Pith. sign in

REVIEW 4 major objections 4 minor 28 references

Probabilistic Factorial Experimental Design for Combinatorial Interventions

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper proves that giving every candidate treatment a dosage of $1/2$ is near-optimal for estimating bounded-degree interaction effects from randomly assigned combinations, yielding an $O(k p^{3k}\ln p)$ sample-complexity guarantee.

desk verdict The framework is promising but the central theorem's proof breaks in two places, so the near-optimality and sample-complexity claims should not be cited as proven. read the letter →

arxiv 2506.03363 v1 pith:33PYSRE7 submitted 2025-06-03 cs.LG stat.MEstat.ML

Divya Shyamal , Jiaqi Zhang , Caroline Uhler This is my paper
classification cs.LGstat.MEstat.ML MSC 62K0562K15
keywords combinatorialinterventionsprobabilisticfactorialdesignBooleanfunctionsFourieranalysisoptimalexperimentalactivelearningbounded-degreeinteractionssamplecomplexity
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper introduces a probabilistic factorial design: an experimenter fixes a dosage $d_i\in[0,1]$ for each of $p$ treatments, and every unit independently receives treatment $i$ with probability $d_i$, so combinations are randomized rather than pre-selected. Within a Boolean-function model of outcomes with interactions only up to order $k$, it claims that the half-dosage vector $d=(1/2,\ldots,1/2)$ is optimal up to a factor of $1+O(\ln(n)/n)$ for the expected squared error of the truncated least-squares estimator of the Fourier coefficients, for every $k$. This matters because it turns a potentially expensive design problem into a fixed rule: randomize each treatment independently with probability one half, and the design matrix is as well conditioned as possible. The same argument yields that $O(k p^{3k}\ln p)$ observations suffice to estimate a $k$-way interaction model, and it supports a numerically optimized acquisition function for multi-round adaptive experiments.

What carries the argument

The machinery has four parts. First, the product Bernoulli dosage model, in which each unit's combination $x_m\in\{-1,1\}^p$ is sampled independently with $P(x_{m,i}=1)=d_i$; this generalizes full and fractional factorial designs. Second, the Fourier basis $\phi_S(x)=\prod_{i\in S}x_i$ for real-valued Boolean functions, with coefficients $\beta_S$ that encode $k$-way interactions under the bounded-order assumption $\beta_S=0$ for $|S|>k$. Third, the truncated OLS estimator, which returns zero when the inverse eigenvalue sum exceeds $B^2/\sigma^2$, giving two-sided bounds on the expected error in terms of $\lambda_i(X^\top X)$. Fourth, the expected Gram matrix $\Sigma(d)$ with entries $\Sigma(d)_{S,S'}=\prod_{i\in S\triangle S'}(2d_i-1)$, whose trace is $K=\sum_{i=0}^k\binom{p}{i}$ for every $d$, so eigenvalue balancing via Cauchy-Schwarz identifies $d=(1/2,\ldots,1/2)$ as the unique minimizer of the idealized inverse-trace objective. Concentration of $(1/n)X^\top X$ around $\Sigma(d)$, established in Lemma B.5 through an $\epsilon$-net and Hoeffding argument, transfers this to the actual random design and supplies the $1+O(\ln n/n)$ factor.

What would settle it

Run Monte Carlo for a fixed small model, say $p=8$, $k=2$, $\sigma=1$, and compare the truncated-OLS expected squared error at $d=(1/2,\ldots,1/2)$ with the best dosage found by search at the same $n$; if the ratio exceeds $1+O(\ln n/n)$ for large $n$, Theorem 4.2 fails. A direct check of the proof would evaluate the failure probability in Lemma B.5 at $\delta=\sqrt{2\ln n/n}$ and verify that the resulting term $n^{1-1/(4K^2)}$ actually vanishes fast enough to keep the upper-to-lower bound ratio at $1+O(\ln n/n)$.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is Theorem 4.2: for the truncated OLS estimator, the design $d=(1/2,\ldots,1/2)$ minimizes $\mathbb{E}_D[\lVert\hat{\beta}-\beta\rVert_2^2]$ up to a multiplicative factor $1+O(\ln(n)/n)$ whenever the outcome is a bounded-degree Boolean function of the received combination, and any optimal design lies in an $\ell_\infty$ ball of radius $O(\sqrt{\ln(n)/n})$ around the half dosage. The intuition is that the expected Gram matrix $\Sigma(d)$ has trace $K$ for every $d$, so by Cauchy-Schwarz its inverse-trace is minimized exactly when all eigenvalues equal one, which happens only at the half dosage. The proof transfers this intuition to the random design via a concentration bound on $(1/n)X^\top X$ and a truncated estimator that avoids exploding inverse-eigenvalue terms. A corollary states that the estimation error decays as $\leq (2K\sigma^2+1)/n$ for $n>n_0$ with $n_0=O(K^3\ln K)$, giving the $O(kp^{3k}\ln p)$ sample-complexity statement, and Theorem 4.4 provides a near-optimal acquisition function for the multi-round active setting.

