acceptodds
Under review as a conference paper at ICLR 2027

GenomeCoder: Neural-Guided Inference of Executable Biological Programs

Abstract

Many latent quantities in biology are short, executable procedures: an immune receptor records which gene segments were joined and trimmed, an expressed gene is a set of transcript isoforms, a genealogy is a sequence of local trees linked by recombination. Mechanistic models assign probabilities to such programs, while the candidate space grows combinatorially, so exact inference is affordable only in a narrow band. We present **GenomeCoder**, which trains a recognition model to construct a compact support of executable latent programs, scores that support under the mechanistic model, and, when generative parameters are unknown, learns them from observations without latent labels. Where the exact posterior is enumerable, GenomeCoder reaches the *Bayes-optimal recovery rate* in V(D)J recombination, against of it for beam search, and reaches of the splicing ceiling in exact evaluations. For a fixed mechanistic model, the restricted posterior's error is exactly the posterior mass the support omits, measured exactly in V(D)J and splicing: at evaluations GenomeCoder retains of the splicing posterior against for beam search. The same marginal likelihood selects how much generative-parameter updating the reads support, halving the distance to the generating parameters at matched compute. On the published IGoR benchmark, GenomeCoder matches IGoR in complete-scenario recovery at mean rate . Under the same learned generative model, its retained support matches full enumeration recovery while scoring orders of magnitude fewer programs per read.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
programs, genomics, bayesian, inference

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.25 over 4 reviews, above 2% of the others.
Reviewer_guFzOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

This paper presents GenomeCoder, a method for trying to score/identify likely latent variables/explanations (PP) when given observations (RR) for various genetics/genomics-related processes. For example, trying to find the likely exons which are responsible for an observed transcript. The method assumes there is already a mechanistic model which can score p(R∣P)p(R|P), and it attempts to then find p(P∣R)=p(R∣P)p(P)/p(R)p(P|R) = p(R|P)p(P)/p(R), where Z=p(R)Z=p(R) is the normalization constant. The method does this by training a neural network to propose/prioritize a small set of PPs to score, and then applies MAP to find the best PP. The neural network is trained to take in observations RR and produce likely explanations PP, via EM to maximize the final p(R)p(R). This allows giving an estimate of what the missing probablistic support of p(P∣R)p(P|R) is.

Critical strengths

Neural network for prioritization is a nice separation

The neural network only decides what candidate PPs get scored, and the downstream correction for scoring a limited set of PPs is a clever workaround for intractability.

Upfront reporting of experimental results

The results are reported in an honest and clear way. The scope of the paper is also stated, in that evaluation is limited by the mechanistic model which also defines the posterior.

Critical weaknesses

Methodology seems somewhat circular

The method being presented relies heavily on the (known) mechanistic model, p(P)p(R∣P)p(P)p(R|P), both as the source of training data, and also as the evaluator. That is, p(P)p(R∣P)p(P)p(R|P) consists of simulated data which is used to train the neural network which proposes PPs given RR. Then, it is used again to provide the exact scores used for inference, and also defines the Bayes ceiling. It is also used to generate test data. So of course, it makes sense that in this setup, GenomeCoder performs well because it's trained specifically to recapitulate the data exactly. Thus, the reported advantage over baselines (e.g. Kallisto/Salmon) is only in the context of being in-distribution for this specific simulated dataset. Its advantage is likely to shrink once we move to real data.

Benchmarks are limited to simple simulations

Almost all accuracy results in the paper are measured against simple simulations (which the method was trained to match). These simulations are generally quite simple (e.g. at most 3 - 5 isoforms per locus for transcript inference). Furthermore, the setup of the method relies heavily on having a well-defined mechanistic model (as described above). So this method has not been demonstrated on real-world data.

Paper writing is unclear

Much of the main text depends on definitions and details that are hidden in the appendices or are missing. It is difficult to decipher what the methodology is, what each piece is required for (i.e. motivation for each part of the methodology). The paper would certainly benefit from a more structured walkthrough of what the proposed method is. Many of the steps are not clearly defined, and their role in the method is also not clear. Some examples of this are listed in the questions below.

