acceptodds
Under review as a conference paper at ICLR 2027

FORGE: Reaction-Guided Generative Design of Ionizable Lipids

Abstract

Ionizable lipids are essential for RNA delivery, but their discovery largely depends on synthesizing and screening libraries assembled from predefined reactants. This strategy preserves a clear synthesis route while limiting the chemical space that can be explored. *De novo* molecular generation could expand this space, yet existing whole-molecule models do not retain the reaction information needed to assemble their designs. We introduce FORGE (**F**low-matched, **O**pen-ended, **R**oute-resolved **G**eneration and **E**xploration), a generative model that uses known reaction chemistry to design ionizable lipids without restricting generation to a fixed reactant catalogue. FORGE represents each synthesis route through precursor roles, reaction-core positions, transformation order, and coarse molecular architecture, while generating the complete molecule at atom-and-bond resolution. The model does not receive component identifiers, stored precursor graphs, or fragment tokens. We evaluated FORGE on Ugi, repeated aza-Michael addition, and repeated reductive amination. Across three independent seeds, FORGE achieved exact level-one (exact-L1) transform-consistent yields of 96.4 ± 2.1%, 72.2 ± 1.6%, and 52.3 ± 3.4%, respectively. Matched whole-molecule generation followed by *post hoc* reaction filtering yielded 34.8 ± 7.1%, 19.2 ± 4.1%, and 37.0 ± 2.2%. FORGE also generated exact-L1 products with precursor identities absent from the corresponding training component catalogues across all three reactions, a capability unavailable to finite-catalogue assemblers. In an *in vivo* reporter mRNA-LNP study, three synthesized FORGE lipids demonstrated higher potency than the clinical benchmark MC3, with FORGE-3 showing approximately 13-fold higher signal. These results establish that known reaction chemistry can guide *de novo* molecular generation while preserving the information needed to verify assembly and demonstrate that generated lipids can achieve strong *in vivo* performance.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
discrete flow matching, molecular generation, synthesis-aware generation, reaction-aware conditioning, graph generative models, ionizable lipids, lipid nanoparticles, RNA delivery

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_kDy2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper targets synthesis-aware molecular generation for ionizable lipids, where existing methods either assemble from a fixed reactant catalogue or generate whole molecules freely. FORGE conditions a discrete flow-matching graph generator on a structured "reaction program" that specifies the assembly chemistry, depth, per-atom precursor roles, reaction-core positions, and coarse per-role morphology counts, but not precursor identities. Molecules are generated at atom-and-bond resolution, then verified by reverse decomposition followed by exact forward replay.

One shared model is evaluated on Ugi 3-CR, repeated aza-Michael addition, and repeated reductive amination across three seeds, against an unconditioned baseline, a cyclically mislabelled control, a finite-catalogue oracle, and three external generators. The paper reports higher exact-L1 yields for the conditioned model than for the unconditioned baseline on all three reactions, and recovery of precursors outside the training catalogues. Three generated lipids were synthesized and compared to DLIN-MC3-DMA in a mouse FLuc reporter study.

Critical strengths
  1. The verification protocol is careful and reusable. The paper distinguishes the recovered witness set Ê_c(G) from the true preimage E*_c(G) and claims only the inclusion that bounded reverse search licenses (Eq. 25). It defines strong catalogue escape (Def. C.5) and then states plainly that its component-novelty metric does not establish it. Decomposition coverage, the replay invariant, abstention, and exact ambiguity are reported separately rather than collapsed into one "synthesizability" number. This protocol is worth adopting independently of the generative model, and it is a substantial part of why my recommendation is not lower.

  2. The representation is a sensible fit to the domain. Encoding lipids as a spanning tree plus at most three closure edges supports the full 194-atom range without an O(N²) edge state, which matters for molecules of this size and near-acyclic topology. Combined with program-fixed coordinates held constant throughout sampling, the construction yields zero fixed-coordinate violations and zero support overflows across 27,648 attempts, so the conditioning mechanism itself is demonstrably doing what it claims.

Critical weaknesses

1. The headline gain is a validity effect, and no experiment isolates the contribution of reaction chemistry.

Table 3 gives both validity and exact-L1, which permits the conditional quantity the paper never reports. For Ugi, FORGE is 96.5% valid / 96.4% exact-L1 and the unconditioned model is 35.9% / 34.8%, so exact-L1 given validity is 99.9% vs 96.9%. For reductive amination it is 86.6% vs 88.3%, favoring the unconditioned model. The 61.6-point headline gap is therefore almost entirely about whether a valid connected graph is emitted at all, not about transform consistency, which is near-free in both cases. The abstract's claim that conditioning increased "exact transform-consistent generation" does not hold in that form.

The cyclic control cannot repair this. Since the program already fixes per-atom roles, core positions, and morphology counts, and Π_P pins the core throughout sampling (Prop. C.6), the reaction-identity token is largely redundant. Its null effect is equally consistent with "chemistry matters but is already encoded elsewhere in P" and with "chemistry does not matter." The ablation that separates these, conditioning on chemistry and depth alone without roles, core, and morphology, is absent, and it is exactly the claim in the title.

2. The generalization claim is contradicted by the paper's own held-component result.

The stated advantage over catalogue methods is reaching precursors outside the inventory. On the component-disjoint fold FORGE recovers 1.4 ± 1.7 held components per 1,000 attempts (per-seed 3.3 / 1.0 / 0.0), while unconditional DeFoG recovers 15.5 ± 2.4, roughly eleven-fold more. The contribution bullet rests instead on component novelty, a much weaker predicate that any atom-level perturbation satisfies; by rate, DeFoG is higher there too (97% of verified products vs FORGE's 87%).

Table 10 suggests the expansion runs away from realistic lipids rather than toward them: descriptor-manifold precision 29.5 for FORGE vs 54.6 for the catalogue oracle, fingerprint precision 0.0, and diversity 0.643, the lowest of any method in Table 2. The parsimonious reading is local perturbation around training templates, not catalogue escape.

Questions for authors

Answers to 1–3 could change my assessment.

