acceptodds
Under review as a conference paper at ICLR 2027

PATHCAST: PATHWAY-STRUCTURED GENE EXPRESSION FORECASTING IN CARDIAC ORGANOIDS

Abstract

Forecasting gene expression during cardiac organoid development can support experimental planning and the study of developmental programs. However, sparse, irregularly sampled RNA measurements provide limited evidence for the coordinated expression changes that emerge at later stages. We introduce PATHCAST, a pathway-structured Transformer that integrates biological pathway memberships and gene descriptions to forecast group-level expression beyond the observed RNA history. The central design couples complementary representations of pathway dynamics: a history branch uses temporal and cross-pathway differential attention, while a state branch summarizes reconstructable pathway states and recent trends. Their fusion predicts continuous-time state increments, which a pathway-constrained dynamic decoder translates into future expression changes with explicit pathway and gene-level contributions. Across within-condition and reference-assisted forecasting on three cardiac organoid datasets and one hepatic differentiation dataset, PATHCAST achieves the lowest overall mean squared error among the reported baselines. Ablations assess the contributions of pathway structure and complementary modeling components. Pathway analyses and uncertainty-based risk ranking characterize forecast behavior. Together, these findings support pathway-structured representations for interpretable forecasting from limited transcriptomic observations.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Gene expression forecasting, pathway-informed modeling, transcriptomics

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 3 reviews, above 27% of the others.
Reviewer_Edd2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PATHCAST, a pathway-structured Transformer designed to forecast future group-level gene expression in cardiac organoids from sparse, irregularly sampled bulk RNA-seq histories. The model takes historical expression profiles alongside external priors (specifically gene text descriptions from NCBI, UniProt, and GO encoded via SapBERT, and pathway membership masks from Hallmark and Reactome) and feeds them into a dual-branch architecture consisting of a history branch using differential Transformers and a state branch projecting pathway states. The fused representations predict continuous-time state increments, which are decoded into additive pathway, auxiliary, and gene-level contributions. Experiments on three cardiac organoid datasets and one hepatic differentiation dataset show lower MSE compared to several general time-series and trajectory baselines.

Critical strengths
  1. The targeted problem of forecasting future transcriptomic profiles from short sampled histories is practical for experimental planning in organoid development. Decomposing the predicted expression changes into additive pathway, auxiliary, and gene-level terms provides a way to inspect which biological modules drive the model's outputs.

  2. The empirical evaluation covers multiple scenarios, including within-condition forecasting, reference-assisted prediction, and cross-study transfer. Incorporating an uncertainty head for risk ranking is also a useful addition for identifying reliable gene predictions in downstream analysis.

Critical weaknesses
  1. The core methodological novelty and motivation are unclear, and the baseline comparisons appear unfair. Rather than offering a fundamental algorithmic advance, the architecture reads like an over-engineered assembly of existing modules. Moreover, PATHCAST relies heavily on external biological priors and engineered historical summary features that standard time-series baselines do not have access to.

  2. The presentation and visual design of the figures need substantial improvement. Figure 1 and especially Figure 2 are overly crowded and very difficult to read. The excessive use of different colors, nested boxes, and dense mathematical notations in Figure 2 is visually overwhelming and easily causes misunderstanding about how tensors actually flow between the shared inputs, the history branch, and the state branch. In addition, while the whole paper emphasizes pathway-structured interpretability, the cross-study transfer experiment abruptly switches to pretraining a "direct-gene variant" rather than the main pathway-structured model, which undermines the consistency of the core story.

Questions for authors
  1. Since PATHCAST takes rich external priors and hand-crafted historical statistics, how do the authors ensure a fair comparison with general time-series baselines? If baselines like iTransformer are also provided with the same hand-crafted slope/mean features and text embeddings, does PATHCAST still hold a meaningful advantage?

  2. In Table 5, removing the state branch, sustained trend, gene correction, or text modules results in very tiny MSE differences. How do the authors justify the heavy architectural and hyperparameter complexity given such marginal gains? Furthermore, why was the "direct-gene variant" used for cross-study transfer instead of the proposed pathway-structured model?

