PaperlayerView exact revision on arXiv

Source-linked AI summary

Language-Informed Flow Matching for Trend-Guided Structure-Based 3D Molecular Generation

Tianyu Gao, Zhikai Su, Jiashu Li, Wenjun Gao, Zichuan Ying, Zhe Zhao, Fei Zhang, Ye Wei

arXiv:2608.31009v1cs.LG

TL;DR

SBDD requires molecules that are both chemically valid and geometrically compatible with protein pockets, while existing controllable methods may require fine-tuning or external guidance. LiFT uses language-derived SMILES priors, a frozen chemical encoder, and state-aware routing within Flow Matching to guide 3D generation. On CrossDocked2020, it maintains competitive distribution matching and structural validity while improving medicinal chemistry profiles under task steering without generator fine-tuning.

  • Problem

    SBDD must jointly satisfy 3D pocket compatibility and 1D chemical validity, while existing 1D–3D methods do not fully translate task preferences, pocket context, and property trade-offs into semantic conditions.

  • Method

    LiFT converts pocket context and design preferences into target-aware SMILES, encodes them as continuous semantic priors, and uses zero-initialized conditioning plus SCDR to guide Flow Matching.

  • Results

    LiFT achieves competitive distribution matching and structural validity while providing effective trend steering and improving medicinal chemistry profiles across de novo design and scaffold hopping without additional generator fine-tuning.

  • Takeaways & Limitations

    Language-derived chemical priors can serve as semantic conditions for pocket-conditioned 3D generation, providing trend-level guidance without geometric-backbone fine-tuning.

  • Takeaways & Limitations

    LiFT provides trend-level rather than exact property or structural control, and its performance may vary with the underlying LLM’s chemical competence and reliability.

Abstract

from arXiv · show

Structure-based drug design (SBDD) requires ligands that satisfy both 3D target affinity and 1D chemical validity. Existing controllable generation methods often rely on task-specific fine-tuning or externally imposed sampling-time guidance, adding cost and potentially conflicting with evolving 3D geometric constraints. We propose LiFT, a language-informed cross-modal framework built on Flow Matching for trend-guided 3D molecular generation across both de novo design and scaffold hopping. LiFT uses a "Sense-Evolve-Assemble" agent to generate target-aware SMILES as intermediate chemical conditions, from which a pre-trained chemical foundation model extracts continuous semantic priors. These priors are integrated into geometric generation through a lightweight semantic projector with zero-initialized adaptive normalization for stable cross-modal conditioning. We further introduce a Self-Conditioned Decoupled Router (SCDR), which modulates the velocity field according to intermediate structural states during ODE integration. Experiments on Cross-Docked2020 show that LiFT achieves competitive distribution matching while improving medicinal chemistry metrics and maintaining competitive structural validity under task-steering settings without additional generator fine-tuning. Our results suggest that language-derived chemical priors provide effective trend-level guidance for 3D molecular generation. Code and released artifacts are available at https://github.com/kasurl/LiFT.

1 Introduction

LiFT addresses the SBDD challenge of jointly satisfying pocket-level 3D compatibility and 1D chemical validity by using language-derived chemical priors to guide target-aware flow-based generation. It supports trend steering across de novo design and scaffold hopping without additional generator fine-tuning.

  • SBDD must balance 3D physicochemical complementarity in protein pockets with 1D chemical validity across a vast molecular space.
  • Existing 1D–3D generators connect molecular strings to geometric models but do not fully translate task preferences, pocket context, and property trade-offs into semantic conditions.
  • LiFT converts pocket information, task preferences, and optional references into target-aware SMILES, then uses a frozen chemical encoder to produce semantic latents for geometric generation.
  • LiFT provides competitive distribution matching and structural validity while enabling effective trend steering across de novo design and scaffold hopping without additional generator fine-tuning.
  • SCDR adapts the influence of language-derived priors to intermediate spatial states while coordinating chemical trends with evolving geometric constraints.

