acceptodds
Withdrawn from ICLR 2027

Leveraging Biokinetic Knowledge Priors for Data-Scarce Bioprocess Modeling

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

While deep learning has accelerated drug discovery, its impact on biomanufacturing, the production stage in which candidate molecules are scaled up in bioreactors, has been considerably more limited. The reason is data scarcity: bioreactor experiments are expensive, require days to weeks to complete, and are rarely shared in public form, so generic neural decoders tend to overfit. Microbial growth dynamics, in contrast, have been described by biokinetic ordinary differential equation (ODE) models for several decades; how this knowledge should be injected into a neural network has not been studied systematically. We compare two orthogonal channels for injecting biokinetic priors on a single task with a shared backbone: simulation pre-training, where synthetic curves drawn from biokinetic ODEs pre-train a generic decoder, and architecture-level priors, where the ODE is embedded directly in the decoder. Across 11 datasets and 7 microbial species, both channels improve over no-prior baselines. Simulation pre-training is the more effective of the two: a generic decoder with simulation pre-training attains R² ≈ 0.515, matching a fully bio-structured decoder (0.554) trained on real data alone. The two channels therefore act as substitutes, and biokinetic specificity is the key factor: random-curve simulation fails, and pre-training outperforms joint training. Together, these results position simulation pre-training as a practical, data-efficient strategy for deploying deep learning in data-scarce bioprocess settings

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
bioprocess, physics informed, growth curve, ordinary differential equation, biokinetic ode, simulation, pretraining, fed-batch process, batch process

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_BsSiOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies learning-based approaches for modeling bioreactor dynamics. These experiments remain highly expensive in practice, limiting the availability of high-quality, real-world data. They learn these dynamics with an encoder-decoder backbone, pretrained on a set of synthetic simulations. To directly embed the underlying dynamics of the system, the paper proposes injecting a biokinetic-driven prior, the governing ODE, directly into the decoder's forward pass. Both approaches provide much stronger avenues to modeling the dynamics than existing neural-surrogates, demonstrating the efficacy of the approach.

Critical strengths
  • Setting and Motivation: The setting is practical and well motivated. The study is effective in investigating the proposed approaches for integrating the biokinetic prior into the learning process.

  • Ablations Are Insightful The set of ablations are informative and well designed. The set of baselines provides good coverage for different approaches to solving the targeted problem (Tables 1 and 2), the set of findings are informative (e.g., broad parameter sampling > narrow parameter sampling), and the results remain interpretable.

Critical weaknesses
  • Novelty: While the comparison itself has merit, the methods tested are well precedented within the literature. Hence, the contribution is reduced to a set of ablations on a niche domain; there is no signal whether these results generalize to other scientific settings, making the implications of the results limited. While this study is interesting in a narrow sense, it does not introduce novel methodological components -- just a new "recipe" for training on simulations in this domain. As is, the contribution remains too limited.

  • Expanding Data-Scarce Ablations: One thing that's missing from the study is more detailed evaluation of how these models perform under data scarcity. This is a point that seems to be built upon in the motivation, but the ablations do not seem to capture truly data-scarce settings. While I understand that the real-data was limited, the paper seems to have missed an opportunity to demonstrate the failure points of (a) using simulation data instead and (b) training on limited size datasets.

Questions for authors
  • Could you please provide more details on the S, M, and L split sizes and their relation to "scarce" data?
  • Where are the cardinal points derived from?
  • How were hyperparameters selected? Are there ablations to support these choices?
Reviewer_PGeeOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper investigates incorporation of biomanufacturing ODE inductive biases into NNs for data-scarce bioprocess modelling. They compare simulation based pretraining against embedding the ODE as a prior into the architecture, proposing BioStruct-ODE as a combination of growth equations with learned neural residual corrections. Across 11 datasets, they find both approaches improve prediction accuracy and show that diverse simulations outperform generic synthetic data.

Critical strengths

The empirical evaluation is a strength of the paper. Covering 11 datasets, multiple architectures, and five independent runs per configuration provides reasonable confidence in the robustness of the results. The consistent evaluation protocol and extensive ablations of architectural priors, simulation strategies, and context conditioning further strengthen the empirical findings.

Critical weaknesses

The contribution appears primarily empirical and incremental. In my view, the proposed architecture-level approach is essentially an application of universal differential equations (UDEs), whilst simulation-based pre-training is well established across a range of AI4Science domains. Although the systematic comparison of these approaches in biomanufacturing is useful, I am not convinced that it constitutes sufficient novelty for the venue.

