acceptodds
Under review as a conference paper at ICLR 2027

Multi-Objective Amortized CDRH3 Sequence Sampling Diversifies Promising Antibody Candidates

Abstract

Antibody therapeutics have transformed treatment across diverse therapeutic areas and have become central to modern medicine. While recent machine learning approaches aim to accelerate antibody discovery through structure-based design and target-affinity optimization, successful candidates must also satisfy broad developability constraints, including structural compatibility, low immunogenicity, and favorable biophysical properties, while remaining sufficiently diverse to provide sufficient coverage of design space to mitigate later-stage development failures. We introduce a preference-conditioned multi-objective generative flow network (GFN) fine-tuning pipeline for fixed-framework CDRH3 loop redesign that fine-tunes a biological sequence model against many inexpensive proxies covering backbone compatibility, sequence likelihood, immunogenicity risk, novelty, and developability liabilities. It combines up to fourteen predictive rewards in a single policy trained to sample CDRH3 loops across 623 SAbDab antibody structures, leveraging a reward reduction strategy to preserve Pareto dominance under this learned map. We also introduce a novel autoregressive span-in-filling variant of AMPLIFY-350M that serves both as a scaffold-conditioned CDRH3 generation prior and as a direct conditional likelihood scorer. We evaluated reward scores, diversity, and design quality on loops sampled on 45 held-out antibody complexes to assess structural consistency, confidence, aggregation propensity, stability, and antigen binding. Compared with the inverse-folding methods ProteinMPNN and ESM-IF, our approach produces roughly more distinct acceptable candidates per design at 50% sequence identity, while holding of the 14-objective Pareto front pooled across methods, trading per-sample quality for greater exploration of novel possible structures. The results suggest that by relying on a more diverse data distribution and reward distribution, we are able to generate a wider array of promising candidates. All code and data will be made publicly available upon publication.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Amortized Sampling, Antibody Design, Multiobjective Optimization, Biological Sequences

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_UTzSOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a preference-conditioned, multi-objective generative flow network (GFlowNet) for redesigning antibody CDRH3 loops within fixed frameworks. The method fine-tunes a biological sequence model using up to 14 predictive rewards spanning structural compatibility, immunogenicity, novelty, and developability. It also introduces an autoregressive span-infilling variant of AMPLIFY-350M for scaffold-conditioned generation and conditional likelihood scoring. On 45 held-out antibody complexes, the approach generates approximately ten times more distinct acceptable candidates per design than ProteinMPNN and ESM-IF at a 50% sequence-identity threshold.

Critical strengths
  1. GFlowNets have been widely applied to biomolecular design because they can sample candidates with probabilities proportional to externally specified rewards. In this study, the authors integrate GFlowNets with multi-objective optimization to construct the diverse reward function incorporating multiple criteria for antibody CDR-H3 design. Therefore, this research work can be regarded as a successful application of GFlowNets in this domain.

  2. The authors progressively optimize a multi-objective GFlowNet through four stages to achieve the ultimate goal of CDR-H3 design. By incorporating multi-objective optimization into the GFlowNet framework, their approach demonstrates a strong foundation in mathematical theory and AI algorithms and provides valuable insights for future research.

  3. Notably, at the experimental level, the proposed approach yields approximately ten times more distinct acceptable candidates per design than the inverse-folding methods ProteinMPNN and ESM-IF, using a 50% sequence-identity threshold.

Critical weaknesses
  1. The framework overview in Figure 1 does not currently convey the algorithmic workflow clearly, particularly for readers who are not already familiar with GFlowNets. I suggest that the authors emphasize both the overall GFlowNet architecture and the 14 reward functions tailored to the CDR-H3 design task. Specifically, the figure could draw on the flow-network schematics presented in Bengio’s blog and prior related work, while highlighting how the adapted reward functions capture the biological objectives relevant to CDR-H3 design.

  2. I recommend that the authors revise the experimental section. The current figure and table captions are somewhat lengthy, and the experiment section lacks ablation studies that would isolate the contributions of individual GFlowNet components, such as the policy model, reward function, and training objective.