2 Related Work

Related work spans geometry-centered 3D generation, sequence-based chemical modeling, and cross-modal conditioning. LiFT builds on these directions by using frozen chemical representations and state-aware modulation within flow-based generation.

  • 3D SBDD models operate directly on atomic coordinates using autoregressive, diffusion-based, and continuous-time Flow Matching approaches.
  • Sequence-based molecular models capture chemical syntax and topology naturally, but connect less directly to target-specific pocket geometry.
  • Cross-modal generators commonly use cross-attention, graph-language interactions, or FiLM- and AdaLN-style modulation to incorporate external information.
  • LiFT uses zero-initialized lightweight modulation to integrate SMILES-derived priors into flow generation with limited architectural overhead.
  • Frozen chemical foundation models provide continuous SMILES representations that LiFT uses as soft chemical priors rather than exact structural targets.
  • Because fixed semantic signals may vary in usefulness during ODE generation, LiFT adapts their influence to intermediate ligand states while preserving the geometric backbone.

3 Methodology

LiFT combines target-aware SMILES generation, continuous semantic encoding, lightweight scalar conditioning, and state-aware routing within a pocket-conditioned Flow Matching model. SCDR uses invariant summaries of evolving ligand states to regulate semantic and geometric updates while preserving equivariant geometry.

  • 3 Methodology: LiFT transports prior noise toward the empirical ligand distribution conditioned on a 3D protein pocket through four stages: SMILES generation, semantic extraction, projection, and state-aware velocity modulation.
  • Sense-Evolve-Assemble LLM Agent: The Sense-Evolve-Assemble agent constructs a pocket profile, performs Pocket-of-Thought reasoning for de novo or scaffold-hopping proposals, and sanitizes the resulting SMILES.
  • Sense-Evolve-Assemble LLM Agent: The Sense phase analyzes an 8.0 Å pocket shell for geometric and physicochemical cues, while Evolve incorporates task-specific preferences and optional references.
  • Semantic Latent Extraction: SMI-TED, pretrained on 91M molecules, maps sanitized SMILES into a continuous semantic vector used as a compact cross-modal prior for the 3D velocity field.
  • Cross-Modal Priming: Semantic conditioning enters only the scalar domain, leaving equivariant vector features unchanged so geometric update rules remain intact.
  • Cross-Modal Priming: Zero-initialized adaptive normalization initially preserves the original scalar pathway and introduces semantic conditioning through learned deviations.
  • Self-Conditioned Decoupled Router: SCDR fuses semantic, temporal, and invariant structural information, then routes it into state calibration and decoupled scalar and vector update pathways.
  • Self-Conditioned Decoupled Router: The vector state bypasses modulation while scalar calibration is bounded within [0.5, 1.5], preserving pocket-centered SO(3) equivariance during generation.

4 Experiments

LiFT is evaluated on CrossDocked2020 for distribution matching, property-oriented generation, ablations, and language-condition diagnostics. It provides competitive distribution matching while steering medicinal-chemistry properties and retaining structural validity without additional generator fine-tuning.