Load-bearing premise

The proof depends on the concentration estimate in Lemma B.5, which says that $(1/n)X^\top X$ stays close to its expectation $\Sigma(d)$ with high probability; the chosen failure probability decays too slowly to sustain the stated $1+O(\ln n/n)$ optimality factor, so if that concentration bound cannot be sharpened, the quantitative claims collapse.

Editorial extensions

If this is right

  • For any bounded-degree interaction model, the same fixed dosage rule—randomize each treatment independently with probability $1/2$—is near-optimal regardless of the interaction order $k$; no first-round design can beat it by more than $1+O(\ln n/n)$.
  • Estimating a $k$-way interaction model requires only $O(kp^{3k}\ln p)$ observations rather than the full $2^p$ factorial enumeration, making experiments with tens of treatments and pairwise interactions feasible.
  • In multi-round active experiments, the acquisition function in Theorem 4.4 is near-optimal and can be solved numerically, so sequential rounds can rebalance dosages toward poorly explored directions.
  • Under a limited-supply constraint $\sum_i d_i \le L < p/2$, the uniform dosage $d_i=L/p$ is near-optimal for the additive model, with simulation evidence suggesting the same for interactive models.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: because the proof relies only on the trace and eigenvalue balance of the expected Gram matrix, the same near-optimality should extend to the ridge estimator analyzed in Lemma B.4 and to other linear estimators whose error bounds are driven by $\sum_i 1/\lambda_i$.
  • Editorial inference: the product-Bernoulli assumption excludes interference and censoring; in settings where treated cells die or neighbors affect outcomes, the effective design matrix is no longer product-Bernoulli, and a dosage schedule maximizing coverage of observable combinations could differ from $1/2$.
  • Editorial inference: a sharper spectral concentration bound would likely replace the $\ln(n)/n$ factor with a smaller term and improve the hidden constant in $n_0=O(K^3\ln K)$, since the current slack comes from balancing the failure probability against the eigenvalue error.
  • Editorial inference: the closed-form result suggests a practical rule for pooled genetic screens—set the multiplicity of infection so each guide is present with probability $1/2$ per cell, then estimate interaction coefficients by Fourier regression—which is testable against existing perturbation datasets without running new assays.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper introduces probabilistic factorial experimental design, in which an experimenter chooses dosages d and each unit independently receives a product-Bernoulli combination of treatments. Within a bounded-degree Boolean interaction model, the paper studies estimation of Fourier coefficients by a truncated OLS estimator. Its main passive claim is that the half-dosage design d=(1/2,...,1/2) is optimal up to a factor of 1+O(ln(n)/n) with respect to the expected squared error, with a claimed sample complexity O(kp^{3k} ln p). It also proposes an active acquisition function with a claimed analogous near-optimality guarantee, several extensions, and simulation experiments.