  3. In the Related Work section (Page 3, Line 120-121), according to my knowledge, the product of experts refers to the normalized product of the probabilities assigned by two models—specifically, the sum of the log-probabilities from ProGen2 and the Potts model followed by renormalization—rather than to ProGen2 alone. The authors should revise this description slightly to ensure that the distinction is clear and technically accurate

  4. A persistent challenge in applying GFlowNets is training instability during flow matching. How do the authors weight the 14 reward functions, and how stable is the training process throughout this research study?

Questions for authors

I suggest that the authors expand the experimental evaluation to clarify which of the proposed model’s novel components contributes most to its final performance. Such analysis would provide stronger evidence for the advantages of the proposed method.

4 comments

The updated version of our manuscript is available - all changes are marked in RED. Note that the majority of the new analysis is in the appendix currently, in it's own section "Preliminary Results"

Thank you for these concrete suggestions. We revised Figure 1 to show partial-sequence states, amino-acid actions, terminal rewards, the policy and the 14-reward preference mapping, and shortened captions while specifying what their bars represent. The related-work description now identifies the normalized product of ProGen2 and antibody-domain Potts probabilities.

Component contributions. Appendix H compares learned and fixed projection on identical rewards (Figure 24), novelty removal with projection refitting and original REINFORCE (Figure 25; Table 19), heated controls (Figure 26), and non-GFN reward selection (Table 16). Learned projection improves shared reward coverage but provides no clear shortlist quality/diversity advantage. On the selected 20 targets, the no-novelty arm retains substantial raw diversity and has higher mean fractions passing the common quality thresholds than D5, without a clear diversity or energy gain in the Rosetta-best 13 shortlists (Tables 18–19). Original REINFORCE has concentrated sampling; different exploration settings prevent attributing this solely to the optimizer. We did not ablate the AMPLIFY backbone, so these comparisons do not identify one component as the principal cause of final performance. Added main Figure 4 shows differences in raw sequence neighbourhoods; main Figure 5 shows varied predicted contact patterns among designs passing common quality thresholds. These findings suggest potentially functional variation, but practical usefulness requires experimental validation, hence our "promising" framing.

Weighting and stability. For normalized rewards r̂, positive row-stochastic projection A and sampled slot preferences w, the scalar reward is R=wᵀAr̂, with effective objective weights Aᵀw; the target sampling distribution is proportional to R⁸. We sample preferences over the reduced reward slots; the projection and sampled preferences jointly determine the weights on the fourteen original rewards. Figure 28 tracks how the probabilities assigned to a fixed set of sequences change during training. For each model, we selected 64 CDRH3 sequences from its final generated pool for each of 12 antibody frameworks, reusing these sequences under five fixed preference vectors. Figure 28 shows 26 saved main-trajectory checkpoints—13 per model—with stored normalized rewards and the final projection held fixed. Median centered trajectory-balance residual SD decreases from 1.477 to 0.850 for D4 and 1.475 to 0.764 for D5. This supports closer agreement between relative sequence probabilities and the final reward target on these fixed test cases. The bands describe variation across framework/preference combinations, not independent training runs. The online min–max normalizer and learned projection change the reward target during training before each is frozen. We did not separately ablate these safeguards, and the diagnostic does not establish global convergence or robustness across training seeds.

We will reply on the last day of the discussion period as further results become available, but appreciate that you might only have time to consider the results available today. Thank you for your suggestions, which have materially strengthened the precision of the claims made in the paper.

Should you have further suggestions around clarity of the experimental write up, we would be happy to include those clarifications in time for any camera ready version of the paper.

AuthorsOpenReview ↗

Supplement with Preliminary Results Updated, including UMAPs in a Supplemental .zip

Please note our general comment above "Supplement with Preliminary Results Updated, including UMAPs in a Supplemental .zip"

We look forward to your feedback regarding our work, and would be happy to answer any final questions you might have. Thank you for your careful review and insightful feedback.

I thank the authors for their response. After carefully considering the responses, my overall assessment of this paper remains unchanged, and I therefore retain my original score.

AuthorsOpenReview ↗

Any specific concerns I failed to address?

Are there any specific concerns from your review that my revisions failed to address?

Reviewer_s4jUOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper is proposed to address the question of generating massive qualified and diversified antibody candidates instead of a few structurally similar candidates with high rewards such that the subsequent screening can have diverse choices and options.