  • Evaluation setup: Experiments use refined CrossDocked2020 with 100K complexes and compare LiFT against autoregressive, diffusion, flow-matching, and LLM-based SBDD methods.Evaluation covers binding, medicinal-chemistry, topology, validity, and filter-compliance metrics.
  • Evaluation setup: Reference-free and reference-guided variants are evaluated against the empirical ligand distribution, alongside property-oriented prompting variants.Reference-free settings use pocket-derived specifications and task preferences; reference-guided settings use a scaffold-hopping ligand SMILES.
  • Distribution matching: LiFT achieves competitive Wasserstein Distance performance, while reference-free variants outperform the ligand-reference variant on several binding-efficiency and ring-distribution metrics.Higher QED and SA WD can reflect improved pharmacological profiles that shift generated molecules away from the training distribution.
  • Property-oriented generation: Vina-oriented prompts bias generation toward higher binding affinity, whereas QED-centric prompts improve drug-likeness while preserving competitive binding and structural-validity indicators.The conditioning functions as trend-guided rather than deterministic optimization.
  • Ablations and diagnostics: Removing zero-initialization slightly weakens several metrics, while removing SCDR causes larger drops in QED, RDKit, and REOS.The ablations associate zero-initialized modulation with stable prior injection and SCDR with drug-likeness and filter compliance.
  • Ablations and diagnostics: GPT-4o, Claude-4-Sonnet, and DeepSeek-V3 show relatively small performance variation, with GPT-4o selected for QED and DeepSeek-V3 strongest on filter compliance.GPT-4o is used as the default backbone.
  • Ablations and diagnostics: Language-condition analyses compare random, retrieved, prompted, ablated, and full Sense-Evolve-Assemble conditions for semantic effects under a fixed 3D generator.They examine property correspondence and semantic preservation after cross-modal projection.

5 Conclusion

LiFT uses language-derived chemical priors as semantic conditions for pocket-conditioned 3D generation. The study reports that reference-free prompting supports valid chemical exploration and property steering while maintaining competitive binding and structural validity.

  • 5 Conclusion: LiFT introduces a flow-matching framework that uses language-derived molecular conditions for trend-guided 3D generation.The framework targets pocket-conditioned generation across the study’s evaluated settings.
  • 5 Conclusion: Strict ligand-reference anchoring can constrain exploration, whereas reference-free de novo prompting enables exploration of valid chemical regions.This conclusion is reported alongside property-oriented prompting results.
  • 5 Conclusion: Property-oriented prompting improves medicinal-chemistry profiles while maintaining competitive binding and structural validity without geometric-backbone fine-tuning.The conclusion frames language-derived priors as semantic conditions rather than final molecule proposals.

Limitations

LiFT provides trend-level language conditioning but remains limited in geometric precision, dependence on language-model quality, evaluation consensus, and benchmark generalization.

  • Method scope: LiFT provides trend-level guidance rather than exact property or structural control because natural language lacks atomic-level geometric granularity.The authors identify intermediate representations as a future route for bridging language intent and evolving 3D states.
  • Method scope: Performance may vary with the chemical competence and reliability of the underlying LLM, especially where text-based chemical knowledge is sparse or biased.This limitation concerns LiFT’s reliance on LLM-derived symbolic priors.
  • Evaluation scope: SBDD evaluation lacks consensus on balancing distributional fidelity with absolute property-oriented improvement.The authors specifically contrast distributional measures such as Wasserstein Distance with property gains.
  • Evaluation scope: CrossDocked2020 enables controlled comparison but does not establish generalization across broader pocket distributions or capture real biological dynamics.The authors call for validation on additional targets, datasets, and wet-lab feedback.
  • Supporting analyses: The supplementary material expands diagnostics, reproducibility details, prompting protocols, validity analyses, mathematical formulations, and related-work comparisons.These appendices support analysis and reproducibility but do not remove the stated scope limitations.

A Extended Analysis of Language-Derived Molecular Conditions

The analysis tests whether different SMILES-condition sources produce chemically meaningful intermediates and distinct downstream 3D behavior under a shared generation pipeline. The full Sense-Evolve-Assemble agent provides balanced condition quality and downstream trade-offs, while arbitrary or retrieved priors are less effective for target-aware guidance.

