PaperlayerView exact revision on arXiv

Source-linked AI summary

LLM-to-Flow Distillation: Teaching Natural Language Goals to Scientific Generators

Riccardo De Santi, Federico Di Gennaro, Angelo Gnazzo, Sophia Tang, Pranam Chatterjee, Fanny Yang, İsmail İlkan Ceylan, Michael M. Bronstein, Andreas Krause

arXiv:2610.17246v1cs.LG

TL;DR

Scientific generators and LLMs offer complementary capabilities, but aligning a generator to natural-language discovery goals is difficult when proxy rewards are unavailable or unreliable. LFD uses a frozen LLM to construct sub-goals and distill candidate preferences into the generator. Across peptide and small-molecule design, it outperforms reward-guided baselines on independent in silico evaluations.

  • Problem

    Scientific generators lack natural-language grounding, while proxy rewards for experimental discovery goals are often unavailable or unreliable.

    Source: §1 · pp. 1–2

  • Method

    LFD uses a frozen LLM to propose capability-aware intermediate sub-goals, judge current samples, and iteratively distill preferences into a domain-specific generator.

    Source: §3 · pp. 4–5

  • Results

    Across therapeutic peptide and small-molecule design, LFD surpasses state-of-the-art reward-guided baselines on independent in silico evaluations and shows stronger performance on out-of-distribution discovery tasks.

    Source: §7 · p. 11

  • Takeaways & Limitations

    LFD provides a form of generative control that couples high-level LLM scientific knowledge with domain-specific generative priors for discovery tasks expressed in natural language.

    Source: §7 · p. 11

    The practical setting assumes a frozen, API-access LLM and computable observables that provide only a partial representation of each design.

    Source: §2 · p. 3

Abstract

from arXiv · show

Generative discovery in science requires coupling high-level scientific reasoning with the ability to generate samples respecting the structure of a design space. Yet these capabilities remain largely separated: foundation LLMs encode broad chemical and biological knowledge, while domain-specific flow- and diffusion-based generators capture the priors needed to produce valid sequences and geometries. We study how to align such scientific generators to high-level natural-language discovery goals without requiring an explicit proxy reward, which is often unavailable or poorly captured by learned predictors in experimental sciences. We introduce LLM-to-Flow Distillation (LFD), where a frozen LLM acts as an adaptive teacher for a domain-specific generator. Given a goal such as 'generate peptides with promising anticancer activity', the teacher constructs a curriculum of intermediate sub-goals aware of the flow's current capabilities, and progressively directs them toward the original discovery goal. At each round, the LLM judges sampled candidates under the current objective. These judgments induce preferences that are iteratively distilled into the generator through DPO. This procedure provides dense learning signals even when the initial model places negligible probability on designs satisfying the final goal, enabling adaptation toward poorly covered, out-of-distribution regions. Experiments on therapeutic peptides and small-molecule structure design demonstrate that LFD enables a new form of generative control: aligning arbitrary domain-specific scientific generators to scientific discovery tasks expressed via high-level natural-language goals.

1 INTRODUCTION

Scientific discovery needs both high-level scientific reasoning and generators that produce valid designs, but these capabilities are often separated. LFD uses an LLM teacher and a generator-dependent curriculum to align specialized generators to natural-language goals without proxy rewards; the paper reports gains over reward-guided baselines across peptide and molecular design tasks.

1 INTRODUCTION

  • LLMs provide scientific knowledge but may not reliably produce valid designs, while domain-specific generators capture design priors but lack high-level knowledge and natural-language grounding.LLMs can concretize goals and assess candidates using observable structural measurements.Source: §1 · pp. 1–2
  • Existing reward-guided fine-tuning depends on proxy rewards that are often unavailable or unreliable, while discovery goals are commonly expressed in natural language.The paper asks how to align a domain-specific generator directly to such a goal.Source: §1 · p. 2
  • LFD uses an LLM as an adaptive teacher, constructing generator-dependent sub-goals and distilling its candidate judgments into the student to steer it toward the discovery goal.The curriculum addresses cases where the initial generator places negligible probability on designs satisfying the final goal.Source: §1 · p. 2
  • Across therapeutic peptide and molecular design tasks, LFD outperforms state-of-the-art reward-guided baselines on held-out evaluations and adapts to highly out-of-distribution goals.The reported advantages also include reduced proxy overoptimization and interpretable curricula.Source: §1 · p. 2