The proposed method is called Preference-conditioned Multi-Objective GFlowNet (MOGFN). There are three steps: (1) train a antibody sequence generative model based on AMPLIFY-350M with autoregressive span-in-filling; (2) do multi-objective optimization with GFlowNet using as many as 14 rewards with a learnable projection layer to compress them into 4-5 preference slots such that multiple sequences can be sampled from the model with different preferences; (3) generate massive sequence candidates and then use evaluation tools like ABodyBuilder3, Chai-1, Rosetta, FoldX to evaluate their structure quality, binding energy, etc. to filter out potential good ones.

Three main findings: (1) the generated candidates have significantly higher diversity; (2) diversity increased but the candidate quality decreased; (3) multi-objective steering is effective but not accurate enough.

Critical strengths
  • This paper addresses a meaningful challenge in practice, i.e. antibody design requires the trade-off between multiple objectives rather than optimizing a single structural or binding score.
  • The framework is reasonably comprehensive, combining multiple reward models, preference-conditioned generation, and a multi-stage evaluation workflow for antibody structures.
  • There is extensive experiments and analysis with a honest statement on the quality-diversity trade-off. The reported 10$\times$ gain in distinct useful candidates is interesting.
Critical weaknesses
  • The proposed method is more like a combination of existing techniques rather than an innovation. The contribution lies primarily in the integration of these techniques for antibody design.
  • The proposed method consistently underperforms inver-folding baselines on structural confidence and predicted binding energies. Although diversity is an important metric, it's still not clear whether more candidates provide larger gains beyond just covering the sequence space.
  • More ablated comparisons would help whether the gains are specifically from multi-objective GFlowNet training, rather than the generative prior or increased exploration. In particular, diversity-encouraging baselines and clearer ablations of the learned projection layer of rewards would make the conclusion more solid.
  • This is not a criticism on the missing wet-lab validation. However, it's pretty hard to solidify the believe that quality-diversity improvement can help realistic antibody functionality. This is particularly relevant since the generator does not directly condition on the antigen.
Questions for authors
  • How does the method compare against diversity-augmented versions of ProteinMPNN or the AMPLIFY prior under matched sampling budgets and quality thresholds?

  • What is the empirical benefit of learned reward projection over simpler fixed groupings, especially given that the D=4 projection largely suppresses the novelty objectives?

  • Can the authors better characterize whether the additional diverse candidates offer improvements in predicted functional outcomes, rather than primarily sequence novelty?

7 comments

The updated version of our manuscript is available - all changes are marked in RED. Note that the majority of the new analysis is in the appendix currently, in it's own section "Preliminary Results"

Thank you for your insightful questions. The new control experiments we have run in response to your queries support the general framing of our work with greater precision: we have evaluated the components of the multi-objective sampling workflow and span-infilling prior, and identify distinct computational trade-offs rather than general superiority of the various approaches.

Under identical reward-based selection, the AMPLIFY prior achieves higher mean reward, while GFN samples retain more sequence clusters after liability filtering (Table 16). Our original REINFORCE model produces little diversity, but different exploration settings prevent attributing this difference to the training objective alone (Figure 25). We are currently running a second REINFORCE baseline with sampling temperature and replay settings matched to the GFN.

Diversity-augmented baselines. Figure 26 compares a fixed-semantic 5-element reward grouping (which is not learned) and learned preference GFN with the novelty terms removed (12 rewards total), with ProteinMPNN and the AMPLIFY prior at T=1.25, all of them using 32,768 proposals, 250 Chai screens and 50 scored finalists per method/target (these reduced budgets are in light of computational constraints - for any final version, these budgets would be matched to the original paper at a minimum). As stated previously, GFNs have lower qualifying fractions compared to these baselines. GFN diversity persists among structurally screened, Rosetta-ranked shortlists (Figures 3 and 25), although a within-set diversity advantage over heated ProteinMPNN is not established when imposing common absolute confidence and energy thresholds (Figure 26). Because both heated baselines also yield more passing designs, these results show a trade-off rather than an overall quality–diversity advantage for the GFN. This is a limited evaluation due to our inability to test a wide range of baseline temperatures effectively. At matched generation budgets, GFNs retain substantially more sequence clusters after liability screening than the heated baselines, and this broad exploration persists without the two explicit novelty rewards (Table 19).

