LiFT

Ask a language model for a molecule that binds a kinase and it will write a plausible SMILES string with sensible valency, sane rings and drug-like substituents — and no idea what the pocket looks like. Ask a pocket-conditioned 3D generator for the same thing and the geometry will be right while valency, ring topology and medicinal-chemistry preferences remain implicit in the coordinates. The usual bridges between the two either fine-tune the geometric generator for each new task or bolt guidance onto sampling time, where a fixed chemical target pulls against a ligand whose 3D configuration is still changing.

LiFT's premise is that the language model's molecule should not be the answer at all. It is a trend. A "Sense-Evolve-Assemble" agent reads the pocket — a bounding box, principal geometric axes, and the hydrophobic, acidic, basic and metal-coordinating features inside an 8.0 Å shell around the pocket centroid — plus a stated task preference and an optional reference ligand, and emits target-aware SMILES. Those strings are sanitized by RDKit and then discarded as molecules: a frozen chemical foundation model compresses them into a single continuous vector that conditions the 3D generator toward a region of chemistry, never toward a specific structure.

Released in August 2026 by City University of Hong Kong with the University of Hong Kong and Stanford University, and accepted to Findings of EMNLP 2026, LiFT is built on the DrugFlow flow-matching backbone with new conditioning modules and its own trained checkpoint. Because steering happens through the condition embedding rather than the weights, one checkpoint covers de novo design and reference-guided scaffold hopping across four preference settings.

#Key Features

• Language as a soft prior, not a proposal: agent-written SMILES are encoded by a frozen SMI-TED molecular-language model into a 768-dimensional semantic latent, so chemical knowledge enters as a direction rather than as a structure the generator must reproduce.
• Zero-initialized cross-modal priming: a lightweight projector maps the latent to adaptive-normalization parameters initialized at zero, so conditioning starts as the identity and the pretrained flow backbone is never destabilized at the outset of training.
• Equivariance preserved by construction: priming is applied only to the invariant scalar channel; the rotation-equivariant vector features are left untouched, keeping the pocket-centered SO(3) symmetry of the geometric backbone intact.
• Self-Conditioned Decoupled Router (SCDR): a DeepSets readout summarizes the partially built ligand into invariant statistics, and dual gates rescale semantic influence along the ODE trajectory — so chemical guidance yields where it would conflict with emerging pocket geometry.
• Task steering without retraining: Balanced, Distribution, QED and Vina modes, each with and without a reference ligand, are selected by pointing sampling at a different precomputed condition embedding.

#Technical Details

The generator is a five-layer heterogeneous GVP-GNN trained by continuous-time flow matching on the refined CrossDocked2020 set of roughly 100,000 protein-ligand complexes, with batch size 48, learning rate 8e-4, 500 epochs and 500 ODE steps at inference; the released checkpoint is epoch 399. The SMI-TED encoder that supplies the priors is pretrained on 91 million molecules and stays frozen. On CrossDocked2020 distribution matching, the Vina-No-Reference variant reaches Wasserstein distances of 0.031 for Vina and 0.019 for Gnina and the best ring-frequency alignment in the table (0.008 and 0.004 for rings > 0 and > 10). Property-oriented sampling pushes mean QED to 0.757 and SA down to 2.659 in the Balanced-No-Reference setting, with RDKit and REOS filter pass rates above 71% and PoseBusters validity up to 73.56%. Ablations show SCDR matters more to QED and filter compliance than zero-initialization does. Swapping the condition-writing LLM among GPT-4o, Claude 4 Sonnet and DeepSeek-V3 changes results modestly; GPT-4o is the default.

#Applications

LiFT targets the early stage of structure-based drug design where a team has a pocket and a preference rather than a lead: generating diverse 3D-plausible starting points for a target, or hopping scaffolds away from a known binder while keeping the pharmacophore. Because the steering instruction is written in natural language and consumed as an embedding, a chemist can ask for higher drug-likeness or stronger predicted affinity and get a different sample distribution without touching training code.

#Impact

LiFT is an argument about where language models belong in molecular generation: upstream, as a source of trend-level chemical priors, rather than as the generator of record. The evidence is in-silico only — CrossDocked2020 benchmarks, no experimental validation — and LiFT does not lead every distribution-matching column, with autoregressive and diffusion baselines still ahead on some Wasserstein distances. The condition-generation stage also depends on an external commercial LLM API, though the released archive ships the eight condition files and two precomputed embedding sets so inference can run without it. Code is MIT-licensed and the paper checkpoint is downloadable, making the recipe easy to graft onto other pocket-conditioned generators.