2 CHALLENGE: ALIGNING FLOW MODELS TO NATURAL-LANGUAGE GOALS

The section formalizes alignment as shifting a generator’s density toward designs satisfying a natural-language goal, under a setting with a frozen LLM and computable observables. It identifies unreliable absolute scores and sparse goal-region samples as challenges, motivating ordinal feedback and objectives that distinguish current samples.

2 CHALLENGE: ALIGNING FLOW MODELS TO NATURAL-LANGUAGE GOALS

  • The practical setting assumes a pre-trained flow, a frozen API-access LLM, and computable observables that represent designs without supplying a scalar proxy reward.For molecules, observables can include SMILES strings and structural descriptors; they provide only a partial representation.Source: §2 · p. 3
  • Aligning a generator to a natural-language goal means redistributing its density to better cover the goal’s discovery region.The problem is framed for flow models but stated as generally applicable to other generators.Source: §2 · pp. 2–3
  • LLM absolute scores can be far from ground truth, but their induced ordering remains informative, with pairwise accuracy consistently above chance across three scientific domains.The method therefore uses goal-conditioned ordinal judgments as supervision.Source: §2 · p. 4
  • When the generator assigns negligible probability to the goal region, its samples may rarely satisfy the goal, leaving direct feedback non-discriminative and without a useful learning signal.The paper notes this regime is common for narrow, out-of-distribution discovery goals.Source: §2 · p. 4
  • These challenges motivate ordinal rather than cardinal feedback and objectives that distinguish among the student’s current samples, implemented through a generator-dependent curriculum of intermediate sub-goals.The curriculum progressively tightens toward the original goal.Source: §2 · p. 4

3 METHOD: LLM-TO-FLOW DISTILLATION

LLM-to-Flow Distillation (LFD) adapts a pretrained flow to natural-language goals through an online curriculum of capability-aware subgoals. Teacher judgments on current samples become preferences that iteratively update the student.

3 METHOD: LLM-TO-FLOW DISTILLATION

  • LFD steers a pretrained flow toward a natural-language goal through an online curriculum of capability-aware subgoals and candidate judgments.The teacher progressively tightens subgoals toward the original goal.Source: §3 · p. 4
  • Each round samples the current generator, proposes an attainable, scientifically connected subgoal, elicits judgments, and updates the student from preferences.The teacher uses observables of current samples to understand the generator’s covered design region.Source: §3 · pp. 4–5
  • A subgoal is accepted only when it yields enough positive and negative current-batch candidates to discriminate, and the teacher may abstain on ambiguous designs.Source: §3 · p. 5
  • Positive-versus-negative judgments, excluding abstentions, form preference datasets; one-pass batch classification avoids the quadratic API cost of pairwise duels.Source: §3 · p. 5
  • LFD updates the student with teacher-derived dDPO preferences, resampling each round and using the current model as reference to regularize updates.The paper notes that terminal log-densities are intractable and optimizes a denoising-loss surrogate.Source: §3 · p. 5
  • LFD relies on ordinal judgments as usable adaptation signals; Figure 3 compares four elicitation mechanisms by pairwise accuracy against a 50% reference.Source: §3 · p. 5, Figure 3 · p. 6

4 HOW RELIABLE IS SCIENTIFIC FEEDBACK FROM LLMS?

Across three scientific domains, LLM ordinal judgments were more informative about relative than absolute design quality. All four tested feedback mechanisms achieved above-chance pairwise accuracy, and batch classification was chosen for preference supervision.