2 comments

We appreciate the time you took to review our manuscript and provide detailed feedback. Our point-by-point responses follow.

W1 / Q1: Methodological contribution and fairness of baseline comparisons

Motivation and contribution. PATHCAST addresses forecasting from sparse, irregularly sampled RNA histories with many correlated genes and few observed time points. Its methodological contribution is a pathway-structured increment model: observed expression is jointly projected into reconstructable pathway states, history and recent state changes condition continuous-time forecasts, and membership-constrained decoding maps state increments into explicit pathway contributions to gene-expression changes. Auxiliary and gene-level terms accommodate variation outside the pathway representation. This connects the representation, temporal prediction, and interpretation within the forecast computation.

The ablation evidence supporting these design choices is discussed in W2/Q2 below.

Fairness of information access. Historical means and slopes are computed from the same observed expression available to the baselines; their explicit encoding can nevertheless affect performance. Pathway annotations and gene descriptions provide additional biological information. To test the effect of historical statistics and text access, we supplied iTransformer with the same gene-level historical statistics and fixed gene-text embeddings, projected into its gene representations. A separate matched raw-input control uses the same RNA input adapter and training configuration as the augmented variant.

The feature-augmented iTransformer and PATHCAST rows below report mean ± standard deviation across three runs on EB-hCO.

MethodMSE ↓MAE ↓Top50 MSE ↓r ↑
iTransformer: original paper setting0.0502 ± 0.00550.1425 ± 0.00800.1175 ± 0.00600.9494 ± 0.0061
iTransformer + historical statistics + gene-text0.1060 ± 0.01220.2315 ± 0.01160.4159 ± 0.10120.8879 ± 0.0132
PATHCAST0.0322 ± 0.00120.1117 ± 0.00250.1275 ± 0.01700.9680 ± 0.0016

PATHCAST outperforms the feature-augmented iTransformer configuration. Providing historical statistics and text therefore does not automatically improve forecasting.

W2 / Q2: Architectural complexity, figure clarity, and the direct-gene transfer variant

Architectural complexity. Table 5 shows that the components contribute at different scales and across different metrics. Removing history conditioning increases MSE from 0.0322 to 0.0924, while removing auxiliary forecast outputs increases it to 0.0472. Other components have benefits that are more apparent beyond overall MSE: removing the sustained trend and gene-level correction increases Top50 MSE by 19.4% and 18.4%, respectively. Overall MSE alone therefore gives an incomplete picture of these components’ contributions. We will clarify their respective roles and the associated hyperparameter settings.

Figure clarity. We will simplify Figures 1 and 2, explicitly labeling the shared inputs and the connections between the history branch, state branch, and output reconstruction.

Cross-study transfer. Table 3 tests whether cross-study pretraining improves forecasting when the target study provides limited observations. The direct-gene variant provides a common gene-level representation for this experiment, separating the transfer question from the pathway-specific representation. It retains PATHCAST’s temporal modeling and observation-anchored forecasting. All four strategies use the same architecture, gene panel, observation window, and forecast dates, so their comparison isolates the effects of pretraining and adaptation within this variant. Table 3 thus complements the pathway-based evaluations with a controlled study of transfer. We will make the variant and the scope of this conclusion explicit in the Table 3 caption.

Thank you for the rebuttal, which addresses some of my concerns. However, my main concerns about methodological novelty and the adequacy of the baseline comparisons remain. Considering other reviewers’ comments, I am inclined to maintain my original score.

Reviewer_SxgoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes PATHCAST, a pathway-structured Transformer for forecasting group-level bulk RNA-seq profiles from short, irregularly sampled histories. It combines pathway membership and gene-description embeddings with a history branch, a reconstructable pathway-state branch, and a continuous-time decoder that produces pathway, auxiliary, and gene-level contributions. Evaluations on cardiac organoid and hepatic differentiation datasets report strong MSE results against several forecasting baselines, alongside ablations, cross-study adaptation, and uncertainty-based risk ranking. The authors clarify that pathway contributions describe model behavior rather than biological causality.