  • Evaluation setup: The study compares six SMILES-condition sources before and after 3D generation while holding the downstream encoder, generator, sampling, and evaluation pipeline fixed.The sources range from random and retrieved training SMILES to direct prompting, ablations, and the full Sense-Evolve-Assemble agent.
  • Condition sources: The six sources span random valid SMILES, retrieved training molecules, direct prompting, removal of pocket sensing, removal of Pocket-of-Thought reasoning, and the full agent.C0–C5 isolate generic validity, retrieval, direct prompting, pocket sensing, structured reasoning, and their combination.
  • 1D condition quality: Validity is saturated across variants, while retrieved conditions have zero scaffold novelty and lower scaffold diversity than generated conditions.The comparison therefore distinguishes chemically valid but non-novel retrieval from more diverse generated molecular priors.
  • 1D condition quality: The full agent achieves the highest uniqueness, high scaffold uniqueness, strong scaffold novelty, and a balanced property profile without collapsing to retrieved training-set scaffolds.Direct prompting instead achieves strong QED and SA but lower scaffold uniqueness and simpler ring profiles.
  • Downstream 3D effects: Random SMILES produce the weakest downstream performance, whereas retrieved SMILES improve ring-frequency statistics but do not provide the strongest binding-related guidance.The results indicate that injecting any valid molecular string is insufficient and that reusing training-set priors is not optimal for target-aware generation.
  • Downstream 3D effects: The full agent achieves the best Gnina efficiency and highest PoseBusters interaction-energy validity while maintaining strong QED, SA, and ring-frequency statistics.It also better preserves topological richness than Direct LLM and w/o PoT, and avoids empirical-prior collapse relative to Retrieved SMILES.
  • Trend correspondence: Across more than 6,000 ligands from 100 CrossDocked2020 pockets, generated ensembles show consistently positive correlations between text-derived targets and final molecular properties.The analysis evaluates whether property tendencies in intermediate SMILES remain visible after 3D generation.

B.2 Topo-Chemical Distribution of Agent-Generated SMILES

The appendix examines whether agent-generated SMILES occupy chemically structured regions and how SCDR gates vary during generation. Reference-guided conditions broadly overlap empirical ligand fingerprints, while reference-free conditions shift but remain near them; gate behavior is heterogeneous and trajectory-dependent.

  • Topo-chemical distribution: The analysis compares agent-generated SMILES with 100 held-out ligand SMILES using 2,048-bit Morgan fingerprints projected into two dimensions with t-SNE.The visualization is intended to assess relative overlap, separation, and neighborhood structure within each panel.
  • Topo-chemical distribution: Reference-guided LiFT produces SMILES whose ECFP4 embeddings broadly overlap the ground-truth ligand distribution.The passage attributes this overlap to the explicit molecular anchor provided in the reference-guided setting.
  • Topo-chemical distribution: Reference-free LiFT shows a clearer distributional shift but remains near the empirical fingerprint distribution rather than forming a detached cluster.The shift is consistent with greater freedom to propose alternative scaffolds without a reference molecule.
  • Interpretation: The t-SNE analysis provides a qualitative check of chemical structure in the intermediate conditions, not proof of exact ligand recovery or chemically quantitative distances.Absolute coordinates should not be compared across panels because t-SNE depends on the stochastic full input matrix.
  • SCDR diagnostics: Gate distributions vary across latent channels and target pockets rather than collapsing to uniform modulation.State calibration and scalar-update distributions are heterogeneous, with pocket-dependent differences that are strongest in intermediate trajectory stages.
  • SCDR diagnostics: SCDR diagnostics show distinct gate evolution during training and ODE inference, with scalar and vector updates following different temporal trends.Scalar updates are stronger earlier and relax later, while vector updates rise mid-trajectory before declining near the end.
  • Overall interpretation: Together, the property, fingerprint, and gate analyses support interpreting language-derived conditions as trend-level signals in the 3D generation process.This conclusion remains diagnostic rather than a claim of exact chemical recovery or a specific physical mechanism.

C.4 Implementation and Hyperparameters

LiFT combines a frozen SMI-TED semantic encoder with a heterogeneous GVP-GNN flow-matching generator and SCDR/AdaLN conditioning. The implementation reports runtime, uncertainty, docking-score, and information-boundary procedures for reproducible evaluation.