Significance. If the quantitative claims were correct, the paper would provide a clean and practically relevant design principle: for bounded-degree interaction models, a uniform half-dosage product-Bernoulli design is near-optimal, with sample complexity polynomial in p for fixed interaction degree k. The paper is also explicit about its estimator and limitations, and it ships reproducible simulation code. However, the central proof of Theorem 4.2 contains two independent errors, and the claimed optimality factor and sample-complexity rate do not follow from the arguments in the appendix. Because Theorem 4.2 and its corollary are the load-bearing results of the paper, the quantitative contributions are not established.

major comments (4)
  1. [Appendix B.3, Eq. (11), with §4.2 proof sketch] Theorem 4.2's claimed factor 1+O(ln(n)/n) is not established. With the proof's choice delta=(2 ln n / n)^{1/2}, the bad-event term in Eq. (11) is nB^2 exp(K ln 9 - n delta^2/(8K^2)) = B^2 9^K n^{1-1/(4K^2)}, which grows with n for every fixed K>=2, so the upper bound is not K sigma^2/(1-delta)+o(1). Even if one uses the unnormalized display in the main-text proof sketch, where the prefactor n is absent, the failure term is B^2 9^K n^{-1/(4K^2)} and the denominator 1-delta contributes a relative error of order sqrt(ln n/n); neither yields the stated 1+O(ln(n)/n) optimality factor. The final sentence of the appendix proof, 'Choosing delta_1=delta_2=(2 ln(n)/n)^{1/2}, we have optimality to a factor of 1+O(ln(n)/n)', therefore does not follow.
  2. [Appendix B.3, Eq. (12) and the following display] The lower-bound step is invalid. The proof asserts sum_i 1/(lambda_i(Sigma)+delta) >= 1/(c+delta)+(K-1)/(1+delta) from lambda_min(Sigma)<=c, but knowing one eigenvalue is at most c does not imply the remaining K-1 eigenvalues are at least 1. A concrete counterexample is p=3, k=1, d=(0.5,0.5,0.9), delta=0, where the eigenvalues of Sigma are {1.8,0.2,1,1}; the left-hand side is approximately 7.56 while the right-hand side is 1/0.2+3=8. Thus the lower bound on the expected error for arbitrary dosages is not proved, and the subsequent optimization over c cannot be used.
  3. [Appendix B.4, Eqs. (14)-(15)] Theorem 4.4, the active-setting near-optimal acquisition function, inherits the same concentration defect. The proof chooses delta_1=delta_2=(2 ln n/n)^{1/2} and claims a factor 1+O(ln(n)/n), but the bad-event term in Eq. (14) is again B^2 9^K n^{1-1/(4K^2)} up to constants, and the denominator terms in the ratio introduce a relative error of order sqrt(ln n/n). Since the active acquisition result is a central contribution and rests on this argument, Theorem 4.4 is also unsupported.
  4. [Corollary 4.3] The sample-complexity claim O(kp^{3k} ln p) does not follow. The corollary is stated as a consequence of the proof of Theorem 4.2, but that proof supplies no O(1/n) bound on the bad-event term at n=O(K^3 ln K); with the chosen delta the failure term is of order B^2 9^K n^{1-1/(4K^2)} (or n times that, depending on the normalization), which is not bounded by 1/n. The claimed threshold n0=O(K^3 ln K) is therefore not attained by the argument given.
minor comments (4)
  1. [§4.2 proof sketch versus Appendix B.3] The main-text proof sketch displays the bad-event term for d=(1/2,...,1/2) without the prefactor n that appears in Eq. (11) of Appendix B.3; the two displays lead to different decay rates and the manuscript should reconcile which normalization is intended.
  2. [§4.3] The stated complexity O(np^{3k+3}+p^{6k+3}) is not consistent with K=O(p^k) for fixed k. With O(nK^2+K^3) per iteration and O(p^3) iterations, the more direct expression is O(np^{2k+3}+p^{3k+3}); the manuscript should either correct the estimate or define K differently.
  3. [Corollary 4.3] The equality O(K^3 ln K)=O(k p^{3k} ln p) is stated too loosely; for fixed k the factor k is not obtained from K^3 ln K alone, and for growing k the comparison should be written as an upper bound rather than an equality.
  4. [§6.2] The experiments show a qualitative U-shaped agreement with the half-dosage optimum, but they do not probe the asymptotic factor 1+O(ln(n)/n). I recommend an explicit sentence clarifying that the simulations are not a direct test of the quantitative optimality factor.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the half-dosage optimality result is derived from the spectrum of Sigma(d) and a concentration bound, not from a fitted parameter or self-citation.