  1. Tables 1/3 and 2/7 disagree by ~3× on the same configuration. Ugi "Shared null" is 35.9% valid / 34.8% verified in Table 3, but "Shared-null + post-hoc" is 194.7 ± 15.6 valid / 114.4 ± 10.9 verified per 1,000 in Table 2. Same models and same fold? If so, one is in error; if not, which supports the 34.8% in the abstract?

  2. Can you report exact-L1 conditional on validity for every row in Table 3, and is a (chemistry + depth only) ablation feasible during discussion? If the contrast largely vanishes under conditioning, how should the transform-consistency claim be restated?

  3. Reproducibility. A.7 states that no source fingerprint match was recovered for seeds 1 and 2 and that no complete training manifest was recovered. Since every headline number is a mean over those three seeds, can they be re-run under a pinned implementation before camera-ready?

  4. In vivo. How many animals and injection sites produced the four ROI values, and which limb is which formulation? RLU is a luminometer unit, so were these IVIS acquisitions, and can calibrated radiance or total flux be reported with a color scale and ROI definition? Why a single 4-hour timepoint? How were FORGE-1/2/3 selected from ~2,900 verified products, and was the D.2 role-resolved SMILES screen used? Was an animal-use protocol approved, and by which body?

  5. Architecture and theory. Given the negative conditional-dependence diagnostic, FACT outperforming the full model on reductive amination (520.3 vs 318.8 per 1,000), and PCGrad removal improving all three means, what justifies joint whole-molecule denoising? Relatedly, with training times confined to [0.02, 0.98] and PCGrad not a gradient of any scalar loss, what does Theorem 3.1 establish about the system actually trained?

8 comments

We thank the reviewer for the careful analysis and for recognizing the verification protocol and molecular representation. To address the concerns, we reverified the final checkpoint, completed 35 new training runs under a pinned implementation, and added the requested chemistry and catalog-escape experiments. We also clarify how we selected, synthesized, and tested the three generated lipids in the in vivo study. Comments 2 and 3 cover reproducibility, novelty, and the animal experiments.

Q1. Resolving the checkpoint discrepancy

Tables 1/3 and 2/7 disagree by ~3× on the same configuration.

We traced the reporting error to the intermediate shared-null checkpoint that we had used in Tables 2/7. For Tables 1/3, we had used the final checkpoint.

Ugi shared-null checkpointValid / 1,000Exact-L1 / 1,000
Intermediate checkpoint (Tables 2/7)194.7 ± 15.6114.4 ± 10.9
Final trained checkpoint (Tables 1/3)359.0 ± 73.6348.1 ± 70.7

Mean ± SD across three seeds, 3,072 attempts each. We independently reverified all 9,216 final-checkpoint attempts, confirming the abstract's 35.9% validity and 34.8% exact-L1. We will correct Tables 2, 7, and 8 and the associated text. Against this final baseline, FORGE achieves 96.5% validity and 96.4% exact-L1, a 61.6-point increase in verified-generation yield.

W1 + Q2. Isolating learned reaction chemistry

Can you report exact-L1 conditional on validity for every row in Table 3, and is a (chemistry + depth only) ablation feasible during discussion?

We obtained 100% replay precision for every configuration in Table 3. Exact-L1 among valid molecules therefore equals decomposition coverage:

Exact-L1 among valid molecules (%)FORGEShared nullCyclic
Ugi99.997.099.8
Aza-Michael100.069.5100.0
Reductive amination86.688.485.7

We report the per-seed aggregated values from Table 3. In the original Ugi comparison, we supplied both models with the same reaction core and structural constraints, and conditioning primarily increased the yield of valid molecules. To isolate the contribution of learned chemistry, we then trained matched models receiving chemistry and depth only, or neither input. We disabled core pinning and optional chemistry-specific decoder corrections in both models.

ReactionConditioningExact-L1 / attempt (%)Exact-L1 / valid (%)
UgiNone37.60 ± 5.6140.10 ± 5.45
Chemistry + depth55.04 ± 3.9458.87 ± 3.66
Aza-MichaelNone26.68 ± 5.6029.08 ± 5.56
Chemistry + depth65.86 ± 3.8468.99 ± 3.27
Reductive aminationNone70.88 ± 4.0484.70 ± 3.24
Chemistry + depth73.76 ± 2.5191.34 ± 1.26

Mean ± SD across five paired seeds, 8,192 attempts per reaction per seed. By adding chemistry and depth, we increased reaction compatibility among valid molecules by 18.77, 39.91, and 6.64 points. These gains directly demonstrate a learned chemical contribution, with the largest per-attempt improvements for Ugi and aza-Michael.

We also found that we had permuted labels consistently during training and evaluation in the original cyclic control. To test reaction identity directly, we instead changed only the label at inference, holding the trained model and other inputs fixed and disabling core pinning. Switching Ugi to aza-Michael changed Ugi compatibility from 44.10% to 0.00% and aza-Michael compatibility from 0.04% to 62.73%. Switching aza-Michael to reductive amination changed the corresponding rates from 57.85% to 0.82% and 11.76% to 31.45%. Switching reductive amination to Ugi increased Ugi compatibility from 0.00% to 22.62%, with reductive-amination compatibility at 43.55% (five paired seeds, 512 attempts per reaction per seed).

We further compared full and reduced neural-program conditioning while holding production decoding and structural constraints fixed:

Ugi metricFull conditioningReduced conditioning
Qualified distinct new products with novel precursors / 1,000895.31 ± 18.04655.08 ± 34.29
O–O-containing outputs (%) ↓0.783.32
ECFP4 diversity ↑0.6580.726

Five seeds, 512 matched attempts each. We qualified products by exact replay, structural screening, absence from the training-product union, and a novel recovered precursor. Full conditioning increases qualified novelty yield, with slightly lower overall validity, and reduces undesirable O–O-containing outputs. Reduced conditioning yields higher ECFP4 diversity. Together, these experiments isolate learned chemistry and show how the complete program improves qualified discovery under fixed constraints. We will clarify both contributions in the abstract and manuscript.