Critical strengths

1 PATHCAST provides an explicit pathway-level decomposition of gene-expression forecasts. The ablations suggest that pathway modeling and temporal history both contribute to its performance.

2 The paper compares against diverse baselines across several public datasets. PATHCAST achieves the lowest reported MSE in the within-condition and reference-assisted tables, and on the hepatic dataset.

Critical weaknesses

1 The evaluation covers a small number of condition-level trajectories and only a few held-out future time points. The paper reports averages over three random seeds but not variability for the main benchmark scores; it also does not clearly state how many independent groups or biological replicates contribute to each result. Because genes within a profile are correlated, gene-wise errors do not by themselves establish robust generalization.

2 Table 2 relies on reference-condition profiles at forecast times, but the method section and architecture figure do not explain how these profiles enter PATHCAST or the baselines. This is central to interpreting the task and assessing whether the comparison is reproducible and fair.

Questions for authors

For Table 2, what exact future-time reference profiles are available to each method, and how are they incorporated into PATHCAST and the baselines?

How many independent groups or biological replicates contribute to each evaluation? Can you report the per-seed results for the main benchmark metrics and explain how replicate profiles are aggregated?

3 comments

Thank you for your careful review and detailed feedback. We respond to each of your comments below.

W1 / Q2: Evaluation scope, replication, and variability

Evaluation scope. Our benchmark evaluates three forecasting tasks across four studies: EB-hCO, Cardioids, cVO, and H1-Hep. These studies cover cardiac organoid development and hepatic differentiation. The six condition-level time courses test temporal extrapolation, reference-assisted condition prediction, and cross-study transfer. Each setting uses a fixed split between observation and forecasting, with two to four held-out future dates, making the evaluation scope explicit.

Replication and aggregation. EB-hCO, Cardioids, and cVO use three replicate expression profiles per condition and time point. H1-Hep provides one expression profile per time point. For analyses based on replicate means, we use expression profiles averaged across the three replicates within each date.

Run-to-run variability. To enable a direct assessment of performance variability, we report PATHCAST’s individual results from three runs for Tables 1–4 below, using head adaptation for Table 3. We have also added the corresponding means and standard deviations to Appendix G, Tables G1–G4, complementing the individual-run results with an aggregate summary.

Table 1: Within-condition forecasting.

SettingRunMSE ↓MAE ↓Top50 MSE ↓Pearson r ↑
EB-hCORun 10.033050.113530.141980.96634
EB-hCORun 20.032760.112750.131730.96815
EB-hCORun 30.030820.108790.108760.96962
CHIR4Run 10.028380.091720.098010.96679
CHIR4Run 20.028270.091330.100780.96670
CHIR4Run 30.028020.090710.097740.96697
CHIR8Run 10.042630.118700.096890.95481
CHIR8Run 20.042230.118040.094220.95504
CHIR8Run 30.042580.118660.097150.95494

Table 2: Reference-assisted forecasting.

SettingRunMSE ↓MAE ↓Top50 MSE ↓Pearson r ↑
cVO: Unvasc → VascRun 10.050920.125850.237740.95222
cVO: Unvasc → VascRun 20.050940.125810.240000.95182
cVO: Unvasc → VascRun 30.050930.125790.238890.95209
cVO: Vasc → UnvascRun 10.050450.124800.237860.94969
cVO: Vasc → UnvascRun 20.050510.125030.238140.94974
cVO: Vasc → UnvascRun 30.050500.125010.236810.94973
Cardioids: CHIR4 → CHIR8Run 10.033660.107950.101180.96667
Cardioids: CHIR4 → CHIR8Run 20.033720.108140.100940.96674
Cardioids: CHIR4 → CHIR8Run 30.033950.108320.101970.96664
Cardioids: CHIR8 → CHIR4Run 10.033530.107630.105030.95985
Cardioids: CHIR8 → CHIR4Run 20.033550.107630.104390.95980
Cardioids: CHIR8 → CHIR4Run 30.033410.107490.105100.95997

Table 3: Cross-study transfer, head adaptation.