Questions for authors
  • What is x (the context)? Is it ever used in practice (if so, what form does it take on?) or is it just a mathematical concept used for the theorems?
  • What is the goal of the EM algorithm? It seems like it's used to learn p(R∣P)p(R|P) but isn't this given directly from the biological model (i.e. the dataset)?
Reviewer_UzyCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes GenomeCoder, a neural network framework for inferring biological programs. The framework employs a neural network to generate candidate programs, followed by exact mechanistic scoring to construct a restricted posterior distribution. The proposed method is evaluated across several biological inference tasks. While the paper is well-written and clearly organized, the proposed framework lacks sufficient novelty, as it largely follows a conventional two-stage candidate proposal-and-scoring paradigm without substantially proposing and addressing the fundamental challenges of program inference.

Critical strengths
  • Well-written and clearly organized. The paper is well-written and clearly structured, with detailed explanations of the proposed framework , making it easy to follow.
Critical weaknesses

-Lack of novelty. The proposed framework follows a conventional two-stage paradigm, in which a neural network generates candidate programs and a mechanistic model evaluates them. This approach offers limited algorithmic novelty and makes little progress toward addressing the fundamental challenges of program inference.

-Lack of ablation studies. The main text lacks ablation studies to evaluate the individual contributions of the proposed components, including neural propose candidates, and mechanistic rescoring. More comprehensive ablation experiments are needed to demonstrate the necessity and effectiveness of each component.

-Limited performance gain. The proposed method fails to demonstrate substantial performance improvements over existing approaches. For instance, GenomeCoder achieves the same recovery rate as IGoR (0.1330) on V(D)J.

Questions for authors
  • Could the authors clarify the key methodological innovations of GenomeCoder compared with existing neural network inference approaches?
  • Could the authors provide more ablation studies to quantify the individual contributions of the proposed component?
  • Given that GenomeCoder achieves the same recovery rate as IGoR (0.1330) on V(D)J inference, could the authors clarify the practical advantages of the proposed framework over existing approaches?
Reviewer_GLgeOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

GenomeCoder uses learned proposals to discover finite supports of executable biological latent programs. It deduplicates candidates, scores them under a mechanistic model, and normalizes the retained scores. It also uses posterior responsibilities for generative-parameter updates. The evaluation covers V(D)J recombination, isoform inference, small-sample genealogy reconstruction, and a separate supervised structural-variation task (pp. 3-8, Sections 2-5). The strongest evidence concerns support discovery and exact retained-mass accounting under specified models. The reported external V(D)J recovery averages 0.1330 for both GenomeCoder and IGoR across five fresh-recognizer splits (p. 18, Table 8).

Critical strengths
  1. The paper makes support truncation measurable. Exact retained scores give TV error equal to omitted posterior mass under the specified model. The V(D)J normalizer and splicing regrouping provide useful ways to quantify that mass. Linear interpolation of the supplied summaries reproduces coverage at 7,000 evaluations: 0.953 versus 0.906 for beam and 0.779 for prior sampling. The supplement's separate comparison using only measured points within that budget favors beam, 0.904 versus 0.806. GenomeCoder's next measured point is at about 7,020 evaluations. These are different budget summaries, not contradictory measurements, and Figure 8 explicitly states that it uses interpolation (pp. 3-5, Eqs. 3-6; p. 19, Figure 8).

  2. The evaluation includes meaningful external and representation controls. The fresh-fit IGoR comparison uses five held-out splits with the same 400-read denominator. A matched-capacity SV control improves conditional multiset recovery from 0.245 to 0.617, although detection F1 falls from 0.546 to 0.467. The recovery rates use each model's own matched calls, 200 and 214, rather than identical detected variants. Splicing and small-n ARG results add useful breadth (p. 18, Table 8; p. 21, Table 12; p. 7, Section 5.1). These are scoped benefits, not evidence of universal superiority.