Q3. Reproducibility of the main results

Since every headline number is a mean over those three seeds, can they be re-run under a pinned implementation before camera-ready?

We completed a new five-seed study comprising 35 training runs across seven configurations, recording the source code, training splits, configurations, random seeds, and checkpoint hashes.

ReactionExact-L1 success (%)
Ugi93.68 ± 3.00
Aza-Michael92.66 ± 0.80
Reductive amination85.84 ± 1.68

We evaluated five independently trained models over 8,192 attempts per reaction per seed. We retained the submitted architecture, expanded the lipid data for aza-Michael and reductive amination, and ran additional updates. The original three-seed means were 96.4%, 72.2%, and 52.3%, respectively. Through this new study, we documented the training procedure, obtained high reaction-compatible yields across all three families, and completed the controlled ablations in the first comment. We report the new and archived studies separately and specify the data and training differences.

W2. Structural novelty and generalization beyond the catalog

The generalization claim is contradicted by the paper's own held-component result.

In the designated held-out experiment, FORGE recovered 1.4 excluded precursors per 1,000 attempts and DeFoG recovered 15.5. To examine the broader question of generation beyond the training catalog, we added precursor-distance, novelty-denominator, and exhaustive catalog-reachability analyses.

For the original 87% versus 97% comparison, we divided distinct component-novel products by all verified attempts, which made the rates sensitive to duplicate generation. To compare novelty among unique products, we recalculated the rates using distinct verified products as the denominator:

Ugi novelty metricFORGEDeFoG
Novel precursor rate among distinct verified products (%)97.2796.98
Similarity to nearest training product (lower is more novel)0.6600.773

Across three Ugi seeds, we measured ECFP4 Tanimoto similarity to the nearest training product. FORGE's lower similarity indicates greater average structural separation. We further compared precursors within each reaction role and found greater distances for FORGE isocyanides (0.464 versus 0.215) and aldehydes (0.213 versus 0.173), with comparable amine distances (0.316 versus 0.312).

To test catalog escape directly, we exhaustively enumerated 485,625 Ugi precursor combinations under the registered reaction, yielding 488,250 distinct reachable products. We counted a generated product as an escape when it passed exact verification and structural screening and fell outside this complete set (Definition C.5).

Qualified Ugi catalog escapes / 1,000 attemptsFORGEDeFoG
Mean ± SD833.3 ± 23.2258.4 ± 22.9

In this separate three-seed Ugi evaluation, FORGE produces 3.2 times as many qualified escapes per attempt. Because we enumerated the full reachable set, these results directly establish generation beyond every product obtainable from the specified finite inventory.

In Table 10, we compared the generated structures with the published-lipid distribution. FORGE achieved descriptor-manifold precision of 29.5 versus 54.6 for the catalog oracle, fingerprint precision of 0.0 under the reported metric, and diversity of 0.643 versus 0.727 for DeFoG. We then used exhaustive reachability analysis to establish expansion of the accessible product space. To assess biological function, we also synthesized three selected lipids and measured reporter expression in vivo, as detailed in the next comment. Together, these evaluations measure structural novelty, catalog expansion, and experimental performance.

Q4. In vivo experimental design and candidate selection

In vivo. How many animals and injection sites produced the four ROI values, and which limb is which formulation?

How were FORGE-1/2/3 selected from ~2,900 verified products, and was the D.2 role-resolved SMILES screen used?

To evaluate the generated designs experimentally, we synthesized three FORGE lipids, formulated each as FLuc mRNA-LNPs, and measured reporter expression in vivo (Figure 2 and Appendix B.5). We tested the formulations in two mice across four intramuscular sites, with one formulation per hindlimb and one site per formulation. At 0.05 mg mRNA/kg, we measured bioluminescence with Xenogen IVIS four hours after administration to assess early reporter expression. We administered D-luciferin approximately nine minutes before imaging.

We display calibrated radiance in photons/s/cm²/sr, with color ranges of 1.00 × 10⁷ to 1.00 × 10⁸ and 2.00 × 10⁷ to 2.00 × 10⁸. To quantify expression, we drew circular ROIs around each injection site and normalized the signals to nonluminescent background ROIs in the corresponding images. We report these background-normalized relative signals below and will use this precise label in the revised figure and text.

MouseInjection siteFormulationBackground-normalized signal
1Right hindlimb, ROI 1FORGE-15.997 × 10⁵
1Left hindlimb, ROI 2FORGE-22.036 × 10⁶
2Right hindlimb, ROI 3FORGE-32.881 × 10⁶
2Left hindlimb, ROI 4MC32.195 × 10⁵

We identify the limbs anatomically from the supine, head-up image orientation. FORGE-3 produced an observed 13.1-fold higher signal than MC3 at separate injection sites in the same mouse. By synthesizing, formulating, and testing the generated lipids, we demonstrated reporter expression in vivo at the four measured sites.

To select candidates for these experiments, we ranked the generated products and screened the leading designs in the following sequence:

StageSelection procedure
GenerationApproximately 2,900 reaction-verified FORGE products generated using reaction programs with defined precursor roles
Activity predictionCandidates ranked using the off-the-shelf LANTERN transfection predictor, with the top 20 prioritized
Precursor analysisAppendix D.2 role-resolved SMILES analysis applied to the top 20 candidates to examine the associated precursor structures
Final selectionThree candidates selected for experimental synthesis based on synthetic accessibility

We specified precursor roles during generation and then applied the Appendix D.2 analysis to the top-20 shortlist. From this screen, we selected three candidates using computational predictions and chemical accessibility before experimental testing. We conducted all animal procedures under an approved institutional animal-use protocol and will provide the identifying institution and protocol details after double-blind review.

Q5. Joint molecular generation and theoretical guarantees

Given the negative conditional-dependence diagnostic, FACT outperforming the full model on reductive amination (520.3 vs 318.8 per 1,000), and PCGrad removal improving all three means, what justifies joint whole-molecule denoising?