The evaluation protocol potentially biases the results in favour of the proposed methods. The authors exclude trajectories that are poorly fitted by Monod–Baranyi kinetics R2<0.4R^2<0.4, as well as three entire datasets on which all baselines perform poorly. This appears to risk selecting for data that already conform to author's assumptions and make me question the generalizability of their findings.

Questions for authors

Have the authors considered the connection to Universal Differential Equations (Rackauckas et al., 2020)? The proposed BioStruct-ODE architecture appears to closely follow this established framework.

Could the authors provide a paired comparison of MLP+Pretrain and BioStruct-ODE+Cardinal on the same seven datasets to substantiate their substitutability claim?

Reviewer_Y3gAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper explores embedding bacterial growth equations in a neural network or achieving the same results by training on synthetic curves. The authors evaluate both approaches on 11 datasets. First, they assess each approach's performance in predicting the growth curve from the experimental parameters. Then the authors estimate the impact of early measurements on forecasting performance.

Based on the results presented by the authors, pretraining the MLP is a strong choice, as the average R^2 increases from 0.168 to 0.515, approaching the 0.521 reported by BioStruct-ODE for training on actual data alone. However, it is unclear whether both approaches will perform comparably in the specific experiment; two models may have a comparable average but differ significantly on particular datasets. A closer analysis of model performance on particular datasets would help determine which models' predictions to prefer in certain cases

Critical strengths
  1. Figure 2 is one of the best parts of this paper. In it, the authors run several thorough experiments, training the same MLP with varying amounts of the simulated data budget and comparing biokinetic model growth curves with smooth curves generated by the Gaussian process. The MLP performs better with the biokinetic curves, with most improvements occurring at a moderate simulation budget, while the Gaussian process offers little additional performance. Also, the GRU (R^2=0.494) result in Table 1 is interesting and useful for comparison with more complex models.

  2. The context experiment in Section 4.3 is very well designed and controlled: the authors evaluate the last 70% of the curve in cases where the model had no early measurements, the first 10%, or the first 30%. The results are easily interpretable. For the pretrained MLP, having 10% of the curve increases R^2 from 0.515 to 0.528, while having all 30% improves the score to only 0.534. The difference is small, especially given the cost of measurement.

Critical weaknesses
  1. It was hard to understand how the authors compare and prove that these two approaches are substitutable. This is illustrated in Section 4.2, where the authors compare the pretrained MLP’s 0.515 (eleven datasets) with Cardinal’s 0.554 (seven), but they provide no standard deviations to show how often the models converge on those seven datasets. The comparison with the base BioStruct-ODE seems natural, as this model was tested on all eleven datasets. However, this model's score is 0.521 in Table 1, whereas the diagonal entries in Table 9 are 0.573 (as the caption notes, these two numbers should match). Moreover, I would like to see the results for smaller training splits in Table 5. For instance, can the MLP pretrained on simulations converge on 20% of the curves as fast as the model trained on 70% of real data? That would more directly show how many experiments simulation could replace.

  2. It is also not quite obvious what kind of curves the benchmark uses. In Appendix A.2, all three datasets were removed because the best baseline performed poorly, yet BL21 was kept even though it was below the same threshold. The authors say they keep only curves that can be modeled with the Monod-Baranyi equation with R^2 >= 0.4, which should select the growth type they are interested in anyway. But if you look at the released code, there is another criterion based on the ratio of the peak value to the initial value. How the priors are fitted is important as well; the scripts fit the priors on seed 0’s training curves and use them whenever the partition changes. If you relied on the default behavior, a test curve at some future seed may have affected the prior used to predict it.

Questions for authors
  1. For the seven data sets provided by the pretrained MLP, BioStruct-ODE, and Cardinal, can the authors tell me what the magnitude of the differences is at the 20%, 40%, and 70% training splits? Is there any value below which the authors regard these models as essentially the same?
  2. Can the authors provide me with the IDs for the curves in the training set, test set, and pre-fitted for each seed value, as well as the filtering criterion that yielded these IDs? It would also be nice to see how the ranking changes once you include the three datasets that were not included.
  3. In Figure 2, both the growth model and the range of parameters sampled are modified simultaneously. With the simulation budget fixed, how do these two methods compare?
  4. How were the ComBase measurements converted before computing the accuracy factor? For example, predicting 7 when the observation is 6 on the log_{10} CFU/mL scale is a 10x error in cell count. Passing 7 and 6 to the released metric gives an accuracy factor of about 1.17. How was this calculation used in Tables 1 and 2?

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

Related papers

Discussion (0)

Sign in to comment.