Projection and novelty. With identical 14-reward sets, learned D5 improves shared semantic hypervolume but lowers qualifying yield, without a clear shortlist-diversity gain (Figure 24; Table 18). This effect is modest overall, and differences between GFN and non-GFN methods persist regardless of specific decisions made regarding the learned projection or novelty reward inclusion. D4 largely downweights novelty, illustrating that dominance-preserving reduction can still leave individual objectives with little practical influence; learned projection therefore does not guarantee balanced coverage of all fourteen objectives (Figure 12). The no-novelty variant has higher mean passing fractions (Rosetta: 11.8% vs 7.83%, FoldX: 15.8% vs 14.448%, comparing the no-novelty variant vs. D=5) while retaining similar shortlist diversity (10.13) and lower median Rosetta energy (3.05 vs 3.71) (Table 18). However, removing novelty reduces diversity among equally sized sets passing the common Rosetta gate. In any case, the method presented is not dependent on any particular set of rewards, but instead should be viewed as a framework for sampling a diverse set of designs informed by a configurable suite of information sources and a rich prior (as measured by the size of the pre-training data).

Potential functional relevance. Added main Figure 4 quantifies local separation from the prior, inverse folders and Uniform sampler. Figure 5 shows more predicted contact patterns than inverse folders, while Uniform has more patterns and the prior contrast remains uncertain. Reward profiles and Pareto coverage provide further evidence of variation in relevant proxies (Figure 29; Table 15). However, fixed-budget mixtures yield fewer qualifying clusters and best energies remain worse (Table 18; Figure 27). These findings suggest potentially functional variation, without demonstrating improved biological function or practical usefulness. New structural comparisons remain exploratory on 20 outcome-selected targets.

Antigen conditioning. Appendix H.6, Antigen Conditioning and Raw Comparisons (Tables 19–20; Figure 29), now reports a variant of our GFN which is directly conditioned on the antigen binding region. Adding antigen conditioning shows no clear change in the assessed structural-quality and shortlist-diversity measures on 20 outcome-selected targets (Table 20). Possible explanations include training limitations or our multitask model already capturing transferable patterns across recurring antigen systems through hundreds of scaffold tasks. Uneven antigen representation in SAbDab is documented (Li et al., 2025). These explanations remain untested; the comparison does not establish equivalence or generalization to new antigen families beyond the generalization seen across all GFNs tested here (note that all reported results from our GFNs are on scaffold tasks not observed during training).

As an aside we note that the motivation of this work is the vast sequence data available likely informs designs unavailable in the current structural databases. The scale difference is enormous: about 1.5 billion unpaired antibody sequences in OAS (Olsen et al., 2022) versus about 23,000 experimentally determined antibody structural instances in SAbDab2 (Capel et al., 2026). This motivates using sequence priors alongside structural information.

We will reply on the last day of the discussion period as further results become available, but appreciate you might only have time to consider the results as they exist today in your evaluation of our work. Thank you for your suggestions, which have materially strengthened the precision of the claims made in the paper.

Thank you for your response and additional experiments. I appreciate your efforts to address my concerns and to characterize the quality–diversity trade-off more carefully. The new results and analysis are informative but several of my core concerns remain:

  1. While GFNs show higher diversity among same-sized sets of qualified candidates, they produce fewer distinct qualified candidates under a given screening budgets. The diversity advantage over temperature-adjusted ProteinMPNN is also unclear. Thus, the practical benefit of the additional diversity remains uncertain.
  2. The learned reward projection improves reward-space coverage but shows no clear advantage in final shortlist quality or diversity over fixed grouping. Additionally, the REINFORCE comparison remains confounded by differences in sampling temperature and replay, making it difficult to isolate the contribution of GFN training itself.
  3. The predicted contact-pattern analysis is a useful addition, but it remains unclear whether the additional diversity translates into improved functional outcomes beyond sequence novelty.