We use joint denoising to exchange information across the headgroup, linkers, and hydrophobic regions during generation. Alongside the conditional-dependence diagnostic, we directly tested this architecture against FACT. Joint generation yields 856.9 versus 56.9 exact-L1 products per 1,000 attempts for Ugi and 767.7 versus 396.4 for aza-Michael, and also exceeds the larger FACT model on both reactions. Reductive-amination yields are 318.8 versus 520.3. These comparisons identify the reaction families that benefit from joint generation. We further removed PCGrad and improved mean performance across all three families.

In Theorem 3.1, we prove recovery of the reaction-conditioned distribution at the population optimum of the reference objective under exact continuous-time sampling. Proposition C.6 further guarantees preservation of fixed coordinates during finite numerical sampling. To evaluate our implementation, which uses training times [0.02,0.98][0.02,0.98], auxiliary objectives, PCGrad, and a capped Euler sampler, we conducted the new five-seed study and matched ablations.

We appreciate the opportunity to resolve the reporting discrepancy and provide the requested evidence. We apologize for the length of the response and thank the reviewer for the careful consideration. In light of the new chemistry ablations, exhaustive catalog-escape results, and clarified in vivo findings, we would be grateful if the reviewer would consider a higher score and renewed confidence in the submission.

Reviewer_Stk5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper essentially is an application of conditional discrete flow-matching over molecular graphs. The conditioning variables are P=(c, d, r, k, M) as defined in the paper, and the workflow generates molecules at atom and bond resolution rather than building one from prior library or inventory, accepting it only if the reverse decomposition and a forward replay reconstructs it— measured via exact L1. Reported scores show improvement on exact L1 over the reaction families including Ugi, aza-Michael, and reductive amination compared to null-plus-filter control. Results also report to recover precursors outside the training library.

The appendix is honest about the limitations of the evidence.

Critical strengths

As far as I’ve understood, the real contribution and novelty lies in proposing a verification protocol via exact forward replay, every attempt in the denominator, seperating \epsilon^{*}{c}(G) and \hat{\epsilon}{c}(G) (Definition A.1), and measuring strong catalogue escape from actual measurements (Definition C.5).

Appendix B.3 reports negative results honestly, something I’d like to credit authors with.

Critical weaknesses

I have five main conceptual reservations:



W1: It seems that the conditioning via reaction name isn’t much useful. The performance difference when given a wrong reaction name (for example in the Ugi case 96.4 to 93.8), isn’t significant implying that the rest of the blueprint is the useful conditional signal not the chemistry, which goes against the paper’s claim of exploiting reaction awareness.



W2: Exact-L1 doesn’t exactly measure whether the generated molecule follows specified reaction program, it simply checks whether some valid precursor decomposition exists that can reconstruct the molecule. There can be cases where for the same molecule/valid precursors some unfeasible reaction route is required, to my understanding.



W3: For the reported reaction-consistency performance, it’s unclear whether the performance gain comes from model understanding chemistry or the imposed constraints, for example in the ugi cases enforcing core H states during decoding increases exact L1 from 75% to 95% while keeping everything fixed.



W4: While FORGE generates components absent from the training catalogue it remains to be evalued whether they were minor variations or drastically new candidates. Importantly while recovering heldout components FORGE has a rate of 1.4/1000 compared to 15.5/1000 of unconditional DeFoG raising questions about generalizability and the purported claims.



W5: Theorem 3.1 establishes consistency under idealized assumptions such as an optimal predictor and exact continuous-time sampling. While the theorem itself appears reasonable, it is unclear what additional guarantees it provides for the practical FORGE model beyond standard discrete-flow-matching theory. It’d be interesting to see whether the guarantees hold/break under real world settings of finite training, Euler approximation error, to name a few.

Questions for authors

Q1: Exact-L1 only requires the existence of a valid decomposition under the specified reaction family, how often does the recovered decomposition actually match the original sampled reaction program? Could the authors report adherence separately for reaction depth, precursor roles, core positions, and morphology? This seems particularly important for aza-Michael and reductive amination, where decomposition ambiguity is high.


Q2: Given that changing the reaction identity while keeping the rest of the program fixed results in only a marginal drop in performance (e.g., 96.4 to 93.8 for Ugi), how much of the improvement can actually be attributed to learning reaction-specific chemistry? Could the authors provide ablations removing the morphology information (M), and separately the precursor-role and core-position information ((r,k)), to understand which parts of the conditioning are responsible for the performance gains?


Q3: Enforcing Ugi core hydrogen states during decoding increases exact-L1 from 74.67% to 95.12% without changing the model weights. How much of the final performance is due to these manually enforced constraints rather than the learned generative model? Could the authors report the main results across all three reactions using a decoder without chemistry-specific corrections?

Q4: FORGE achieves substantially lower recovery of deliberately held-out components compared to unconditional DeFoG (1.4 versus 15.5 / 1,000 attempts). How do the authors bridge this with the claim that FORGE enables open-ended component generation? Could the authors additionally quantify how structurally different the component-novel precursors are from their nearest training-catalogue components, rather than only reporting whether their identities are absent?

Q5: What practical insight or guarantee does the theorem 3.1 provide specifically for FORGE? In particular, does the theoretical analysis explain why reaction-program conditioning should improve exact assembly or generalization beyond the training catalogue?

4 comments

We thank the reviewer for the detailed assessment and for recognizing the value of the verification protocol. To address the concerns, we added program-adherence measurements, controlled chemistry and decoder ablations, precursor-distance analyses, and exhaustive catalog enumeration. We also synthesized three generated lipids, formulated them as mRNA-LNPs, and measured reporter expression in mice in the submission's in vivo study. We present program adherence here, the chemistry and decoder experiments in the second comment, and novelty, theory, and in vivo findings in the third.

W2 + Q1. Adherence to the requested reaction program

Exact-L1 only requires the existence of a valid decomposition under the specified reaction family, how often does the recovered decomposition actually match the original sampled reaction program?

Could the authors report adherence separately for reaction depth, precursor roles, core positions, and morphology?

Exact-L1 verifies that a replayable decomposition exists under the specified reaction family. To test agreement with the requested program, we additionally checked whether one consistent decomposition matches the sampled depth, precursor roles, and core positions, individually and jointly.