full rationale

The paper's central claim, Theorem 4.2, is that d=(1/2,...,1/2) is near-optimal for the truncated OLS estimator under the MSE objective Eq. (4). The proof does not reduce to its inputs by construction. The estimator in Section 4.1 is defined independently of d; its error property (Lemma 4.1) is a bound on the MSE in terms of the eigenvalues of X^T X and the assumed bound ||β||_2 ≤ B. The dosage enters only through the random design matrix, whose expectation is Σ(d) in Eq. (5). Lemma B.1 shows that tr(Σ(d))=K and λ_min(Σ(d)) ≤ 1 with equality iff d=(1/2,...,1/2); this is a mathematical fact about product-Bernoulli Fourier moments, not an assumption that half dosage is optimal. The upper bound for half dosage and the lower bound for arbitrary d are then obtained by the concentration estimate Lemma B.5 and Weyl's inequality, with δ chosen to balance terms. This is a standard concentration-based proof strategy. There are no fitted parameters later reported as predictions, and no load-bearing self-citations: the only author self-citation (Zhang et al., 2023) appears in related work and is not used to justify Theorem 4.2. The choice of the truncated OLS estimator is an explicit modeling decision, and the theorem is explicitly scoped to that estimator, so the claim does not silently redefine the objective. The reader's flagged concerns about the size of the concentration term and the validity of Eq. (12) are potential correctness gaps in the proof, but they are not circularity: they do not make the conclusion equivalent to an input or to a fitted value. Accordingly, no circular step is identified and the circularity score is 0.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

The framework introduces no fitted constants beyond the proof slack delta. The theoretical claims rest on the bounded-order Boolean model, the L2 bound on beta, the Gaussian noise model, and the spectral concentration of the random design matrix. No new entities are postulated. The main proof issue is the strength of the concentration estimate, which is a technical step rather than an additional axiom.

free parameters (1)
  • delta (concentration slack) = sqrt(2 ln n / n)
    Chosen in the proof of Theorem 4.2. With this value the bad-event term in the upper bound scales as n^{1 - 1/(4K^2)}, so the claimed optimality factor does not follow from the presented equations.
assumptions (5)
  • domain assumption beta_S = 0 for |S| > k, with k = o(p) (Assumption 3.1)
    The outcome function is assumed to have bounded-order interactions; this defines the model class being estimated.
  • domain assumption ||beta||_2 <= B (Assumption 3.2)
    The truncated OLS threshold depends on B; the expected-error bounds use B as the worst-case norm.
  • domain assumption y_m = f(x_m) + epsilon_m with independent zero-mean Gaussian noise of variance sigma^2
    Defines the observation model; the estimator's threshold uses sigma^2.
  • domain assumption Each unit receives an independent product Bernoulli treatment vector with no interference between units
    This is the design model (Eq. 1); interference or censoring would break the product structure, as acknowledged in the limitations.
  • standard math Fourier basis {phi_S} is orthonormal and beta_S equals the Fourier coefficient
    Used to represent f and to relate ||fhat-f||_2 to ||beta_hat-beta||_2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Probabilistic Factorial Experimental Design for Combinatorial Interventions." pith.science (2026). https://pith.science/paper/33PYSRE7

@misc{pith2026250603363,
  author       = {Pith},
  title        = {Pith review of: Probabilistic Factorial Experimental Design for Combinatorial Interventions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/33PYSRE7}},
  note         = {Machine review of arXiv:2506.03363}
}
abstract

A combinatorial intervention, consisting of multiple treatments applied to a single unit with potentially interactive effects, has substantial applications in fields such as biomedicine, engineering, and beyond. Given $p$ possible treatments, conducting all possible $2^p$ combinatorial interventions can be laborious and quickly becomes infeasible as $p$ increases. Here we introduce probabilistic factorial experimental design, formalized from how scientists perform lab experiments. In this framework, the experimenter selects a dosage for each possible treatment and applies it to a group of units. Each unit independently receives a random combination of treatments, sampled from a product Bernoulli distribution determined by the dosages. Additionally, the experimenter can carry out such experiments over multiple rounds, adapting the design in an active manner. We address the optimal experimental design problem within an intervention model that imposes bounded-degree interactions between treatments. In the passive setting, we provide a closed-form solution for the near-optimal design. Our results prove that a dosage of $\tfrac{1}{2}$ for each treatment is optimal up to a factor of $1+O(\tfrac{\ln(n)}{n})$ for estimating any $k$-way interaction model, regardless of $k$, and imply that $O\big(kp^{3k}\ln(p)\big)$ observations are required to accurately estimate this model. For the multi-round setting, we provide a near-optimal acquisition function that can be numerically optimized. We also explore several extensions of the design problem and finally validate our findings through simulations.