Overall, the revision strengthens the empirical analysis and better calibrates the claims. Nevertheless, I'm still not fully convinced that the proposed framework offers a clear enough empirical advantage over simpler alternatives, particularly given its limited methodological novelty. Thus, I lean toward keeping my original rating.

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

The paper introduces a preference-conditioned GFlowNet for fixed-framework, fixed-length CDRH3 redesign. It combines an autoregressive AMPLIFY infilling prior with a learned projection of 14 proxy rewards. Generation uses heavy-chain context without antigen or light-chain input (Appendix B.3). Evaluation on 45 held-out complexes finds increased diversity and lower predicted aggregation propensity, with reduced structural and energy-based quality relative to inverse-folding baselines (Table 2).

Critical strengths

The diversity advantage survives substantial filtering (Figs. 2-3 and Appendix Fig. 8). Higher embedding diversity persists at matched loop-confidence thresholds and within each arm's best Rosetta quarter. This supports a complementary exploration strategy beyond unfiltered sequence diversity. Matched candidate budgets, target-bootstrap intervals, documented overlap audits, and extensive oracle and hyperparameter analyses make the computational trade-offs assessable

Critical weaknesses

The methodological advantage is not isolated. Fixed and learned GFNs use nine and fourteen rewards, respectively, and no non-GFN reward-optimized comparator is evaluated (Secs. 3.5, 4.1). These comparisons do not isolate learned reduction's benefit or establish an advantage over simpler reward-optimized sampling. Computational shortlist benefit remains unresolved. The large cluster advantage appears mainly before budgeted filtering (Fig. 4). GFNs trail inverse-folding baselines on final energy scores and are close to random on FoldX (Table 2). Matched pLDDT and arm-specific Rosetta quartiles do not establish diversity gains at common energy thresholds.

Questions for authors

With identical reward sets and comparable oracle budgets, do learned projection and GFN training improve upon fixed projection and non-GFN reward optimization? Within the scored B4 pools, at common Rosetta/FoldX thresholds, what are each arm's qualifying-candidate yield and diversity at matched sample sizes?

2 comments

The updated version of our manuscript is available - all changes are marked in RED. Note that the majority of the new analysis is in the appendix currently, in it's own section "Preliminary Results"

Thank you for identifying the two comparisons needed to assess our claims. To address these, we have added identical-reward controls and common-threshold shortlist analyses.

Learned projection and reward optimization. With the same 14 rewards and five preference slots, learned projection increases shared semantic hypervolume from 0.5870 to 0.6048, but lowers the fraction of designs passing the confidence and energy thresholds and provides no clear shortlisted-diversity improvement (Figure 24; Table 18). The AMPLIFY prior generates candidates without reward-preference conditioning. We then use preference-weighted reward scores for post-hoc selection from that pool. Applying identical selection to an equally large fixed-semantic GFN pool yields slightly lower mean reward but retains more distinct liability-surviving sequence clusters (Table 16). MOGFN targets reward-proportional sampling for each preference; these selection controls do not measure fidelity to that target distribution.

Our original REINFORCE recipe produces highly concentrated samples (Figure 25; Table 19). We are testing a follow-up with higher sampling temperatures during training and a replay buffer, aiming to report its evaluation before discussion closes. Our current comparisons do not isolate the effect of the training objective. Suitably formulated entropy-regularized RL can target the same reward-proportional distribution as a GFlowNet (Tiapkin et al., 2024); whether it reproduces our empirical quality–diversity trade-off remains untested. Section H.3 reports nominal oracle budgets; Table 17 separates measured training components from estimated screening costs. Total compute of these two approaches is unequal.

Yield and diversity at common thresholds. Table 18 applies H3 pLDDT≥80 with Rosetta≤0 or FoldX≤0. On the original 45 targets, 249 scored draws produce 21.36 qualifying D5 designs versus 42.71 for ProteinMPNN under the Rosetta gate. D5 also produces fewer qualifying clusters. Conversely, equally sized sets of designs passing the same confidence/energy thresholds have higher mean embedding diversity for GFNs on supported targets. Figure 27 retains the adverse best-energy results. GFNs therefore offer more variety among equally sized sets of passing designs, but their lower pass rate means fewer qualifying designs and distinct clusters at a fixed draw budget.