ReactionDepth (%)Roles (%)Core (%)Joint adherence (%)
Ugi100.097.7100.097.7
Aza-Michael87.286.887.286.8
Reductive amination94.890.480.980.8

We evaluated exact-L1 products from five seeds and 8,192 attempts per reaction per seed, requiring the same decomposition to satisfy all three conditions. FORGE achieves 97.7%, 86.8%, and 80.8% joint adherence, respectively, directly demonstrating agreement with the sampled program.

To assess architectural control, we next compared the generated morphology with the requested numerical targets:

ReactionMorphology featureRequested meanGenerated meanDifference
UgiAttachment count1.0751.076+0.001
Cycle rank0.1790.1790.000
Exterior nodes9.6499.6490.000
Junction count1.2250.682−0.543
Aza-MichaelAttachment count1.5001.5000.000
Cycle rankn/rn/r0.000
Exterior nodes24.40724.717+0.310
Junction count2.0221.686−0.335
Reductive aminationAttachment count1.3201.221−0.099
Cycle rank0.4900.485−0.005
Exterior nodesn/rn/r+0.102
Junction count3.2032.920−0.283

n/r denotes an unreported mean. We report each difference as generated minus requested.

FORGE closely matches the requested attachment counts, cycle ranks, and exterior sizes. It generates 0.543, 0.335, and 0.283 fewer junctions than requested for Ugi, aza-Michael, and reductive amination. Together, these evaluations directly measure adherence to the reaction program and the accuracy of molecular architecture control.

W1 + Q2. Learned reaction chemistry and program inputs

Given that changing the reaction identity while keeping the rest of the program fixed results in only a marginal drop in performance (e.g., 96.4 to 93.8 for Ugi), how much of the improvement can actually be attributed to learning reaction-specific chemistry?

To address this concern, we audited the original cyclic control and found that we had permuted labels consistently during training and inference rather than only at inference. Thus, it did not serve as an appropriate control. We therefore changed only reaction identity at inference and added controlled ablations of the program inputs.

We first evaluated the complete production program, supplying the corresponding core and structural constraints for each chemistry. We then checked the generated products against all three reaction families:

Reaction program suppliedUgi (%)Aza-Michael (%)Reductive amination (%)
Ugi94.22 ± 3.150.000.08 ± 0.11
Aza-Michael0.0092.97 ± 1.050.00
Reductive amination0.002.85 ± 0.6682.58 ± 1.51

Mean ± SD across five seeds, 512 attempts per reaction per seed. FORGE predominantly generates products compatible with the requested reaction family under the complete program.

To isolate the reaction-identity input, we next trained a model with chemistry and depth alone, disabled core pinning, and used generic decoding. With the correct labels, we obtained:

Requested reactionUgi (%)Aza-Michael (%)Reductive amination (%)
Ugi44.100.0426.09
Aza-Michael0.0057.8511.76
Reductive amination0.000.5953.71

Across five paired seeds and 512 attempts per reaction per seed, we then changed only the reaction label. Switching Ugi to aza-Michael reduced Ugi compatibility from 44.10% to 0.00% and increased aza-Michael compatibility from 0.04% to 62.73%. Switching aza-Michael to reductive amination changed the corresponding rates from 57.85% to 0.82% and 11.76% to 31.45%. Switching reductive amination to Ugi increased Ugi compatibility from 0.00% to 22.62%, with reductive-amination compatibility at 43.55%. Thus, changing the conditioning label directly changes the chemistry of the generated products.

To quantify the contribution of chemistry and depth, we also trained matched models receiving both inputs or neither input. We disabled core pinning in both and measured the changes after removing chemistry and depth:

ReactionExact-L1 decrease among valid (pp)Exact-L1 decrease per attempt (pp)
Ugi−18.8−17.4 ± 2.6
Aza-Michael−39.9−39.2 ± 3.0
Reductive amination−6.6−2.9 ± 4.0

Five paired seeds, 8,192 attempts per reaction per seed. Removing these inputs reduces reaction compatibility even among valid molecules, with the largest effects for Ugi and aza-Michael. This directly supports the contribution of learned chemistry conditioning.

We separately removed morphology and role/core-position conditioning while holding the remaining inputs and generic decoding fixed:

ReactionRemoving morphology (M)Removing roles and core positions (r,k)
Ugi+3.32 pp−11.3 pp
Aza-Michael+5.8 pp−4.4 pp
Reductive amination−1.1 pp−8.9 pp

Changes in exact-L1 per attempt. Removing role and core-position conditioning reduces exact-L1 across all three reactions. We use morphology conditioning to control architecture and quantify that control in the first comment.

W3 + Q3. Decoder contributions

Could the authors report the main results across all three reactions using a decoder without chemistry-specific corrections?

To separate decoder corrections from the neural predictions, we evaluated four decoding configurations with identical predictions. We used the updated five-seed checkpoints, retaining the submitted architecture and expanding the lipid training data and training updates.

ReactionGeneric decoder (%)Hydrogen only decoder (%)Structural only decoder (%)Full model (%)
Ugi71.13 ± 3.9181.56 ± 2.8084.61 ± 5.1494.22 ± 3.15
Aza-Michael87.58 ± 2.5387.54 ± 2.7192.97 ± 1.0592.97 ± 1.05
Reductive amination84.41 ± 2.0783.95 ± 2.4185.43 ± 1.8885.43 ± 1.88

Exact-L1 per attempt, mean ± SD across five seeds, 512 paired attempts per reaction per seed. With generic decoding, FORGE achieves 71.13%, 87.58%, and 84.41% reaction-compatible yield. Adding the production corrections further improves performance, especially for Ugi. Together with the experiments that disable core pinning, these comparisons identify the contributions of learned conditioning and decoder corrections.

W4 + Q4. Structural novelty and exhaustive catalog escape

Could the authors additionally quantify how structurally different the component-novel precursors are from their nearest training-catalogue components, rather than only reporting whether their identities are absent?