SettingRunMSE ↓MAE ↓Top50 MSE ↓Pearson r ↑
D11 historyRun 10.031260.116140.069980.96561
D11 historyRun 20.030620.114990.068780.96713
D11 historyRun 30.031140.116240.064180.96536
D7 historyRun 10.091030.212070.222450.89429
D7 historyRun 20.092920.215850.222070.89226
D7 historyRun 30.089870.212280.209290.89626

Table 4: H1-Hep forecasting.

SettingRunMSE ↓MAE ↓Top50 MSE ↓Pearson r ↑
H1-HepRun 10.050920.151570.336370.94990
H1-HepRun 20.042430.139650.193210.95650
H1-HepRun 30.047220.150190.267970.95236

W2 / Q1: Future-time reference profiles and their incorporation

For Table 2, PATHCAST and all baselines receive the same known reference expression at the forecast dates, while the target condition provides only early observations. The reference profiles come from Unvasc./Vasc. for cVO and CHIR4/CHIR8 for Cardioids. Reverse prediction exchanges the reference and target conditions.

PATHCAST uses historical target–reference expression differences to predict an increment to the latest observed condition difference. The updated difference is then added to the forecast-date reference expression to reconstruct the target prediction. Future reference profiles therefore enter only at output reconstruction.

The baselines likewise forecast future condition differences and add back the same reference values. All methods are evaluated on the reconstructed target expression, ensuring a common evaluation target under the same available reference information.

We will make this information flow explicit in the method section and architecture figure, and document the reference dates and baseline adaptations in the appendix.

Thank you for the response. After reviewing, I decide to keep my scores.

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

PATHCAST combines expression history, pathway states, and recent trends to forecast future expression changes. Its additive outputs support inspection of pathway and gene-level contributions, and the overall MSE results are promising across four datasets.

Critical strengths

The model shares trajectory parameters across forecast horizons and enforces zero increments at zero horizon. This provides a useful structural constraint when temporal observations are limited. Pathway contributions are explicit terms in the prediction, supported by output-removal tests, matched gene-set controls, and checks across random seeds. Overall MSE improves across the reported datasets.

Critical weaknesses
  1. The reference-assisted prediction procedure is insufficiently specified. The paper does not explicitly define the task-specific input and base profile in Equation 7, or explain how each baseline receives the reference measurements. These choices determine whether the model predicts full expression or differences from a known reference trajectory. The adaptations of single-cell methods to bulk expression forecasting also need explanation.
  2. The evaluation does not establish the gain over preserving the latest expression profile or following the reference trajectory with a fixed early condition offset. Last-velocity is the only simple baseline. The learned-baseline comparisons and ablations are encouraging, but do not answer this question. Moreover, the overall MSE advantage does not extend to EB-hCO's Top-50 subset, where MSE is 0.1275 versus 0.1012 for scGPT-temporal.
  3. The direct-gene ablation changes both biological membership and the grouping and parameter sharing introduced by the pathway architecture. It therefore does not isolate the value of biologically correct pathways. The matched output-removal controls are useful, but they do not replace retraining with alternative groupings under the same architecture.
Questions for authors
  1. Please define the inputs, base profile, and prediction target for each setting. How does each baseline use the permitted reference measurements, and how are single-cell methods adapted to bulk expression profiles?
  2. How does PATHCAST compare with last-value prediction and reference prediction with a fixed early condition offset? Please also report the correlation between predicted and observed changes relative to the last observation.
  3. Does PATHCAST retain its advantage over a retrained control with shuffled pathway memberships, while matching group sizes, overlap, and model architecture? This would clarify the contribution of biological pathway knowledge.
2 comments

Thank you for taking the time to carefully review our work and for the detailed comments. We address your concerns point by point below:

W1 / Q1: Forecasting setup and baseline adaptations

Table 2 evaluates reference-assisted forecasting. PATHCAST and the baselines use the same forecast-date reference values, and all reported metrics are computed on the reconstructed target expression.