4 HOW RELIABLE IS SCIENTIFIC FEEDBACK FROM LLMS?

  • The evaluation tested three API-accessible LLMs on peptide hydrophobic moment, relaxed crystal energy above the convex hull, and QM9 MolLogP, with ground truth hidden during judging.Source: §4 · p. 6
  • All four ordinal-feedback mechanisms achieved above-chance pairwise accuracy despite unreliable raw scores; linear-cost methods matched or exceeded quadratic-cost duels.Source: §4 · p. 6
  • Batch classification was selected for LFD preference supervision because it scales linearly and matches or exceeds duel accuracy.Source: §4 · p. 6
  • LLM scientific judgments tended to be more reliable in relative than absolute terms.Source: §4 · p. 6

5 EXPERIMENTAL EVALUATION OF LLM-TO-FLOW DISTILLATION

Across peptide and small-molecule design tasks, LFD adapts generators using LLM judgments and intermediate sub-goals rather than task-specific proxy rewards. It reports strong held-out activity and docking results, a hit for a sparsely covered molecular goal, and interpretable traces for human oversight.

5.1 CASE STUDY I: THERAPEUTIC PEPTIDE DESIGN

  • For anticancer peptides, LFD achieves the highest held-out AntiCP2-alternate predictor score (53%) with the lowest marginal toxicity increase (42%).The teacher uses sequence and generic physicochemical descriptors; the baselines directly optimize the ACP predictor.Source: Table 1 · p. 7, §5.1 · p. 7
  • LFD exposes interpretable sub-goals and candidate judgments: scientists can inspect the curriculum and traces and intervene during adaptation.The curriculum progresses from positive charge to hydrophobic residues and aromatic interfacial anchors; a rejected peptide is flagged for a prolonged hydrophobic stretch, with hemolysis and aggregation measurements unresolved.Source: §5.1 · p. 7, Figure 5 · p. 7, Figure 4 · p. 7
  • For anticancer design, GUIDANCE saturates its optimized ACP proxy at 100% but scores only 35.94% on held-out AntiCP2-alt; LFD never optimizes either anticancer predictor.LFD achieves the best held-out score using LLM judgments, consistent with the reported concern that proxy optimization can produce reward hacking.Source: §5.1 · pp. 7–8

5.2 CASE STUDY II: DE NOVO DESIGN OF SMALL MOLECULES

  • Across GSK3β docking, LFD leads all four baselines on held-out evaluators, reaching 6.77 pK GNINA CNNaffinity, −8.84 kcal/mol Vina, and −12.87 rDock.It preserves 96% RDKit validity and 84% PoseBuster conformer quality.Source: Table 2 · p. 8, §5.2 · p. 9
  • On D2 docking, LFD leads on all three evaluators: 6.54 pK GNINA CNNaffinity, −8.48 kcal/mol Vina, and −9.89 rDock.It retains 98% validity and 89.5% PoseBusters conformer quality; Figure 7 reports the strongest top-sample quality/diversity trade-off.Source: Table 2 · p. 8, §5.2 · pp. 8–9, Figure 7 · p. 9

5.1 CASE STUDY I: THERAPEUTIC PEPTIDE DESIGN

  • For cell-penetrating peptides, LFD raises pLM4CPPs from 8.81 to 69.06 versus 57.04 for the best baseline, while lowering HemoPI2 toxicity from 40.84 to 37.02.The LLM supervises adaptation without the pLM4CPPs evaluator or a proxy reward.Source: §5.1 · p. 8

5.2 CASE STUDY II: DE NOVO DESIGN OF SMALL MOLECULES

  • For the complex kinase-focused molecule goal, LFD produces a full-goal hit in a 512-sample batch at round 6, while the pretrained model yields none in 38,500 samples and reward-guided baselines fail.The baselines receive an LLM-generated binary reward, but positive examples are initially too sparse to provide useful learning signal.Source: §5.2 · pp. 9–10, Figure 9 · p. 9

6 RELATED WORKS

Prior flow and diffusion adaptation methods generally assume a specified reward, supervision signal, or distributional utility. LFD instead starts from a natural-language scientific objective and uses generator-dependent sub-goals and ordinal judgments to adapt a separate domain-specific generator.

