Understanding the Noncoding Genome: Scale, Variant Effects, and Causal Identification in Long-Range Regulatory Models

1. Long-range context changes the regulatory scope a model can see

Enformer improves predictions from DNA to expression and chromatin tracks by integrating distal information; its comparisons address input span and the enhancer distances that can be integrated. AlphaGenome further combines high-resolution outputs across multiple modalities. What matters here is not that longer inputs are always better, but whether the information needed for the task falls within the available window, whether the network actually integrates it, and whether it reaches the output at sufficient resolution. Inputs of the same length can correspond to different effective receptive fields; the same variant may also produce different effects through local splicing and distal regulation, which cannot be replaced by a single overall track score.[2] Effective gene expression prediction from sequence by i[1] Advancing regulatory variant effect prediction with Alp

2. Multitrack learning places heterogeneous experiments in a shared representation

The original AlphaGenome paper describes 11 modality classes and thousands of human and mouse tracks covering expression, splicing, and chromatin-related observations. This joint prediction provides a shared representation for conditional analysis, but experiments differ in measurement noise, tissue origin, and interpretable units. This article recommends storing each track's samples, normalization, resolution, and replicate agreement in a task dictionary. Strong performance on many easily predicted tracks does not establish equal reliability for underrepresented tissues or particular splicing events; evaluation must preserve within-modality distributions rather than report only a single average.[1] Advancing regulatory variant effect prediction with Alp

3. Allelic contrasts impose stronger demands than track fitting

A sequence model can receive reference and alternate sequences separately, after which predictions for an output window or gene can be compared. Such contrasts help formulate mechanistic hypotheses, but errors may be amplified by differencing or appear to cancel because of shared bias. High correlation in expression across genes often reflects baseline differences between genes and does not imply accurate ranking of multiple variants within the same gene. Enformer's eQTL analysis also discusses validation difficulties arising from linkage disequilibrium. We therefore separate baseline tracks, within-gene allelic contrasts, effect direction, and effect-size calibration into four endpoints, each evaluated independently.[2] Effective gene expression prediction from sequence by i[1] Advancing regulatory variant effect prediction with Alp

4. Derivation in this article: the same regulatory increment has different effects at different baselines

Suppose a regulatory variable z, used solely for analysis, determines a normalized output through a sigmoid, and a variant increases z by 1. At baseline z=0, the output changes from 0.5 to approximately 0.731, a difference of 0.231; at baseline z=3, it changes from approximately 0.953 to 0.982, a difference of only about 0.029. This background dependence caused by saturation shows why a locus's local contribution cannot be detached from its tissue or sequence background and treated as a universal fixed effect. No biological parameters are fitted in this example; it simply demonstrates that equal input contributions do not guarantee equal changes on the observation scale. The figure displays both the absolute difference and the local derivative to facilitate examination of the source of nonlinearity.作者分析与复核建议 / Author analysis

5. Derivation in this article: paired variants require testing the nonadditive term

Constructing a reference sequence, a single substitution A, a single substitution B, and a double substitution AB allows a double difference to isolate the interaction term relative to the chosen output scale. Even without an explicit interaction inside the model, a nonlinear output can produce a nonzero double difference, so any claim of discovering an interaction must specify the measurement scale. Conversely, zero interaction in the output does not establish that the molecular mechanisms are independent. This article recommends retaining raw tracks, normalized values, and contrasts together, without selecting the most significant transformation scale after the fact. Genuine mechanistic inference also requires interventions capable of ruling out alternative pathways, rather than simply observing model responses.作者分析与复核建议 / Author analysis

6. Association, intervention, and transferability cannot share a single label

Research on causal representation emphasizes that stable patterns in observational distributions do not automatically determine intervention outcomes. In genomic tasks, linkage, cell composition, and tissue state can all alter observed relationships. This article calls computational sequence substitutions within a model 'within-model perturbations'; only measurements generated by external interventions constitute the corresponding experimental support. If a locus changes expression in one tissue, transferring that finding to another also requires considering whether the regulatory background is preserved. Models can narrow the set of hypotheses worth testing, but attention, gradients, or virtual deletions alone cannot establish a causal pathway.[3] Towards Causal Representation Learning[2] Effective gene expression prediction from sequence by i