Prediction setup.

  • Within-condition forecasting (Tables 1 and 4): PATHCAST takes observed expression history as input. Equation (7) uses the latest observed expression as the base profile, to which the model adds predicted expression changes.
  • Cross-study transfer (Table 3): PATHCAST takes the target study's history, averaged across replicates at each date. It adds predicted changes to the latest observed profile to forecast future expression.
  • Reference-assisted forecasting (Table 2): The target condition provides only early observations. PATHCAST takes historical target-minus-reference expression differences matched by recorded group as input. The known base in Equation (7) is the forecast-date reference expression plus the latest observed condition difference. The model predicts the subsequent change in this difference and adds it to that base. Forecast-date reference expression enters only at output reconstruction.

The reference conditions and forecast dates are listed below. Reverse prediction exchanges the reference and target conditions:

StudyReference conditionTarget conditionForecast dates
cVOUnvasc.Vasc.D10, D12, D16
cVOVasc.Unvasc.D10, D12, D16
CardioidsCHIR4CHIR8D6.5, D7.5, D10
CardioidsCHIR8CHIR4D6.5, D7.5, D10

Baseline adaptations.

Each Table 2 baseline predicts future condition differences and adds back the reference values, converting predictions to gene-expression units where necessary. The baseline implementations are summarized below:

  • Last-velocity: Simple extrapolation from the latest observed profile, with the slope estimated from the two latest observations.
  • iTransformer: For within-condition forecasting, fixed-length windows of observed gene expression, together with observation-time and forecast-horizon covariates, are used to directly predict expression at each target date.
  • ReIMTS–GraFITi and ASTGI: PCA is fitted only to historical profiles, and forecasts are inverse-transformed to gene-expression units.
  • Time-MoE: Only the observed history is interpolated onto a 0.5-day grid before extrapolation.

The single-cell methods are adapted to historical bulk expression profiles as follows:

  • scGPT-temporal: The pretrained encoder and added temporal module and decoder are fine-tuned on historical bulk profiles; in within-condition forecasting, the decoder directly predicts future expression. In reference-assisted forecasting, the latest target and reference profiles are encoded separately; their embedding difference conditions the temporal module, and the decoder predicts the future condition difference, which is added to the forecast-date reference expression.
  • MIOFlow: PCA and latent dynamics are fitted to historical bulk profiles. The learned ODE extrapolates from the last observed profile, followed by inverse transformation to gene space.
  • ARTEMIS: The VAE and bridge model are trained on historical bulk profiles. As an additional extrapolation adaptation for this task, the score-network time input is fixed at the historical endpoint during SDE integration, and 16 decoded samples are averaged.
  • scMix: Historical bulk profiles form empirical distributions for training-time resampling, and prediction uses up to the four most recent dates. Resampling does not increase the number of independent biological samples.

Manuscript revisions. We will make this protocol explicit in Sections 3.1, 3.5, and 4.2.2 and the Figure 2 caption, with reference dates and baseline adaptations documented in the appendix.

W2 / Q2: Simple baselines and prediction dynamics

The added baselines test whether PATHCAST improves upon preserving the latest expression or following the reference trajectory with a fixed early offset. For reference-assisted forecasting, we also evaluate whether its predicted expression changes agree with the observed changes.

Last observation predicts the latest observed expression at every future date. Reference + fixed early offset adds each gene's average early target-minus-reference difference to the reference expression at each forecast date.

Results. PATHCAST results are reported as mean ± standard deviation across three random seeds. The simple baselines are deterministic. Bold indicates the better mean within each setting; r denotes the Pearson correlation of expression values. Table 3 reports head adaptation.

