SOLD

SOLD (SELFIES-based Objective-driven Latent Diffusion) is a generative model for target-conditioned de novo drug design that produces candidate ligands by running diffusion in a learned latent space over one-dimensional SELFIES strings. Most structure-based drug design generators sample molecules directly in 3D atomic space conditioned on a binding pocket, which requires a resolved or predicted receptor structure and typically an equivariant graph neural network to handle the geometry. SOLD takes the opposite route: molecules are compressed by a transformer encoder into a compact latent tensor, and a 1D U-Net diffusion model generates new latents conditioned on the target protein.

The conditioning signal is a protein sequence embedding rather than a pocket geometry. Targets are encoded with ESM-2, so SOLD needs only an amino acid sequence at inference time and can be pointed at proteins with no solved structure. Because SELFIES uses a formal grammar in which every string decodes to a chemically valid molecule, the generator never has to spend capacity learning syntax validity — a known failure mode of SMILES-based generative models.

SOLD was developed by Elbert Ho, working as an independent researcher with computing resources provided by the Pingry School, and posted to arXiv in September 2025. It is a preprint that has not been peer reviewed, and the reported evaluation is entirely in silico.

#Key Features

• Sequence-only target conditioning: Proteins enter the model as 1280-dimensional ESM-2 embeddings, whose weights are fine-tuned during diffusion training, so generation requires no binding-pocket structure and extends to targets without crystal or cryo-EM data.
• 1D SELFIES latent space: Diffusion runs over 128-position latents with 256 channels rather than 3D atomic coordinates, avoiding equivariant architectures and shortening sampling time.
• Multi-task molecular encoder: The encoder-decoder transformer is trained jointly on SELFIES reconstruction and regression to five PCA-selected RDKit descriptors (spatial score, MinEStateIndex, exact molecular weight, Balaban J index, VSA_EState6), shaping a latent space organized by chemical properties.
• Modified multi-task loss balancing: A dual-balancing gradient scheme is extended with the Adam second moment when normalizing per-task gradients, which the author reports substantially improved convergence.
• Classifier-free guidance: Twenty percent of training examples are conditioned on a null target; a guidance weight of 5 gave the best docking scores in a sweep from 0 to 11.

#Technical Details

The molecular encoder is a two-layer transformer with embedding dimension 256 and four attention heads, trained on 10,000 ChEMBL molecules that were converted to randomized SMILES, then to SELFIES, then tokenized with a byte-pair encoder of vocabulary size 256 and padded to length 128. Training ran 200 epochs with a 20-epoch reconstruction-only warmup and cosine-annealed learning rate from 1e-4; a VAE encoder was tried first and abandoned after reconstructing 0% of sequences correctly. The diffusion model follows the GLIDE architecture with a 1D U-Net and 1000 timesteps, totals 360M parameters, and was trained for 1000 epochs at batch size 32 in 16-bit precision on a single NVIDIA RTX 4090 using 15,000 protein-ligand pairs from PDBbind. On the SARS-CoV-2 3CL protease, a held-out target, SOLD reports a mean AutoDock Vina score of -4.498 against -7.318 for TargetDiff and -7.333 for DiffSBDD, but a top-10% Vina score of -9.830 that is close to DiffSBDD's -9.927; it also reports the highest diversity (0.946 versus roughly 0.72-0.74 for the 3D baselines) and the shortest generation time (960 s per 100 molecules). QED (0.373) and normalized synthetic accessibility (0.451) fall below every 3D baseline, and an evolutionary noise-and-denoise property optimization scheme degraded docking scores and was dropped.

#Applications

The sequence-only conditioning makes SOLD most useful for early hit generation against targets where no reliable pocket structure exists, and for rapidly enumerating chemically diverse starting points that can be triaged by docking and medicinal-chemistry filters. Its speed and diversity profile suit a generate-then-filter workflow — sample broadly, keep the top decile by predicted affinity — rather than careful optimization of a single lead, which the paper's own results show it does not currently support.

#Impact

SOLD is an argument that the 1D chemical string representation, paired with latent diffusion and a protein language model for conditioning, remains competitive on the metric that matters most for hit finding — the quality of the best molecules generated — while being cheaper to train and faster to sample than pocket-conditioned 3D generators. The caveats are substantial and stated plainly by the author: the training corpus is small relative to available protein-ligand data, evaluation rests on docking scores for a small number of held-out targets with no experimental validation, and average affinity, drug-likeness, and synthetic accessibility all trail established baselines. The preprint reports that source code is on GitHub but includes no link, and no public repository, released weights, model card, or license accompanies the paper, so the results are not currently reproducible by third parties.