7. Derivation in this article: repeated measurements provide a reference for predictive correlation

Suppose the observation y is the true signal s plus independent, zero-mean noise, and the predictor recovers s exactly. The correlation between the prediction and y then equals the square root of the fraction of total variance attributable to signal. This ideal relationship shows that observational reliability across tracks affects attainable correlation; it also shows why a theoretical upper bound alone cannot excuse a poorly performing model. Independent noise, stable variance, and recovery of the true signal are all strong assumptions. In practice, reliability should be estimated from replicate data, followed by checks stratified by tissue and locus; the figure varies only the assumed noise variance and does not simulate any real experiment or model.作者分析与复核建议 / Author analysis

8. Test splits should guard against leakage through homology, proximity, and sample origin

Genomes contain repetitive sequences, homologous genes, neighboring regulatory regions, and similar cellular origins, so random window splitting may leave training and testing with extensive shared information. This article proposes three levels of holdout: locus holdout to evaluate sequence extrapolation, tissue or sample holdout to evaluate transfer across conditions, and temporal holdout to evaluate new experiments. Normalization, feature selection, and thresholds must be determined only on training or validation sets. Comparisons between models should use the same samples and loci, with intervals calculated by clustering on independent loci; treating millions of adjacent bases as independent samples produces overly precise statistical significance.作者分析与复核建议 / Author analysis

9. A research agenda worth testing

A more valuable next step is to test whether joint multimodal learning improves allelic effects across tissues, rather than merely improving track fitting. New loci, single variants, and combined variants can be specified in advance, with direction, magnitude, and cross-modal consistency tested separately; external observations should be collected only after the model is frozen. If a joint model leads only in baseline expression but does not improve contrasts, conclusions should be limited to representation and track prediction. If different modalities yield conflicting effects, those conflicts should also be retained and investigated for tissue, measurement, or mechanistic explanations, rather than selecting a track that supports the original hypothesis.作者分析与复核建议 / Author analysis

Technical supplement A · From local convolution to distal information integration

Convolutional layers progressively expand local receptive fields, while attention can directly exchange distal information between already compressed representations. Their tradeoffs involve resolution, distance range, and computational cost: downsampling too early may lose local splicing or motif details, whereas preserving full base-level resolution increases the cost of long-range processing. This article recommends masking proximal and distal context separately for the same variant, holding the window and normalization fixed, and observing prediction changes. This tests which information the model relies on; it still does not prove that those sequence regions necessarily have causal effects in real tissues. Input masking may itself also create sequences outside the training distribution.[2] Effective gene expression prediction from sequence by i[1] Advancing regulatory variant effect prediction with Alp

Technical supplement B · Contrast error depends on the correlation between the two predictions

Let the reference and alternate prediction errors be er and ea, respectively. The allelic contrast error is then ea−er, whose variance includes the covariance between the two errors. Shared biases may cancel in the contrast, while uncorrelated noise may accumulate. There is therefore no fixed ratio between track error and contrast error. This article proposes retaining both absolute predictions and paired comparisons in independent experiments to examine whether large baseline errors coexist with useful contrasts, and vice versa. When comparing models, a single absolute-track correlation cannot establish better variant prioritization; direct evaluation and calibration against contrast endpoints are required.作者分析与复核建议 / Author analysis

Technical supplement C · Agreement in direction can coexist with severely incorrect magnitudes