6 RELATED WORKS

  • Reward-guided and distributional flow or diffusion methods optimize specified rewards, preferences, constraints, or distributional utilities, whereas LFD starts from a scientific objective expressed only in natural language.Examples of prior approaches include expected-reward, entropy, CVaR, and constrained adaptation methods.Source: §6 · p. 10
  • Unlike LLM-based candidate editing or optimization, LFD fine-tunes a separate domain-specific generator from ordinal judgments and does not have the LLM generate candidate designs.This separation also distinguishes LFD from on-policy distillation approaches that require teacher and student to share an output space.Source: §6 · p. 10
  • Unlike preference methods with a fixed goal, LFD generates sub-goals online from the current generator’s distribution and distills preferences that distinguish its samples.Its progression is not derived from a specified reward, verifier, executable task, or predefined score threshold.Source: §6 · pp. 10–11
  • LFD shares adaptive progression with curricula and rare-event sampling, but uses generator-dependent natural-language scientific sub-goals rather than predefined thresholds or executable environments.Related methods select tasks or regions according to learner capabilities or current samples.Source: §6 · pp. 10–11

7 CONCLUSION

LFD adapts domain-specific generators to natural-language scientific discovery goals without task-specific proxy rewards. Across therapeutic peptides and small-molecule design, it outperforms reward-guided baselines on independent in silico evaluations and shows stronger performance on OOD discovery tasks.

7 CONCLUSION

  • Across therapeutic-peptide and small-molecule design tasks, LFD surpasses state-of-the-art reward-guided baselines on independent in silico evaluations.The conclusion also reports higher interpretability and stronger performance on OOD discovery tasks.Source: §7 · p. 11
  • LFD adapts domain-specific generators to natural-language scientific discovery goals without requiring task-specific proxy rewards.The paper presents it as generative control that couples LLM scientific knowledge with domain-specific generative priors.Source: §7 · p. 11
  • Integrating wet-lab feedback into LFD is identified as a natural next step.Source: §7 · p. 11

AI USE STATEMENT

The authors report using AI tools during manuscript development and retain responsibility for the final content, including AI-assisted artifacts.

AI USE STATEMENT

  • The authors used AI tools to generate code, analyze and browse existing literature, and improve the manuscript exposition.Examples of exposition work include rephrasing sentences and modifying figures.Source: AI USE STATEMENT · p. 11
  • The authors take full responsibility for the final content, including text, claims, and artifacts produced with AI assistance.Source: AI USE STATEMENT · p. 11
  • AI-assisted work covered both research-support tasks and manuscript presentation.Source: AI USE STATEMENT · p. 11

A LLM-AS-AN-EXPERT: DOMAINS AND ELICITATION PROTOCOL

This appendix evaluates LLMs as scientific experts and describes the LFD protocol. It reports that cardinal score magnitudes are poorly calibrated, while ordinal elicitation strategies can provide useful preferences more efficiently than exhaustive pairwise comparisons.

A LLM-AS-AN-EXPERT: DOMAINS AND ELICITATION PROTOCOL

  • LLM-as-an-Expert evaluates candidate designs on domain-specific quantities from their representations and the model’s scientific prior, without explicit predictors.The evaluation spans molecules, inorganic crystals, and peptides, with ground-truth scalar values hidden from the models.Source: A LLM-AS-AN-EXPERT: DOMAINS AND… · p. 19
  • Across domains, 1–100 LLM scores frequently fail to reproduce the proportions of ground-truth differences, so they should not be treated as calibrated property-difference estimates.Their induced candidate ordering may still be informative.Source: A LLM-AS-AN-EXPERT: DOMAINS AND… · p. 19
  • Cheaper score, ranking, and set-classification strategies can match or sometimes exceed pairwise-duel accuracy, while producing many preferences from few listwise API calls.Exhaustively querying every duel scales quadratically with sample count.Source: A LLM-AS-AN-EXPERT: DOMAINS AND… · p. 19
  • The elicitation study compares independent scores, listwise rankings, positive/negative/abstain classifications, and pairwise duels using ground-truth pairwise accuracy and preference coverage.Source: A LLM-AS-AN-EXPERT: DOMAINS AND… · p. 19

