SynthFormer

A generative model can propose a molecule that docks beautifully and is still useless, because nobody can make it. Most 3D generative models handle this with the synthetic accessibility (SA) score, a heuristic that does not reliably separate makeable compounds from unmakeable ones. SynthFormer removes the question by construction: it never emits an atom graph. It emits a synthetic tree — an alternating sequence of commercially available Enamine building blocks and reaction templates — so the output is the recipe, and the molecule is whatever that recipe produces.

The other half of the design is what the model is conditioned on. Structure-based generators such as TargetDiff need a protein pocket, which GPCRs, ion channels, multi-domain complexes and phenotypic screening hits often cannot supply. The authors note that unrefined predicted structures from AlphaFold 2 or Boltz-1 generally underperform experimental ones in virtual screening, which keeps ligand-based design relevant. SynthFormer therefore conditions on a pharmacophore — the 3D point cloud of hydrogen-bond donors and acceptors, aromatic rings, hydrophobic centres and ionizable groups read off a known active. An E(3)-equivariant graph neural network encodes that geometry and a transformer decoder translates it into a synthesis route, placing the model beside pharmacophore-conditioned generators like PharmaDiff but with reactions, rather than atoms, as its output alphabet.

The model comes from Zygimantas Jocys and colleagues at the University of Southampton, with the ALBORADA Drug Discovery Institute at the University of Cambridge. It appeared as a preprint in October 2024 and in Artificial Intelligence in the Life Sciences in December 2025.

#Key Features

• Synthesis as the output alphabet: The decoder predicts building blocks and reactions rather than atoms and bonds, so each compound comes with an interpretable route through purchasable reagents.
• Equivariant pharmacophore encoder: A seven-layer EGNN treats the pharmacophore as a fully connected graph of typed points with Cartesian coordinates, so generation follows spatial arrangement rather than a bare feature list.
• Ligand-based conditioning: One crystallographic ligand, reduced to its pharmacophore, is enough input — no pocket, no docking grid, no reliance on a predicted fold.
• Hit expansion from a fingerprint: Seeding the decoder with a hit's Morgan fingerprint grows analogs around it; across ten targets, 11.4% docked better than the original hit at an average Tanimoto similarity of 0.67.
• Reagent-level optimization: A genetic algorithm swaps reagents inside the predicted route and resamples reactions over three cycles, so optimization never leaves the synthesizable space.

#Technical Details

Training data is generated rather than curated. Random reaction rollouts over 251,222 Enamine building blocks and a public set of 58 SMARTS reaction templates produce synthetic trees of up to five building blocks; each final molecule gets one MMFF-optimized RDKit conformer, is filtered to QED above 0.5, and yields a pharmacophore-to-tree training pair. The encoder is a seven-layer EGNN over six pharmacophore types plus 3D coordinates, projected to 512 dimensions and cross-attended by a seven-layer, eight-head, 512-dimensional transformer decoder whose input tokens are 4096-bit radius-3 Morgan fingerprints. The loss sums a cosine-similarity term for building-block selection and cross-entropy for reaction choice.

Evaluation uses ten PDB structures spanning nine Pfam families, with generated molecules redocked by SMINA and QVina into the originating site. Against ligand-based 3D generators, SynthFormer reports a ΔDock of 2.46 kcal/mol — comparable to SQUID at 2.39 and better than LigDream at 2.78 — with markedly lower structural overlap with the reference ligand (ΔTanimoto 0.09 versus 0.24 and 0.22) and 100% in-silico synthesizability against 23.6% for SQUID and 32.4% for LigDream. An ablation isolates the geometry: removing equivariance degrades ΔDock from 2.39 to 2.75, dropping 3D coordinates gives 2.77, and a decoder-only baseline reaches only 3.45.

#Applications

The model targets campaigns where a few actives are known but a usable structure is not. From one hit's pharmacophore it produces scaffold-hopped compounds whose routes already exist in a commercial catalogue. Hit expansion and reagent-level lead optimization use the same frozen checkpoint, so a medicinal chemistry team can move from hit finding to analog series to property tuning without retraining, and can inspect the proposed reaction sequence before committing bench time.

#Impact

SynthFormer joins 3D pharmacophore conditioning to synthesis-constrained decoding, and argues a useful point: enforcing synthesizability by construction costs little docking performance while removing a class of unusable outputs. The limits are real. Synthesizability is in-silico, bounded by the 58-template reaction set and the Enamine catalogue, and no compound in the paper was made or assayed; the optimization cycles improved logP by 0.21 and drug-likeness by 0.09 while making average docking energy 0.03 kcal/mol worse; and the authors state that hit expansion and optimization have no established benchmark, so those results stand uncompared. The code is public, no trained weights are distributed, and reproducing the model requires registered downloads of the Enamine building-block set and PDBbind.