On further examination, we believe the original held-out experiment was not best suited to evaluating open-ended generation. It rewards exact recovery of 68 designated precursor identities, giving no credit for other new, reaction-compatible precursors. The score measures recovery of selected identities, not breadth of chemical exploration.

DeFoG performs better on that recovery task (15.5 versus 1.4 per 1,000 attempts), but this does not establish which model is better at open-ended discovery. We therefore added direct assessments of structural novelty and catalog escape.

We measured ECFP4 Tanimoto distance to the nearest training precursor of the same reaction role (higher indicates greater structural separation).

Ugi precursor roleFORGEDeFoG
Amine0.316 ± 0.0190.312 ± 0.006
Aldehyde0.213 ± 0.0030.173 ± 0.003
Isocyanide0.464 ± 0.0130.215 ± 0.012

Three seeds, 3,072 attempts per seed, with matched decomposition and distance calculations. FORGE showed greater aldehyde and isocyanide distances, with comparable amine distances.

We then exhaustively enumerated 485,625 Ugi precursor combinations under the registered reaction, obtaining 488,250 distinct reachable products. Catalog escapes passed exact verification and structural qualification and fell outside this complete set (Definition C.5).

Qualified Ugi catalog escapes / 1,000 attemptsFORGEDeFoG
Mean ± SD833.3 ± 23.2258.4 ± 22.9

In this separate three-seed Ugi evaluation, FORGE generated 3.2× as many qualified catalog escapes per attempt, beyond every combination in the specified finite catalog.

We also revisited Table 10: its lower descriptor precision (29.5 versus 54.6 for the catalog oracle), fingerprint precision (0.0), and diversity (0.643 versus 0.727 for DeFoG) show weaker matching to one study's held-out lipid distribution, not necessarily implausible lipids or minor training-set variations.

We reanalyzed the same Ugi outputs using shared LNPDB-derived lipid-architecture criteria and catalog escape:

MethodArchitecture-compatible among exact-L1 (%)Compatible catalog escapes / 1,000
FORGE80.45 ± 1.47674.70 ± 16.02
DeFoG71.54 ± 2.31192.49 ± 12.39
Catalog oracle89.13 ± 0.370
Learned inventory selector89.17 ± 1.400

Mean ± SD across three seeds. Inventory methods show higher architecture compatibility but cannot escape their catalogs. FORGE generated 3.5× as many architecture-compatible catalog escapes as DeFoG, expanding beyond the catalog while retaining relevant lipid architecture. We will clarify these distinct evaluation criteria in the revision.

W5 + Q5. Theoretical guarantees and measured performance

What practical insight or guarantee does the theorem 3.1 provide specifically for FORGE?

We incorporate the reaction program by restricting atom, bond, and topology states and fixing the specified core coordinates throughout sampling. In Theorem 3.1, we prove that the resulting flow recovers the intended conditional distribution at the population optimum under exact continuous-time sampling. Proposition C.6 further guarantees that program-fixed coordinates remain unchanged during finite numerical sampling, regardless of predictor accuracy. To evaluate the trained model, we then measured conditioning effects, program adherence, and exhaustive catalog escape. Together, the theoretical analysis and new experiments connect the constrained formulation to practical generation performance.

In vivo expression from generated lipids

To test the generated designs experimentally, we synthesized three FORGE lipids, formulated each as FLuc mRNA-LNPs, and measured reporter expression in mice (Figure 2 and Appendix B.5). We tested the three formulations and MC3 across four intramuscular sites in two mice, with one site per formulation. At 0.05 mg mRNA/kg and four hours, FORGE-3 produced an observed 13.1-fold higher background-normalized signal than MC3 at separate sites in the same mouse. These experiments carry the generated designs through synthesis and formulation to measured biological function in vivo.

Together, the controlled experiments, verified catalog expansion, and in vivo findings strengthen the evidence for FORGE and address the reviewer's concerns. We apologize for the length of the response and thank the reviewer for the constructive feedback. In light of these additions, we respectfully ask the reviewer to consider increasing the score and revisiting confidence in the submission.

Dear Reviewer Stk5,

Thank you again for your thoughtful review and helpful suggestions.

As the author–reviewer discussion period concludes on November 18 (AoE), we wanted to check whether you've had an opportunity to review our detailed responses and new experiments. We hope these additions help resolve your concerns and provide stronger evidence for our contributions. We would greatly appreciate any further feedback and would be happy to address outstanding questions before the deadline.

The Authors

Reviewer_FD56OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the synthesizable molecular generation problem. Typical deep generative models for molecules perform well on de novo molecular generation. Reinforcement learning algorithms could shift the generative distribution toward high-reward molecular spaces (e.g., high binding affinity). However, most of these methods do not consider synthesizability. Several recent approaches have proposed constraining the generative process with predefined molecular building blocks to guarantee a synthetic pathway, yet these methods usually sacrifice generation flexibility and performance.

This paper proposes FORGE for synthesizable and high-performance molecular generation. The authors use predefined reactions as conditions for a generative model (discrete flow matching) and train it with a standard maximum likelihood-style objective on a fixed dataset. The performance seems promising, but the paper does not compare against proper baselines (e.g., RGFN, S3GFN).

Critical strengths

They provided wet-lab validation of three generated ionizable lipids, including in vivo experiments.

Critical weaknesses
  • The method is not principled. There is no explicit training objective for synthesizability constraints. Although Theorem 3.1 provides an asymptotic guarantee for reaction consistency, it relies on strong assumptions (e.g., an unrestricted predictor achieving the population optimum) that do not hold in practice.

  • Missing baselines: For example, S3-GFN [1] should be included as a baseline that incorporates soft-constraint regularization for synthesizability directly into the training objective.


[1] Kim, H., Kim, M., Roget, C., Biton, D., Vaillancourt, L., Brun, Y. V., Bengio, Y., & Hernández-García, A. (2026). Synthesizable Molecular Generation via Soft-constrained GFlowNets with Rich Chemical Priors. Proceedings of the 43rd International Conference on Machine Learning (ICML 2026), PMLR 306, 57977–58001.