  • Implementation: LiFT uses a heterogeneous GVP-GNN backbone, frozen SMI-TED encoder, continuous flow matching, and 500 ODE steps.The model generates ligand coordinates, atom types, and bond types through the flow-matching framework.
  • Runtime and efficiency: The 3D sampling stage accounts for approximately 89.5% of runtime, with a 33.1% sampling-time overhead versus DrugFlow under matched settings.LiFT requires 2 h 38 min 48 s versus 1 h 59 min 20 s for DrugFlow; the comparison excludes DrugFlow’s language-generation and semantic-encoding stages.
  • Uncertainty reporting: Pocket-level uncertainty uses sample SD, across-pocket variance, and bootstrap 95% confidence intervals computed by resampling complete pockets.These estimates characterize across-pocket heterogeneity and sampling uncertainty, not retraining-seed variance.
  • Uncertainty reporting: Small numerical differences require caution because several docking and pose-validity confidence intervals overlap with strong 3D baselines.For example, QED-NoRef has a QED 95% CI of [0.723, 0.763] and an SA 95% CI of [2.602, 2.851].
  • Operational protocol: The language stage produces valid SMILES conditions through fixed prompting, deterministic parsing, validation, and semantic extraction procedures.Prompt variants change only task directives and do not alter generator weights or add downstream evaluation feedback.
  • Information boundary: LiFT uses fixed information boundaries: reference-free prompting receives pocket and task information, while reference-guided prompting additionally receives ligand SMILES and coarse size constraints.Downstream metrics are used only for evaluation, not to select, revise, or rank SMILES conditions.

D.6 Task-Specific Property Optimization Directives

Task-specific directives steer LiFT’s SMILES-derived semantic conditions while keeping the downstream generator and validation pipeline fixed. The resulting analyses combine aggregate chemical-filter outcomes with rule-level diagnostics.

  • Directive design: Property-oriented variants change only the LLM directive while keeping information boundaries, parsing, semantic extraction, and the 3D generator unchanged.This isolates the effect of balanced, Vina-oriented, and QED/safety-oriented task directives.
  • Evaluation: Aggregate RDKit, REOS, and PoseBusters pass rates summarize structural-alert, medicinal-chemistry, and 3D pose-validity compliance.Rule-level diagnostics are needed because aggregate scores do not identify which chemical or geometric constraints drive the results.
  • Chemical-filter diagnostics: Rule-level audits show LiFT no-reference variants are strong on ZINC and competitive on Glaxo, Inpharmatica, and PAINS-family pass rates.The localized alert profiles differ across model families, so favorable behavior is not attributed to a single filter family.

E.2 Fine-grained PoseBusters Checks

Fine-grained PoseBusters checks complement aggregate chemical-filter and pose-validity scores by separating local ligand geometry from protein–ligand compatibility. The diagnostics show why isolated geometry checks cannot substitute for overall 3D validity.

  • Diagnostic checks: PoseBusters evaluates aromatic-ring flatness, bond-angle geometry, double-bond flatness, internal clashes, protein–ligand distance, and protein-volume overlap.All reported values are pass rates, so higher percentages indicate fewer violations.
  • Interpretation: Language-only baselines can pass several local geometry checks while retaining very low overall PoseBusters pass rates.Thus, isolated sub-checks do not guarantee a valid 3D pose.
  • LiFT comparison: LiFT variants retain stable protein-overlap behavior and competitive geometry-related sub-checks alongside their RDKit/REOS compliance patterns.These results support interpreting aggregate gains together with both chemical-rule and 3D pose-validity diagnostics.

F Qualitative Case Study on Representative Targets

The qualitative study compares LiFT with autoregressive, diffusion, and flow-matching baselines across four challenging protein pockets. The broader formulation uses heterogeneous continuous and categorical dynamics with state-aware semantic routing.