Figures

Figures reproduced from arXiv: 2506.03363 by the authors.

Figure 1
Figure 1. Simulation 1. Average ∥βˆ − β∥ 2 2 over 2000 different observation sets generated from 100 different dosages at each given distance. The bars correspond to ±.5 std over the 2000 observations. The curves are generated with values p = 10, k = 2, n = 200; p = 20, k = 2, n = 1000; and p = 30, k = 2, n = 1000. 7 [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Simulation 2. Approximation error of uniform dosages. Bars correspond to ±.5 std over 500 observations per dosage. The curves are generated with values p = 10, k = 2, n = 200; p = 20, k = 2, n = 1000; and p = 30, k = 2, n = 1000. We show these results for three different sets of p, k, and n in [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 4
Figure 4. Active setting with relative small n, high noise. Results are averaged over 50 trials, where p = 5, k = 1, n = 16, σ = 5. Bars correspond to ±.1 std. where there are not many samples (compared to features) per round, we find that the optimal acquisition strategy more clearly outperforms the half strategy. This is because when we have a smaller number of samples, we will need to“correct” as the distribution of combin… view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: P Limited Supply Constraint. Here d needs to satisfy p i=1 di ≤ 2. The x-axis shows the l∞-distance from the L/p uniform dosage . We vary p over 8, 9, 10, keeping k = 2 and n = 1000 constant. 50 different dosages are sampled at each distance, with 40 iterations of each…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 28 canonical work pages

  1. [1]

    Bliss, C. I. The toxicity of poisons applied jointly 1. Annals of applied biology, 26(3):585–615, 1939

    work page 1939

  2. [2]

    E., Hunter, W

    Box, G. E., Hunter, W. G., and Hunter, S. Statistics for Experimenters, volume 664. John Wiley & Sons, New York, 1978

    work page 1978

  3. [3]

    and Cheng, C.-S

    Chang, M.-C. and Cheng, C.-S. A bayesian approach to the selection of two-level multi-stratum factorial designs. The Annals of Statistics, 46(4):1779–1806, 2018

    work page 2018

  4. [4]

    Theory of factorial design

    Cheng, C.-S. Theory of factorial design . Chapman and Hall/CRC Boca Raton, FL, USA, 2016

    work page 2016

  5. [5]

    and Bareinboim, E

    Correa, J. and Bareinboim, E. A calculus for stochastic interventions: Causal effect identification and surrogate experiments. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pp. 10093–10100, 2020

    work page 2020

  6. [6]

    and V oss, D.Design and analysis of experiments

    Dean, A. and V oss, D.Design and analysis of experiments. Springer, 1999

    work page 1999

  7. [7]

    Deming, S. N. and Morgan, S. L. Experimental design: a chemometric approach. Elsevier, 1993

    work page 1993

  8. [8]

    Causation and intervention

    Eberhardt, F. Causation and intervention. Unpublished doctoral dissertation, Carnegie Mellon University , 93, 2007

    work page 2007

Show all 28 references
  1. [9]

    and Scheines, R

    Eberhardt, F. and Scheines, R. Interventions and causal inference. Philosophy of science, 74(5):981–995, 2007

    2007

  2. [10]

    A., and G´en´eticien, S

    Genetician, S., Britain, G., Fisher, R. A., and G´en´eticien, S. The design of experiments, volume 21. Springer, 1966

    1966

  3. [11]

    and Hunter, W

    Fries, A. and Hunter, W. G. Minimum aberration 2k–p designs. Technometrics, 22(4):601–608, 1980

    1980

  4. [12]

    Gunst, R. F. and Mason, R. L. Fractional factorial design. Wiley Interdisciplinary Reviews: Computational Statis- tics, 1(2):234–244, 2009