Questions for authors
  1. Could the authors clarify what makes the proposed method specific to ionizable lipids? This method seems to be a general method.

  2. Are there any ways to use exact-L1 as a more explicit training signal?

3 comments

We thank the reviewer for the careful assessment and for recognizing the wet-lab and in vivo experiments as a strength. To address the methodological questions, we conducted new verification-guided training experiments and five-seed comparisons with S3-GFN and RGFN. We also synthesized three generated lipids, formulated them as mRNA-LNPs, and measured reporter expression in mice, as reported in the submission.

W1 + Q2. Explicit reaction-verification training objective

There is no explicit training objective for synthesizability constraints.

Are there any ways to use exact-L1 as a more explicit training signal?

We designed FORGE to generate complete molecular structures from a specified reaction program while leaving precursor identities open to de novo generation. During training, we condition the model on reaction annotations. During evaluation, we independently verify the generated structures through exact-L1 forward replay. This links the generative objective to a concrete, independently checkable assembly criterion.

To incorporate the reviewer's suggestion, we added exact-L1 feedback to the training objective and conducted matched experiments within the same FORGE architecture. Specifically, we added a verification term to the flow-matching loss:

\mathcal{L}&=\mathcal{L}_{\mathrm{FM}}+\mathcal{L}_{\mathrm{ver}},\\\\ \mathcal{L}_{\mathrm{ver}}&=-\frac{\beta}{M}\sum_{i=1}^{M}\operatorname{sg}(r_i-b_{-i})\sum_t\log K_{\theta,t}(x_{t+1}^{(i)}\mid x_t^{(i)},c,d). \end{aligned}$$ Here, $M$ is the number of sampled trajectories, $r_i$ is the binary exact-L1 reward, $b_{-i}$ is a leave-one-out baseline, $\beta$ weights verification feedback, and $\operatorname{sg}$ denotes stop-gradient. The learned distribution $K_{\theta,t}$ gives the transition from molecular state $x_t^{(i)}$ to $x_{t+1}^{(i)}$, conditioned on reaction chemistry $c$ and depth $d$, with parameters $\theta$. The verification term directly updates the molecular generation heads using exact forward-replay outcomes. To isolate the effect of verification feedback, we compared the two objectives from **identical starting weights and matched training updates**. We disabled reaction-core pinning in both conditions. | Reaction | Ordinary training (%) | Exact-L1-guided training (%) | Improvement | |---|---:|---:|---:| | Ugi | 44.49 ± 4.40 | **84.02 ± 5.23** | **+39.53 pp** | | Aza-Michael | 64.80 ± 7.86 | **92.58 ± 0.79** | **+27.77 pp** | | Reductive amination | 82.07 ± 2.00 | **92.66 ± 2.33** | **+10.59 pp** | Mean ± SD across five seeds, 512 attempts per seed. Exact-L1 uses all attempts as the denominator. **All five seeds improved across all three reaction families.** By adding exact-L1 feedback, we increased success to **84.02%, 92.58%, and 92.66%**, respectively, with reaction-core pinning disabled. These matched experiments directly answer Q2 and demonstrate that an explicit verification objective improves learned reaction compatibility. **W1. Connection between the objective and theoretical guarantees** In Theorem 3.1, we prove that the constrained flow recovers the intended conditional molecular distribution at the population optimum under exact continuous-time sampling. Proposition C.6 further guarantees that program-fixed coordinates remain unchanged during finite numerical sampling, regardless of predictor accuracy. To assess the implemented model, we then measured the effect of verification feedback in the matched experiments above. Together, the proof and experiments connect the constrained formulation to improved reaction-compatible generation.

W2. New comparisons with synthesis-aware baselines

Missing baselines: For example, S3-GFN [1] should be included as a baseline that incorporates soft-constraint regularization for synthesizability directly into the training objective.

To broaden the comparison as requested, we evaluated S3-GFN [1] and Reaction-GFlowNet (RGFN) [2] with the same reaction verifier. For this new five-seed study, we retained the submitted FORGE architecture, expanded the lipid training data, and ran additional updates. We evaluated each model over 8,192 attempts per reaction per seed.

Exact-L1 per attempt (%)FORGES3-GFNRGFN
Ugi93.6874.3394.67
Aza-Michael92.6678.2148.24
Reductive amination85.8478.3358.76

FORGE achieves higher exact-L1 rates than S3-GFN across all three reactions and RGFN on aza-Michael and reductive amination. To determine how this translates into molecular discovery, we further counted distinct, qualified new products containing at least one novel recovered precursor.

Qualified novel products / 1,000 attemptsFORGES3-GFNRGFN
Ugi780.6478.47476.22
Aza-Michael618.80133.30236.60
Reductive amination389.58190.16488.96

FORGE also exceeds S3-GFN on qualified novel-product yield across all three reactions. For Ugi, FORGE generates 1.64 times as many qualified novel products per attempt as RGFN. FORGE leads RGFN on both measures for aza-Michael. For reductive amination, FORGE achieves higher exact-L1, and RGFN achieves higher qualified novel-product yield.

RGFN assembles molecules from 445 building blocks in this comparison. We designed FORGE to generate complete structures at atom-and-bond resolution with open precursor identities. By measuring both exact-L1 and qualified novel-product yield, we directly evaluate reaction compatibility and the ability to discover new molecular structures.

Q1. Specificity to ionizable lipids

Could the authors clarify what makes the proposed method specific to ionizable lipids? This method seems to be a general method.

We developed FORGE around the chemical requirements of ionizable lipids, which combine ionizable headgroups, linkers, and long hydrophobic tails in large, flexible structures. To coordinate these components, we incorporated lipid-specific information into the training data, reaction programs, architecture controls, and molecular representation:

ComponentIonizable lipid-specific design
Training dataLipid-derived molecular structures annotated with reaction chemistry and precursor information
Reaction programsUgi, aza-Michael, and reductive amination, established routes for modular ionizable lipid synthesis, with explicit headgroup and tail roles and assembly constraints
Molecular architectureControl over headgroup and tail organization, branching, cyclic structures, and attachment positions
Molecular representationSparse atom-and-bond representation supporting large lipid structures of up to 194 atoms

In vivo reporter expression from generated lipids

To evaluate the generated lipids experimentally, we synthesized three FORGE designs, formulated each with FLuc mRNA, and measured reporter expression in vivo (Figure 2 and Appendix B.5). We tested the three formulations and MC3 across four intramuscular injection sites in two mice, with one site per formulation.

