SAFE-T

SAFE-T is a conditional chemical language model that treats a biological objective — a protein family, a specific target, and a mechanism of action — as a prompt, and models the likelihood of fragment-based molecular sequences given that prompt. Because the same distribution can be evaluated (to score a molecule) or sampled from (to design one), a single checkpoint covers tasks normally split across separate systems: virtual screening, drug–target interaction prediction, activity cliff detection, and goal-directed generation. It was developed by Lu Zhu and Emmanuel Noutahi at Valence Labs, the research arm of Recursion, and posted as a preprint in July 2025.

The design targets two recurring weaknesses of generative chemical language models. Goal-directed generation usually depends on an external reward — a docking score, a QSAR oracle, a reinforcement learning loop — that is unreliable or unavailable for most targets; SAFE-T replaces it with conditioning learned from annotated bioactivity data. Sequence models also rarely explain themselves, and SAFE-T's fragment-level tokenization supports a counterfactual attribution score identifying which fragments carry the conditional likelihood.

Molecules are written in SAFE (Sequential Attachment-based Fragment Embedding), the line notation introduced by the same group, which makes scaffold decoration and linker design ordinary sequence-completion problems. Unlike structure-based scoring functions or the group's later structure-aware affinity model Nesso-1, SAFE-T never sees a protein sequence or a binding pocket — targets enter as identifier tokens.

#Key Features

• Biological context as prompt tokens: Target family, target identifier, and mechanism of action are encoded as bracketed tokens (for example [Kinase], [CHEMBL203], [inhibitor]) prepended to the molecular sequence. Random masking during training lets the model condition on any subset at inference.
• Scoring and generation from one distribution: Normalized conditional log-likelihood ranks candidate molecules for a given biological context, while sampling from the same distribution produces new ones — no task-specific heads or fine-tuning.
• Structure-free throughput: The largest variant scores nearly 800 molecules per second on a single A100 — the whole of ChEMBL (over 2 million compounds) in under an hour per biological context, against roughly 240 molecules per hour for the docking tool GNINA.
• Preference calibration: A third training stage tunes likelihoods on matched active/inactive pairs, lifting activity cliff detection from 0.581 to 0.947 ROC-AUC, at the cost of generation validity, which falls to about 40% in the smaller variants.
• Fragment-level attribution: Replacing one fragment at a time and measuring the drop in conditional log-likelihood recovers known pharmacophoric features in case studies on tofacitinib and ruxolitinib.

#Technical Details

SAFE-T is a LLaMA-style autoregressive decoder: 16 layers, 16 attention heads, hidden dimension 512, 44.97M parameters, with 10.47M and 24.96M variants also reported. Inputs are capped at 512 tokens using a byte-pair encoding tokenizer fitted to SAFE strings and extended with 4,075 biological-context tokens, for a 5,070-token vocabulary. Training runs in three stages: causal language modeling over 11.4M molecules drawn from in-stock, drug-like ZINC20 compounds plus all unique ChEMBL molecules; context-conditioned fine-tuning on MoAT-DB; and preference tuning on ACNet pairs. MoAT-DB, assembled for this work from DrugBank 5.0, Connectivity Map, the Therapeutic Target Database, and ChEMBL34, contains 4,010 targets across 39 target classes, 580,079 compounds, and 22 mechanism-of-action terms in 1.15M samples; mechanism annotations are missing for 85% of entries.

Zero-shot results place SAFE-T at 6.63 EF@1% on LIT-PCBA, ahead of DrugCLIP (5.51) and Glide-SP (3.41), and at 0.706 and 0.759 ROC-AUC on DAVIS and KIBA against 0.686 and 0.706 for a MolGPS embedding plus MLP baseline. On the PMO goal-directed optimization benchmark it reaches top-10 AUC of 0.9950 (DRD2), 0.9927 (JNK3), and 0.9948 (GSK-3β) without reinforcement learning or per-task fine-tuning. Generation retains above 99% validity and above 99% constraint satisfaction on the DRUG benchmark's linker design and scaffold decoration tasks.

#Applications

The practical use is early-stage triage and design when no reliable scoring function exists: ranking a purchasable library for a kinase of interest, prioritizing analogs within a series, flagging activity cliffs during SAR analysis, or generating scaffold decorations conditioned on a mechanism. Because scoring needs only a SMILES string and a target identifier, it suits medicinal chemistry workflows that lack a co-crystal structure or a homology model.

#Impact

SAFE-T shows that conditional likelihood alone — no docking, no reward model, no protein structure — can match or beat structure-based and supervised baselines across screening, interaction prediction, and optimization, at a parameter count small enough for a single GPU. Its limits are equally clear: targets are identifier tokens rather than sequences, so new targets require extending the embedding layer; MoAT-DB is built from active compounds, narrowing exposure to inactive chemical space; preference tuning trades generation validity for ranking accuracy; and in-distribution target-family prediction lags the MolGPS baseline (0.777 versus 0.942 ROC-AUC) even as out-of-distribution performance improves. The work remains a preprint awaiting peer review, the code repository named in the paper is not publicly accessible, and no pretrained weights have been released.