    2009

  5. [13]

    and Lu, K

    Hanrahan, G. and Lu, K. Application of factorial and re- sponse surface methodology in modern experimental de- sign and optimization. Critical reviews in analytical chemistry, 36(3-4):141–151, 2006

    2006

  6. [14]

    Kerr, M. K. Bayesian optimal fractional factorials.Statistica Sinica, pp. 605–630, 2001

    2001

  7. [15]

    Koonin, E. V . Orthologs, paralogs, and evolutionary ge- nomics. Annu. Rev. Genet., 39(1):309–338, 2005

    2005

  8. [16]

    Drug interaction: focusing on response surface models

    Lee, S.-i. Drug interaction: focusing on response surface models. Korean journal of anesthesiology, 58(5):421– 434, 2010

    2010

  9. [17]

    Effect of combinations: mathematical basis of problem

    Loewe, S. Effect of combinations: mathematical basis of problem. Arch. Exp. Pathol. Pharmakol., 114:313–326, 1926

    1926

  10. [18]

    Experimental design and optimization

    Lundstedt, T., Seifert, E., Abramo, L., Thelin, B., Nystr¨om, ˚A., Pettersen, J., and Bergman, R. Experimental design and optimization. Chemometrics and intelligent labora- tory systems, 42(1-2):3–40, 1998

    1998

  11. [19]

    J., Morris, M

    Mitchell, T. J., Morris, M. D., and Ylvisaker, D. Two-level fractional factorials and bayesian prediction. Statistica Sinica, pp. 559–573, 1995

    1995

  12. [20]

    Montgomery, D. C. Design and Analysis of Experiments. John Wiley & Sons, 2017

    2017

  13. [21]

    and Wu, C

    Mukerjee, R. and Wu, C. F. J.A Modern Theory of Factorial Design. Springer Science & Business Media, 2007. O’Donnell, R. Some topics in analysis of boolean functions. In Proceedings of the fortieth annual ACM symposium on Theory of computing, pp. 569–578, 2008

    2007

  14. [22]

    E., Hupalowska, A., and Regev, A

    Rood, J. E., Hupalowska, A., and Regev, A. Toward a foundation model of causal cell and tissue biology with a perturbation cell and tissue atlas. Cell, 187(17):4520– 4545, 2024. 10 Probabilistic Factorial Experimental Design for Combinatorial Interventions

    2024

  15. [23]

    Non- parametric synergy modeling of chemical compounds with gaussian processes

    Shapovalova, Y ., Heskes, T., and Dijkstra, T. Non- parametric synergy modeling of chemical compounds with gaussian processes. BMC bioinformatics, 23:1–30, 2022

    2022

  16. [24]

    and Yamanaka, S

    Takahashi, K. and Yamanaka, S. Induction of pluripotent stem cells from mouse embryonic and adult fibroblast cultures by defined factors. cell, 126(4):663–676, 2006

    2006

  17. [25]

    High-dimensional probability: An introduc- tion with applications in data science, volume 47

    Vershynin, R. High-dimensional probability: An introduc- tion with applications in data science, volume 47. Cam- bridge university press, 2018

    2018

  18. [26]

    Scipy 1.0: fundamental algorithms for scientific computing in python

    Weckesser, W., Bright, J., et al. Scipy 1.0: fundamental algorithms for scientific computing in python. Nature methods, 17(3):261–272, 2020

    2020

  19. [27]

    J., Dey, K., Geiger-Schuller, K., Eraslan, B., Gusev, A., et al

    Yao, D., Binan, L., Bezney, J., Simonton, B., Freedman, J., Frangieh, C. J., Dey, K., Geiger-Schuller, K., Eraslan, B., Gusev, A., et al. Scalable genetic screening for reg- ulatory circuits using compressed perturb-seq. Nature biotechnology, 42(8):1282–1295, 2024

    2024

  20. [28]

    Active learning for optimal intervention design in causal models

    Uhler, C. Active learning for optimal intervention design in causal models. Nature Machine Intelligence, 5(10): 1066–1075, 2023. 11 Probabilistic Factorial Experimental Design for Combinatorial Interventions A. Proofs for Section 3 A.1. Fourier Representation Lemma A.1. f admi...

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.