At 0.05 mg mRNA/kg and four hours after administration, we measured background-normalized reporter signals of 5.997 × 10⁵, 2.036 × 10⁶, and 2.881 × 10⁶ for FORGE-1, FORGE-2, and FORGE-3, respectively, and 2.195 × 10⁵ for MC3. FORGE-3 produced an observed 13.1-fold higher signal than MC3 at separate injection sites in the same mouse. Through synthesis, formulation, and these site-level measurements, we demonstrated reporter expression in vivo with the generated lipids.

[1] Kim et al. Synthesizable Molecular Generation via Soft-constrained GFlowNets with Rich Chemical Priors. ICML, 2026. https://proceedings.mlr.press/v306/kim26ad.html

[2] Koziarski et al. RGFN: Synthesizable Molecular Generation Using GFlowNets. NeurIPS, 2024. https://doi.org/10.52202/079017-1488

Together, the new training experiments and baseline comparisons address the methodological questions, and the in vivo experiments demonstrate biological function for selected generated candidates. We thank the reviewer again for the constructive suggestions. In light of these findings, we would be grateful if the reviewer would consider increasing the score and revisiting confidence in the submission.

Dear Reviewer FD56,

Thank you again for your careful review and constructive feedback.

With the author–reviewer discussion period concluding on November 18 (AoE), we wanted to kindly check whether you've had a chance to consider our responses and additional experiments. We hope these new results address your concerns and strengthen the case for our work. We would greatly appreciate your thoughts and would be happy to clarify any remaining questions before the deadline.

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response

Dear Area Chair,

Thank you for overseeing our submission. The reviewers' feedback motivated substantial new experiments, summarized below.

Generative lipid design with in vivo validation

Ionizable lipids are essential for RNA medicines, but their discovery remains centered on predictive modeling and screening predefined libraries. To our knowledge, FORGE is among the first models to generate ionizable lipids de novo at full atom-and-bond resolution, and the first such generative model to produce experimentally synthesized ionizable lipids. We not only synthesized three generated lipids but also formulated and characterized their mRNA-LNPs and demonstrated functional RNA delivery in vivo with all three designs.

Additional experiments addressing the reviews

Reviewer FD56: Verification-guided training and synthesis-aware baselines. We added exact-L1 feedback, comparing matched training from identical weights without core pinning. Across five seeds, verified-generation yield improved from 44.5% to 84.0% for Ugi, 64.8% to 92.6% for aza-Michael, and 82.1% to 92.7% for reductive amination, with gains in every seed. Against the requested S3-GFN and RGFN baselines, FORGE achieved the highest average exact-L1 and qualified novel-product yields across the three families, exceeding S3-GFN on both metrics in every family.

Reviewer Stk5: Program adherence, reaction specificity, and open-ended generation. Joint adherence to requested depth, precursor roles, and core positions reached 97.7%, 86.8%, and 80.8% among verified products across the three families.

The original cyclic-label control permuted labels in training and evaluation, making it uninformative for testing reaction specificity. We corrected this by changing only the reaction label at inference, holding the trained model fixed. The corrected test showed reaction-label sensitivity, while full-program tests showed near-zero cross-reaction compatibility.

The held-out benchmark was not informative for comparing open-ended generation between FORGE and DeFoG, crediting only 68 designated precursor molecules rather than all novel, reaction-compatible molecules. To correct this, we therefore enumerated all 488,250 products reachable from the reference Ugi catalog. FORGE generated 833.3 qualified catalog escapes per 1,000 attempts versus 258.4 for DeFoG, a 3.2-fold advantage, demonstrating exploration beyond that finite inventory.

Reviewer kDy2: Production-setting conditioning, lipid architecture, and reproducibility. Under production constraints, chemistry and depth conditioning improved exact-L1 by 18.5 and 19.5 percentage points for aza-Michael and reductive amination. For Ugi, the gain was smaller because the pinned core and decoder already strongly constrain its single-step chemistry. This does not diminish FORGE's contribution, but highlights the complementary roles of learning and structural constraints. Exact-L1 measures reaction compatibility alone. When we also assessed lipid structure using the same criteria, conditioning increased the yield of verified products meeting lipid structural criteria from 72.6% to 93.6% for Ugi (+21.0 pp), with similarly substantial gains for aza-Michael (+23.9 pp) and reductive amination (+19.1 pp). These results show that learned conditioning improves the generation of structurally appropriate ionizable lipids beyond satisfying reaction constraints.

In Ugi, 80.5% of verified FORGE products met the architecture criteria, and FORGE generated 674.7 structurally qualified catalog escapes per 1,000 attempts versus 192.5 for DeFoG, a 3.5-fold advantage under the same criteria.

We resolved the checkpoint discrepancy, reverified baseline, and documented 35 additional training runs. As allowed in rebuttal, more training data in two families and further training allowed us improve our results, but FORGE's original method, architecture, objective, and verifier remain unchanged. Ugi received no new data and thus no performance increase, and all three in vivo-tested lipids were Ugi designs. We reran matched ablations appropriately.

Case for acceptance

To our knowledge, FORGE represents a first-of-its-kind demonstration of de novo generative ionizable lipid design, with generated molecules carried through experimental synthesis, LNP formulation, and functional RNA delivery in vivo. Its reaction-program guidance, independently verified reaction compatibility, and exploration beyond fixed precursor inventories establish a distinct methodological contribution. We believe this combination of generative modeling and prospective experimental validation represents a meaningful advance in molecular design for RNA medicines.

We sincerely appreciate the reviewers' careful assessments. We respectfully ask the Area Chair and reviewers to consider the complete evidence, reconsider the scores, and support acceptance of FORGE.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.