uncalibrated VoI under-trigger updates (18.3% vs. 24.8% calibrated). Fitting a shared 2-parameter affine offset calibration on 46 disjoint training donors (applied identically to both methods) restores calibrated posteriors (98.61% marginal / 93.27% donor-panel 90% coverage) and lowers VoI error to 16.60 Hz, beating EIG on both QfullQ_{\text{full}} (16.60 vs. 18.35 Hz) and QactQ_{\text{act}} (16.60 vs. 17.24 Hz).

3. 2×22 \times 2 Reconciliation Table at 20 Hz (142 Cells, 104 Donors)

Bank QQModel0-Query1-Query EIG1-Query VoIPaired VoI−EIG\text{VoI} - \text{EIG}
Full QfullQ_{\text{full}}Uncalibrated21.6620.41 [18.18, 23.12]18.60 [16.57, 22.09]-1.81 [-2.94, -0.72]
Full QfullQ_{\text{full}}46-Donor Calibrated21.4018.35 [15.42, 21.88]16.60 [13.94, 19.99]-1.75 [-2.89, -0.64]
3-Action QactQ_{\text{act}}46-Donor Calibrated21.4017.24 [13.72, 21.74]16.60 [13.94, 19.99]-0.64 [-3.21, +1.29]
3-Action QactQ_{\text{act}}Uncalibrated21.6616.63 [15.55, 17.73]18.60 [16.57, 22.09]+1.97 [+0.42, +4.75]

Added to Appendix T–U.

AuthorsOpenReview ↗

Q2: Decision-directed acquisition versus model-criticism querying

Dear Reviewer PCGF,

Thank you for asking whether acquisition can select measurements that challenge the current model rather than only optimizing decisions within it. Decision-directed acquisition optimizes expected terminal loss reduction under the current predictive model and therefore cannot guarantee selection of measurements whose outcomes lie outside that model's support; our new 2,800-context active discrepancy-probing diagnostic (Table below) shows that explicitly querying for prediction residuals flags 796-811/2,800 contexts at 1×1\times noise and 1,442-1,449/2,800 at 2×2\times noise, demonstrating sensitivity to observation noise rather than specific discovery of omitted biological mechanisms.

Reviewer PCGF (Q2):

Can acquisition select measurements that challenge the current model, rather than only improve decisions within it? Which experiments demonstrate this capability?

Querying modeQuery varianceQueriesFlagged discrepancy contexts / 2,800Stress false accepts / accepted (τfrac=0.25\tau_{\text{frac}}=0.25)Conditional false acceptance
Selective decision value (cost-aware)1x0064 / 12750.39%
EIG (information gain)1x279663 / 12650.00%
Active discrepancy probing1x281153 / 9555.79%
EIG (information gain)2x21,44963 / 12849.22%
Active discrepancy probing2x21,44227 / 5450.00%

In this diagnostic (query cost 0.0035), the cost-aware decision policy selects 0 queries when represented hypotheses already agree on the best action, whereas EIG and active discrepancy probing each spend 2 queries. Flagging standardized innovations z2>2z^2 > 2 increases nearly identically for EIG (796 -> 1,449) and active discrepancy (811 -> 1,442) when observation noise doubles (1×→2×1\times \to 2\times), and in the separate 800-context query-noise stress test (Table 45) the gate defers all 800 cases at τfrac=0.25\tau_{\text{frac}}=0.25 (349/349 high-loss cases intercepted via complete deferral). Thus, discrepancy querying supports conservative deferral under observation perturbation, not discovery of unmodeled mechanisms.

Thank you again for raising this foundational distinction; we have made this scope boundary explicit in the revised manuscript.

AuthorsOpenReview ↗

W2 & Q3 (Part 1): Independent biological validation of inadequate exposure versus nonresponse

Dear Reviewer PCGF,

Thank you for asking which independent biological measurements validate the distinction between inadequate exposure and biological nonresponse. Public multi-dose microglial RNA-seq from 62 hTREM2 knock-in mice (GSE198987, 48 active-arm mice across days 1, 4, and 7; van Lengerich et al., 2023) directly validates that exposure and biological responsiveness are distinct axes: under the source study's receptor-occupancy threshold (KD=2.00 nMK_D = 2.00\text{ nM}) and phagocytic signature cutoff (ΔMME=1.447 log⁡2(TMM CPM+1)\Delta_{\text{MME}} = 1.447\ \log_2(\text{TMM CPM}+1)), all four cells of the 2×22\times 2 exposure-by-response table are populated (20 adequate-exposure & responsive, 4 adequate-exposure & nonresponsive despite saturating 10 mg/kg dosing, 8 sub-threshold & responsive, and 16 inadequate-exposure & nonresponsive; Appendix A.5).

Reviewer PCGF (W2):

Intervention selection depends on the compiled representation and assay likelihoods, whose biological validity remains incompletely established.

Reviewer PCGF (Q3):

Which independent biological measurements validate the distinction between inadequate exposure and biological nonresponse?

Cohort-assigned brain exposure (KD=2.00 nMK_D = 2.00\text{ nM})Measured microglial RNA-seq response (ΔMME=1.447\Delta_{\text{MME}} = 1.447)Active-arm mice / 48 (GSE198987)Regimen breakdown (seed=42, 2,000 resamples)
Adequate (≥KD\ge K_D)Responsive (≥ΔMME\ge \Delta_{\text{MME}})20 (41.7%)10 mg/kg: 8/8 Day 1 (+2.894), 5/8 Day 4 (+2.107), 7/8 Day 7 (+2.398)
Adequate (≥KD\ge K_D)Nonresponsive ($\Delta< \Delta_{\text{MME}}$)
Inadequate (<KD< K_D)Responsive (≥ΔMME\ge \Delta_{\text{MME}})8 (16.7%)1 mg/kg: 5/8 Day 1 (+1.592), 1/8 Day 4, 2/8 Day 7
Inadequate (<KD< K_D)Nonresponsive ($\Delta< \Delta_{\text{MME}}$)

All four 2×22\times 2 categories remain populated under the alternative 4.1 nM exposure cutoff and alternative proliferation or Tspo markers (Appendix A.5). Two additional datasets—mouse microglial pFDG PET (+0.172, +0.249, -0.269 across TREM2 strata, Table 15) and human iPSC-microglia RNA-seq (GSE199154, +0.460 [+0.358, +0.562] 96h G2M contrast across 4/4 batches, Table 16)—confirm state-dependent transcriptional and metabolic responses, while real Allen recordings (142 cells, 104 donors, Table 46) test closed-loop stimulus selection. We explicitly note that GSE198987 assigns exposure from published cohort pharmacokinetics rather than per-animal paired target-occupancy assays, and does not claim prospective clinical AD drug-selection accuracy.

Thank you again for prompting this three-level separation between biological premise validation, model-replay sensitivity, and end-to-end clinical validation.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.