B.1 LFD PROTOCOL

  • Each LFD round samples from the current generator, selects or tightens a goal, forms preferences from teacher judgments, and updates the generator with DPO.The process repeats until the final-goal stopping criterion or another stopping condition applies.Source: §B.1 · p. 22

B.2 IMPLEMENTATION DETAILS OF SETGOAL AND LLM-JUDGE ORACLES

The implementation uses an adaptive sub-goal and categorical-judgment process; the accompanying derivation characterizes a population target under assumptions. That single-round target is not a guarantee of the finite-sample update or full curriculum success.

B.2 IMPLEMENTATION DETAILS OF SETGOAL AND LLM-JUDGE ORACLES

  • The SETGOAL implementation proposes a learnable sub-goal connected to the final goal and plausible as progress, then independently confirms the proposal.The sub-goal is selected using the current batch, its observables, the final goal, and the previous intermediate goal.Source: §B.2 · pp. 22–23
  • The protocol treats sub-goals as revisable steps toward the final goal and keeps anticipated later rungs out of current judgments and training.Later rungs are advisory; tightening retains current requirements and makes a small semantic step toward the final goal.Source: §B.2 · pp. 22–23
  • The judge evaluates samples twice under the current sub-goal, and only stable positive and negative labels enter the preference sets; abstained or unstable samples are excluded.Abstained and unstable samples remain in the denominator when reporting label frequencies.Source: §B.2 · p. 24

C KL-REGULARIZED STEP TO LLM TEACHER-INDUCED IMPLICIT TARGET

  • Under a Bradley–Terry preference assumption and connected comparison paths, the derivation characterizes a KL-regularized population target for one preference update.The analysis concerns ordinal preferences and does not query or directly fit the latent utility.Source: C KL-REGULARIZED STEP TO LLM… · p. 25
  • The single-round derivation does not establish the exact finite-sample update or guarantee that the complete curriculum reaches the discovery region.Finite preference outcomes, model representational limits, optimization budget, and the denoising-loss surrogate can separate the practical update from the population target.Source: C KL-REGULARIZED STEP TO LLM… · p. 25

D EXPERIMENTAL DETAILS

The appendix documents the experimental setup for therapeutic-peptide design and the LFD configuration. It specifies a staged curriculum in which generator updates precede small, reachability-informed goal changes.

D EXPERIMENTAL DETAILS

  • The appendix reports generator choices, candidate representations, evaluation protocols, adaptation settings, and campaign execution; prompts and instantiation details appear separately.PepDFM is the pre-trained peptide sequence generator, and the anticancer goal favors plausible cancer-cell activity while minimizing nonspecific toxicity to healthy human cells.Source: D EXPERIMENTAL DETAILS · p. 26, §D.1 · p. 26, §D.1.1 · p. 26

LFD Configuration

  • The selected PepDFM checkpoint follows eleven accepted DPO updates across seven intermediate goals and two updates on the final goal, using 100-peptide cohorts and 100 sampling steps.Each update uses all stable-positive/stable-negative pairs and a frozen parent reference; stable labels come from two judgment passes, with abstentions and disagreements retained in the denominator.Source: LFD Configuration · pp. 26, 29
  • LFD alternates between adopting a reachable temporary goal and training the generator on categorical positive and negative examples to enrich that goal.The temporary goals are nested relaxations of the final objective; the broadest comes first and the final goal is narrowest.Source: LFD Configuration · pp. 26–27
  • Move toward the final goal only after the generator produces substantially more candidates satisfying the current goal; otherwise, hold that goal fixed and train again.A new goal must be newly reachable and discriminative on the updated generator while having been harder at the prior goal’s adoption.Source: LFD Configuration · pp. 26–27
  • The controller broadens first with a weak, meaningful rung, then tightens by restoring one deferred consideration at a time without claiming assay-level success.Proposals are revised using categorical counts as reachability feedback, not cardinal rewards or quotas; the auditor checks semantic nesting and rejects unsupported proxies as necessary biological conditions.Source: LFD Configuration · pp. 27–28
  • Judgments use the complete peptide sequence and generic observables, while the four developability bands are relative, uncertain predictors rather than measured activity, toxicity, or clinical probabilities.No labelled peptide dataset, target-specific anticancer predictor, or biological validation is supplied to the judge.Source: LFD Configuration · p. 29