Table 1: Within-condition forecasting

DatasetMethodMSE ↓MAE ↓Top50 MSE ↓r ↑
EB-hCOLast observation0.042020.129840.110910.95702
EB-hCOPATHCAST0.0322 ± 0.00120.1117 ± 0.00250.1275 ± 0.01700.9680 ± 0.0016
CHIR4Last observation0.032970.096340.109590.96098
CHIR4PATHCAST0.0282 ± 0.00020.0913 ± 0.00050.0988 ± 0.00170.9668 ± 0.0001
CHIR8Last observation0.044180.119270.111280.95234
CHIR8PATHCAST0.0425 ± 0.00020.1185 ± 0.00040.0961 ± 0.00160.9549 ± 0.0001

Table 2: Reference-assisted forecasting — cVO

For reference-assisted forecasting, Δr measures the Pearson correlation between predicted and observed changes relative to the latest target observation.

Direction (reference → target)MethodMSE ↓MAE ↓Top50 MSE ↓r ↑Δr ↑
Unvasc. → Vasc.Reference + fixed early offset0.074770.158060.498100.926080.3107
Unvasc. → Vasc.PATHCAST0.05093 ± 0.000010.12582 ± 0.000030.23888 ± 0.001130.95204 ± 0.000210.5410 ± 0.0016
Vasc. → Unvasc.Reference + fixed early offset0.073610.157310.480840.925610.4329
Vasc. → Unvasc.PATHCAST0.05049 ± 0.000030.12495 ± 0.000130.23760 ± 0.000700.94972 ± 0.000030.5388 ± 0.0002

Table 2: Reference-assisted forecasting — Cardioids

Direction (reference → target)MethodMSE ↓MAE ↓Top50 MSE ↓r ↑Δr ↑
CHIR4 → CHIR8Reference + fixed early offset0.067650.163150.315700.924310.2173
CHIR4 → CHIR8PATHCAST0.03378 ± 0.000150.10814 ± 0.000180.10136 ± 0.000540.96669 ± 0.000050.5437 ± 0.0007
CHIR8 → CHIR4Reference + fixed early offset0.068020.163290.321720.918900.5129
CHIR8 → CHIR4PATHCAST0.03350 ± 0.000070.10758 ± 0.000080.10484 ± 0.000390.95987 ± 0.000080.5414 ± 0.0012

Table 3: Cross-study adaptation to EB-hCO

HistoryMethodMSE ↓MAE ↓Top50 MSE ↓r ↑
D11 historyLast observation0.035990.120250.093800.96039
D11 historyPATHCAST (Head adaptation)0.0310 ± 0.00030.1158 ± 0.00070.0676 ± 0.00310.9660 ± 0.0010
D7 historyLast observation0.133500.258550.343760.85052
D7 historyPATHCAST (Head adaptation)0.0913 ± 0.00150.2134 ± 0.00210.2179 ± 0.00750.8943 ± 0.0020

Table 4: Forecasting on H1-Hep

MethodMSE ↓MAE ↓Top50 MSE ↓r ↑
Last observation0.057690.155010.130340.94172
PATHCAST0.0469 ± 0.00430.1471 ± 0.00650.2659 ± 0.07160.9529 ± 0.0033

PATHCAST achieves higher expression prediction accuracy than the fixed-offset reference baseline across all four reference-assisted settings.

W3 / Q3: Retrained controls with shuffled pathway memberships

To test the contribution of biological pathway knowledge, we shuffled gene identities in the pathway membership matrix and retrained PATHCAST on EB-hCO. The results are summarized below.

SettingMSE ↓MAE ↓Top50 MSE ↓r ↑
w/ shuffled pathway memberships0.0383 ± 0.00190.1218 ± 0.00350.1691 ± 0.02760.9626 ± 0.0014
Ours (real pathways)0.0322 ± 0.00120.1117 ± 0.00250.1275 ± 0.01700.9680 ± 0.0016

Real pathway memberships outperform the shuffled control on all four metrics, reducing MSE by 15.9% and Top-50 MSE by 24.6%, which indicates that the gains come from biological pathway knowledge rather than the grouping structure alone.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.