The added main Figures 4 and 5 distinguish sampled sequence neighbourhoods and predicted contact-pattern variation. Together with reward coverage, these suggest potentially functional variation, without proving usefulness or improved screening productivity. Figure 4 concerns raw pools; Figure 5 concerns designs passing common gates. Preliminary new-arm comparisons use 20 outcome-selected targets; target-bootstrap intervals do not measure robustness across independent training seeds.

We will reply on the last day of the discussion period as further results become available, but appreciate you might only have time to consider the results as they exist today in your evaluation of our work. Thank you for your suggestions, which have materially strengthened the precision of the claims made in the paper.

AuthorsOpenReview ↗

Supplement with Preliminary Results Updated, including UMAPs in a Supplemental

Please note our general comment above "Supplement with Preliminary Results Updated, including UMAPs in a Supplemental .zip", which details the final experiments added to the revision.

We look forward to your reconsideration of our work in light of these experiments, and would be happy to answer any final questions you might have. Thank you for your thoughtful review and feedback!

Comments on the paper

4 comments
AuthorsOpenReview ↗

Thank you all for your thoughtful comments - a note on compute constraints

Dear reviewers,

I'd like to thank you all for your fair, insightful, and thoughtful reviews. We are working hard now to address as many comments as possible.

I want to flag that our compute infrastructure is under heavy load, and these experiments are unfortunately quite demanding to run. I am flagging now that I am, in good faith, trying to run as many of the requested ablations as possible, including the expensive post-filtering pipeline to evaluate the samples generated from these models, but it remains unclear to me exactly how much of this will be complete by the deadline. For this reason, if there are specific experiments or analyses that you would find particularly compelling, please flag them in response to this comment. In the meantime I will use my best judgement to run the subset of experiments that will hopefully answer your questions by the end of the discussion period.

I would also like to pre-register my commitment to running a more complete form of the preliminary experimental results I'm producing for the discussion period deadline, which will easily be feasible before the camera-ready deadline, should the paper be accepted into the conference. I will make clear in my individual responses how the preliminary results I produce for the discussion will be improved for the potential camera ready version.

EDIT: Please see below for our general reply to all reviewers. The updated version of our manuscript is available - all changes are marked in RED. Note that the majority of the new analysis is in the appendix currently, in it's own section "Preliminary Results".

AuthorsOpenReview ↗

Overview of the Experiments and Analysis Ran

Thank you to all three reviewers. We revised the paper to generally separate three questions: 1) whether different samplers explore different regions of the design space, 2) how those designs satisfy computational quality criteria, and 3) whether their diversity improves practical outcomes. Our present experiments address the first two, but unfortunately do not establish the third.

To address these questions, we analyzed additional samples from both the AMPLIFY prior and ProteinMPNN at a higher sampling temperature (T=1.25), and trained four additional baseline and ablation models: (1) a GFN with fixed semantic grouping of all 14 rewards, (2) a GFN with both novelty rewards removed and a refitted learned 12-to-5 reward projection, (3) a REINFORCE baseline using the same policy architecture and fixed semantic reward grouping, and (4) a GFN with direct antigen conditioning added to the fixed-semantic model.

Published database snapshots illustrate the scale difference: about 1.5 billion unpaired antibody sequences in OAS (Olsen et al., 2022) versus about 23,000 experimentally determined antibody structural instances in SAbDab2 (Capel et al., 2026). This motivates using sequence priors alongside structural information.

In the main text we have added Figure 4, which quantifies local sequence-neighbourhood separation in ESM2 embedding space, including the Uniform sampler. This figure serves as a summary statistic which demonstrates whether "diversity" within a particular sampler is truly exploring distinct regions of embedding space. Our results show that the different classes of models sample partially non-overlapping regions of embedding space, and individual instances of models tend to explore a similar space. The results suggest that 1) the individual changes made to the GFN training procedure only have a minor influence on embedding space exploration, and 2) increasing the sampling temperature of the baseline methods is not sufficient to bridge between methods in general. In addition, main Figure 5 measures predicted antigen-contact variation among equally sized sets of designs passing the same confidence/energy thresholds. Together with reward profiles and Pareto coverage, these results are suggestive of potentially functional variation rather than sequence novelty alone. We acknowledge that a major limitation of this work is our inability to provide functional validation, and believe that this should be the primary purpose of any follow up work.