Sample LLM-Judge Traces

The traces show how the LLM judge distinguishes peptide candidates using sequence organization and uncertain developability evidence. The experimental configuration also records the goal-2 curriculum and baseline settings.

Sample LLM-Judge Traces

  • A positive anticancer trace cites a hydrophobic N-terminal and cationic/aromatic C-terminal organization, while identifying the concentration window for cancer selectivity as unknown.The rationale also notes upper-third non-hemolysis and solubility bands.Source: Sample LLM-Judge Traces · p. 30
  • A negative anticancer trace cites high aromatic residue density and lower-third solubility and non-fouling bands as risks, while leaving cancer-cell versus healthy-fibroblast assays unresolved.The candidate has net charge +6.0 and aromatic residue density 37.5%.Source: Sample LLM-Judge Traces · pp. 30–31

Baselines configuration

LFD Configuration

  • For that campaign, the selected checkpoint follows three updates on G01, two on G02, and two on the final goal, totaling 700 optimizer steps.Each 100-peptide cohort uses the 35/40/25 length mixture at lengths 12/24/38, and stable labels come from two categorical judgment passes.Source: LFD Configuration · p. 31

D.1.2 GOAL 2: CELL-PENETRATING-PEPTIDE DESIGN

  • The cell-penetrating-peptide goal is reached through two temporary goals—membrane interaction and then non-disruptive interaction—followed by two rounds on the final goal.The final goal seeks promising mammalian-cell-penetrating activity while minimizing toxicity to mammalian cells.Source: LFD Configuration · p. 35

Sample LLM judge traces

For dopamine D2 receptor design, the section describes an LFD curriculum that moves from reachable precursor goals toward the final goal, alongside its judging setup and evaluation procedures. It also reports example peptide judgments and baseline and compute configurations.

Sample LLM judge traces

  • The peptide traces show a positive judgment for a +7-charge, 12-residue sequence and a negative judgment for a sequence whose hydrophobicity raised concerns about cytotoxicity and membrane lysis.The positive trace notes cellular uptake and proteolysis as unknowns; the negative trace notes hemolytic threshold and cell viability as unknowns.Source: Sample LLM judge traces · p. 35

Baselines configuration

  • Baseline comparisons use 100 samples per method with a shared 35/40/25 mixture at lengths 12/24/38, while the baselines optimize a CellPPD-label random-forest reward.The downstream pLM4CPPs and HemoPI2 predictors were held out from optimization, but shared historical labels mean the reward and CPP evaluator are not data-independent; the comparison uses one training seed and unequal optimization budgets.Source: Baselines configuration · p. 35

Hardware, computational cost, runtime, economic cost

  • The reported setup used one NVIDIA GH200 GPU with 120 GB and approximately 5.3 allocated hours; recorded API cost was approximately $8.83.The hardware and cost are reported for this experiment’s setup.Source: Hardware, computational cost, runtime,… · p. 36

D.2 MOLECULAR DESIGN

  • Novel-scaffold counts use eligible molecules in each seed’s top 5% by rDock score and count each distinct scaffold absent from the training reference once.Across five sampling seeds, the reported count is the mean.Source: §D.2 · p. 36

D.2.1 GOAL 3: MOLECULAR DESIGN FOR DOPAMINE D2 RECEPTOR DOCKING

  • The final goal is promising drug-like molecules for dopamine D2 receptor binding, balancing plausible receptor recognition, moderate lipophilicity, and chemical viability.Docking scores and target-specific quantities are computed after generation and hidden from the LLM teacher.Source: §D.2.1 · p. 36, §D.2 · p. 36