  • Representative targets: The case study examines targets 1E8H, 1JN2, 2AZY, and 3AF2, spanning deep narrow pockets and wide flat grooves.AR, DecompDiff, PAFlow, and Ground Truth references provide the comparison set.
  • Qualitative comparison: LiFT reports QED scores of 0.893 in both 1E8H and 2AZY, alongside Vina affinity of -11.01 in 1E8H.The case study describes these results as combining strong intrinsic molecular properties with competitive binding affinity.
  • Qualitative comparison: LiFT’s visualizations report binding poses, geometric complementarity, Vina affinity, QED, and SA for generated complexes and Ground Truth references.The selected complexes are top-performing candidates balancing these critical metrics.
  • SCDR routing: SCDR uses a fixed semantic embedding, integration-time information, and an invariant structural statistic to route semantic modulation during generation.The structural statistic is permutation-invariant and geometrically invariant, while the routing module is bounded and geometry-preserving.
  • Generative formulation: The heterogeneous process applies continuous flow matching to coordinates and discrete Markov bridges to atom and bond identities.SCDR modulates the learned dynamics without changing the underlying continuous and categorical transition structures.
  • Scope of verification: The appendix establishes compatibility of semantic routing with the heterogeneous dynamics, while chemical validity, affinity, and controllability remain empirical questions.The symmetry and stability properties alone do not prove those empirical outcomes.

G.6 Equivariance and Stability Proofs

The appendix argues that SCDR preserves the GVP backbone’s permutation and SO(3) equivariance while remaining a bounded, stably initialized perturbation. These properties support semantic modulation without discarding the backbone’s geometric structure.

  • Equivariance preservation: SCDR preserves permutation equivariance and SO(3)-equivariance under the reported reflection-sensitive configuration.The proof uses invariant vector-channel norms, permutation-invariant pooling, invariant AdaLN coefficients, and invariant channel-wise vector scaling.
  • Bounded routing: SCDR yields a bounded local amplification of the original GVP update.The scalar and vector routed updates remain finite multiplicative perturbations of their corresponding backbone updates.
  • Initialization behavior: At initialization, AdaLN recovers the scalar-normalized ligand state while the vector state passes unchanged.Zero-initialized projections and near-identity residual gates provide stable startup behavior; Proposition 2 controls the gate range during training.
  • Overall characterization: Together, the propositions characterize SCDR as invariant, bounded, and stably initialized while enabling semantic modulation.The stated properties preserve the geometric structure of the GVP backbone.

H.1 Systematic Methodological Comparison

The comparison places LiFT at an interface where language-derived chemical semantics guide native pocket-conditioned 3D generation rather than directly specifying a final ligand. Its semantic condition operates inside the flow trajectory, with SCDR adapting influence to time and evolving structure.

  • Systematic methodological comparison: Existing native 3D SBDD methods directly model pocket-conditioned ligand geometry without an open-ended language-derived semantic condition.The comparison includes AR, Pocket2Mol, TargetDiff, DecompDiff, PAFlow, and DrugFlow.
  • Systematic methodological comparison: LiFT converts open-ended design preferences and pocket-derived specifications into a SMILES-derived soft semantic prior, while generating topology and pose in native pocket-conditioned 3D space.The SMILES-derived condition is not the final ligand.
  • Systematic methodological comparison: LiFT’s semantic condition acts inside the flow trajectory, and SCDR adjusts its influence according to flow time and the evolving ligand state.This distinguishes the location and role of control from approaches using other interfaces.
  • ELILLM comparison: ELILLM optimizes a final 1D candidate in an LLM latent space before conventional conformer construction and docking, whereas LiFT conditions native 3D generation directly.The methods therefore address related SBDD goals through different representations and stages.
  • Systematic methodological comparison: Table 21 compares approaches by language input, target representation, 1D representation, native output, trainable components, interaction stage, and trajectory-time control.Native output refers to the generative model’s direct output before optional post-hoc processing.
Loading 2608.31009v1…