Critical weaknesses
  1. The methodological novelty needs a clearer comparison with close peer-reviewed prior work. Memoised Wake-Sleep (Hewitt et al., UAI 2020) discovers, stores, and reuses latent programs with a recognition network while learning a generative model. It is not discussed or compared here. Inference compilation (Le et al., AISTATS 2017), which is cited, already learns proposals for model-based correction. This does not make the algorithms identical. However, the conditional-posterior identities alone do not establish the new contribution. The paper should distinguish its biological support construction and exact normalizers from these existing inference ideas (pp. 4-6, 9, 13, 25).

  2. Central learning and validation claims are not consistently supported. Sections 3.6, 5.3, and 6 present the +0.2568 chr7 splicing result as a sequence-conditioned prior learned from posterior responsibilities without isoform labels. Appendix K.2 instead attributes this number to a fixed conversion of pretrained SpliceAI scores. A conversion hyperparameter was selected by validation usage cross-entropy on training chromosomes. K.2 reports no E-step, M-step, or learning trajectory for this result (pp. 6, 8-9, 34-35, Tables 28-29). K.2 does not show that test-locus isoform labels were used. The problem is that the reported number does not demonstrate responsibility-based prior learning. Separately, the five-split TV mean labeled 0.1824 is actually the mean of three split-0 recognizers; the supplied five-split records average 0.186413. The fresh-fit confidence interval uses the normal critical value, not the t4 convention stated for Table 6. A t4 interval would widen it from approximately +/-0.0155 to +/-0.02195, and the Limitations paragraph itself quotes +/-0.022 (pp. 9, 15, 17-18). Missing current runners prevent resolving these discrepancies from the supplement. The paper also does not report how often the 4-million-program enumeration cap binds under the learned models (p. 16, Appendix C). These problems materially weaken the evidence supporting the recommendation.

Questions for authors
  1. Which unlabeled splicing E-step and M-step produced the +0.2568 result? Please distinguish that experiment from the SpliceAI prior conversion in Tables 28-29 and identify the supervision used for fitting and selection.
  2. What differs algorithmically from Memoised Wake-Sleep? Can existing matched-budget results isolate the benefit of the proposed support construction?
  3. Please reconcile Table 4/Figure 7's 0.1824 aggregation, the two confidence-interval conventions, and the fresh-fit artifacts. For the same-model enumeration comparison, how often does the 4-million-program cap bind under the learned models, and what mass is omitted?
  4. Appendix I reports a 3,141-second end-to-end run versus 154 seconds for IGoR and a separately measured 10.6-fold proposer speedup. What is the complete end-to-end cost with the batched proposer at matched recovery? Does the evaluation-count advantage persist when total runtime is considered?

I would reconsider the current rating if the authors substantiate the claimed splicing learning procedure and reconcile the numerical discrepancies. Evidence that also clarifies the methodological contribution and practical costs could justify a higher score.

Reviewer_oxnaOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduces GenomeCoder, which models biological processes (e.g., V(D)J recombination scenarios and isoform sets) as executable programs, scored by a known mechanistic model. Across V(D)J and splicing inference tasks, the authors report performance parity with IGoR.

Critical strengths
  1. The paper covers several domains with domain-appropriate references (IGoR, Salmon, ARGweaver)
Critical weaknesses
  1. GenomeCoder is slower than IGoR at equal accuracy. The efficiency claims are not demonstrated.
  2. GenomeCoder matches IGoR in V(D)J, ranks second behind Salmon in transcript quantification, and is less accurate than ARGweaver in genealogy inference. The performance benefit also seems limited.
  3. The method closely relates to inference compilation, wake-sleep, and DreamCoder-style recognition models.
Questions for authors
  1. What is the concrete technical contribution relative to the methods mentioned in the weaknesses section?
  2. In a practical setting, why would someone use GenomeCoder over other methods despite being slower?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.