LFD Configuration

  • The teacher makes categorical judgments from molecule cards, using registered molecular observables and deferring final-goal requirements when proposing temporary goals.Proposals are audited for semantic movement and rejected if they rely on numerical descriptor cutoffs, hidden evidence, or claims of experimental activity.Source: LFD Configuration · pp. 37–38

LFD curriculum

  • The DRD2 curriculum progresses from an aminergic scaffold goal (g0, five rounds), to a two-aromatic-ring goal (g1, one round), then the final receptor-design goal (three rounds).The protocol alternates goal movement with generator fine-tuning, moving toward the final goal only after enriching the current goal.Source: LFD Configuration · p. 37, LFD curriculum · p. 38

LFD sample traces

For GSK3β ATP-pocket design, LFD uses temporary goals and categorical molecule-card judgments, with an example positive trace that identifies structural support while noting unresolved binding questions. The section also specifies reward-based baselines and a separate kinase-focused design goal.

D.2.2 GOAL 4: MOLECULAR DESIGN FOR GSK3β ATP-POCKET BINDING

  • The GSK3β final goal combines ATP-pocket binding with reasonable aqueous solubility and chemically plausible structures; the reported curriculum uses four temporary goals before two rounds on the final goal.The listed temporary goals progress from an aromatic ring and hydrogen-bond donor to a heteroaromatic ring, structural-alert avoidance, and solubility-related constraints.Source: §D.2.2 · p. 39, LFD Configuration · p. 41

LFD Configuration

  • LFD judgments use categorical labels and registered molecule-card evidence; the protocol treats missing assays or docking evidence as unknown rather than negative.SMILES or local geometry can support a structural hypothesis but cannot by itself prove GSK3β binding, affinity, solubility, selectivity, safety, or experimental activity.Source: LFD Configuration · pp. 40–41

LFD sample traces

  • The positive GSK3β trace identifies an aminoazine donor–acceptor motif and balanced descriptors as supporting a candidate, while pose, affinity, selectivity, and permeability remain unknown.The negative trace cites a hydrolytically labile exocyclic imine and an intramolecular clash as plausibility liabilities; stability and aqueous-buffer half-life remain unknown.Source: LFD Configuration · pp. 41–42

Baselines configuration

  • The GSK3β baselines use a frozen TDC random-forest probability reward that LFD does not observe, with 100 optimizer steps per round for trained policies.FlowGRPO uses a flow-matching log-density surrogate rather than exact trajectory likelihoods, while guidance fits an 80/20-split reward surrogate.Source: LFD Configuration · p. 42

D.2.3 GOAL 5: MOLECULAR DESIGN FOR KINASE-FOCUSED CAMPAIGN

  • The separate kinase-focused goal requires a neutral, connected molecule with an amino-substituted six-membered aza-aromatic ring and a distinct saturated ring containing a free secondary amine.The saturated amine-containing ring need not be carbon-linked to the aza-aromatic ring.Source: §D.2.3 · p. 42

LFD Configuration

The configuration describes LFD as alternating generator updates on reachable temporary goals with gradual movement toward a final goal. It also gives categorical judgment constraints, a structural hit example, and the reported compute limit.

LFD Configuration

  • LFD alternates goal movement and FlowMol fine-tuning, using nested positive sets and moving one semantic step toward the final goal only after the current goal passes its threshold.The first temporary goal should be a discriminative, assessable precursor; later tightening restores one deferred final-goal consideration, while a carbon-only linkage requirement is explicitly excluded.Source: LFD Configuration · pp. 42–43
  • The round-6 positive example satisfies the stated structural clauses with a 2-amino-1,3,5-triazine motif and a distinct diazepane ring containing a free secondary amine.The accompanying negative example lacks a primary amine directly attached to the aza-aromatic ring and lacks a saturated secondary-amine ring; kinase affinity and cellular activity are unmeasured.Source: LFD Configuration · pp. 44–45

Hardware, computational cost, runtime, economic cost

Related papers

Loading 2610.17246v1…