Appendix H collects the completed projection and novelty ablations, higher temperature controls, simple multiobjective REINFORCE baseline, and prior-selection comparisons, reward/Pareto analyses, cost estimates, common-threshold shortlist and mixture outcomes, training diagnostics, and antigen-conditioning comparison (Figures 24–29; Tables 15–20). We have also produced an expanded rarefaction analysis with these new results which appears in Figure 10. We revised Figure 1, shortened captions and corrected the product-of-experts description. Adverse results as a result of these follow ups are made explicit: GFNs produce lower qualifying yields, worse predicted energies, and we find no obvious general advantage of the learned projection or antigen conditioning in these preliminary experiments.

AuthorsOpenReview ↗

Camera Ready Results

For a camera-ready version, if accepted, we will complete the already-commissioned 45-target expansion, and the temperature/replay-matched REINFORCE evaluation, and any other experiments requested during the end of the discussion period. The preliminary figures in the appendix, where appropriate, will replace those found in the main text, after the full computational results are produced (such that these results are supported by the same number of scored samples as the original submission). We acknowledge that the results as presented might be interesting to the community, but are also less clean than ideal due to the extreme computational cost of these experiments on our limited resources. We are therefore committed to ensure that camera ready version reflects a full effort to do a complete analysis as requested by the reviewers - at minimum using the same sample sizes and targets as the original submission, and we will update the comparisons accordingly. We thank the reviewers in advance for considering the practical limitation of compute costs in their evaluation of the work.

Note: insofar as the current results are preliminary, we have relegated them to the supplement and refer directly to those figures in our responses. For the camera ready, those preliminary figures will be refreshed with the full complement of results matching the compute spent on the main paper figures, and where appropriate the main paper figures will simply be replaced. As the discussion period is quite short given the time required for these experiments, we welcome any further experiments requested by the reviewers at this time, which we commit to running for the camera ready in addition to those experiments currently defined in the paper.

AuthorsOpenReview ↗

Supplement with Preliminary Results Updated, including UMAPs in a Supplemental .zip

Dear reviewers,

I've updated the manuscript with the preliminary results requested across all reviews, which are largely concentrated in Appendix H. We have also uploaded a .zip file under "supplemental materials" which contains UMAP projections for samples from all models. This is an alternative view of the results presented in Figure 4, which provides qualitative insight into how distinct the samples drawn from each method are (i.e., a representation of whether each method produces different designs across design space). This method is flawed due to the dimensionality reduction involved, but we thought you might find it interesting to consider alongside the results presented in the revised version of the paper. Figure 4 makes a similar argument, using summary statistics.

I would finally like to draw your attention to supplemental Figures 25 and 29. Figure 25 shows the quality/diversity trade-off across all methods (and when backfilled with the full set of samples across all 45 targets, will replace main Figure 3). The notable takeaways are that while GFNs produce less high scores under our post-filter oracles, they still produce many high-scoring designs with greater between-sample diversity than all competing methods including the random baseline. The failure of REINFORCE to perform well at this task should be evaluated alongside Figure 29, which demonstrates that the trained models do learn to satisfy the rewards during training, but fail to produce sufficient diversity when sampling, making them a poor choice for discovery problems. I have also added a short methods section for these baselines in the appendix (Section H).

The main paper Figure 4 now shows that the raw pool mixing of both REINFORCE methods remain distant in embedding space from all other methods, and adds in all requested ablations and baselines.

Unfortunately, due to the weight of the post-filtering pipeline and the demand on our compute infrastructure, not all 45 targets are complete, but remain running now. We remain committed to running the full post-filtering pipeline and analysis for all methods discussed during the rebuttal period and replacing the main-text figures with their corresponding Preliminary Results figures in time for the camera ready, should the paper be accepted. There is more than ample time to ensure those experiments are complete by the camera ready deadline. If there are any further analyses you would like to see in time for this, we would be happy to also run those - these suggestions strengthen the work and we believe the results will strengthen the value of the paper to the community.

Thank you all for your careful consideration of our work and insightful questions. I look forward to any final questions you might have given these results.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.