If all true variant effects are positive while a model consistently predicts tiny positive values, directional accuracy can be high even though effect sizes offer almost no practical discrimination. Conversely, for near-zero effects with substantial measurement noise, signs can readily flip even when the model is already close to the measurable limit. This article recommends defining meaningful effect regions in advance and reporting sign, ranking, magnitude, and indeterminate intervals together. When selecting cases with strong effects, thresholds should be fixed on the validation set; selecting a directionally concordant subset after testing changes the question being answered and introduces selection bias.作者分析与复核建议 / Author analysis

Technical supplement D · The complementarity of linkage and intervention data

Population associations can exploit large numbers of natural variants but face shared signals among correlated loci. Controlled interventions come closer to changing a specific position, yet remain constrained by background and measurement conditions. This article argues that the two should be treated as complementary constraints: first use models to propose distinguishable hypotheses about loci and tissues, then examine them in appropriately independent data. If different evidence points to different targets, preserve that conflict and examine the assembly, tissue, haplotype, and label definitions. Compressing all sources into a composite score loses conflict information, even though such conflicts may identify regulatory conditions the model has not yet captured.作者分析与复核建议 / Author analysis

10. Conclusion: turning long-sequence capabilities into cumulative biological knowledge

Sequence models are compressing statistical regularities from many heterogeneous experiments into queryable functions, whose value lies in providing systematic hypotheses about regulatory mechanisms. This article advocates expanding outputs from 'a variant score' to traceable conditional contrasts: which sequence background, which tissue, which measurement, and which effect scale. Evidence from the original studies and analyses in this article should be stored separately, with negative results also linked to the correct versions. Only when model predictions withstand tests using new samples, tissues, and interventions can engineering advances in long-range modeling become reliable mechanistic knowledge, rather than remain demonstrations of large-scale correlation.作者分析与复核建议 / Author analysis

Core equations and derivations

Methods and evidence workflow

Analytical figures and numerical examples

Recomputable worked example

At baseline z=0, the increment is 0.231059, and at z=3, it is approximately 0.029440; when noise variance equals signal variance, the ideal correlation is sqrt(1/2)≈0.707107.

Method comparison

Verification and experimental plan

1. Freeze the reference assembly, sequences, variant representation, model, and track dictionary.
2. Check strand orientation, coordinates, reference alleles, and window boundaries; reject mismatched inputs.
3. Separate training and testing by locus, homologous region, tissue, and sample origin.
4. Evaluate tracks, contrast direction, effect sizes, and combined interactions independently.
5. Use only training data to determine normalization, gene or window selection, and tuning rules.
6. Report intervals using independent loci as statistical clusters, broken down by tissue and modality.
7. Record within-model perturbations and external intervention measurements separately, and define falsification conditions in advance.
8. Publish source and input summaries, computational code, and negative results; do not use predictions for personal variant diagnosis.

Scope and limitations

• The analytical sigmoid and noise model were not fitted to real genomic data.
• This article did not run AlphaGenome or Enformer and conducted no new molecular experiments.
• Multitrack prediction does not mean every possible causal pathway has been observed.
• Assumptions such as independent noise and a fixed background may not hold in real tissues.
• Research interpretation and individual clinical judgment are different tasks.

Evidence from the original studies and its scope

Research questions worth testing

1. Does combining modalities provide an independent gain in allelic contrasts in new tissues?
2. How does the effect scale change the assessment of interactions among combined variants?
3. At the same measurement reliability, which distance ranges still exhibit systematic errors?

Primary sources and corresponding claims

1. Advancing regulatory variant effect prediction with AlphaGenomeAvsec et al. · 2026-01-28Abstract; sequence-to-track prediction, 11 modalities and variant-effect benchmark scope.
2. Effective gene expression prediction from sequence by integrating long-range interactionsAvsec et al. · 2021Results: long-range regulation; eQTL linkage disequilibrium and task-specific validation.
3. Towards Causal Representation LearningBernhard Schölkopf, Francesco Locatello, Stefan Bauer, Nan Rosemary Ke, Nal Kalchbrenner, Anirudh Goyal, Yoshua Bengio · 2021-02-22Abstract and sections on interventions, independent mechanisms and causal representations