FragAtlas-62M

Fragment-based drug discovery starts from very small compounds — typically eight to sixteen heavy atoms — that bind a target weakly but with high ligand efficiency, and then grows or links them into leads. It is a workhorse strategy in medicinal chemistry, but the generative models built for it have lagged: most AI molecular generators are trained on full-sized drug-like molecules, and fragment-specific computational work has concentrated on virtual screening of fixed libraries or on fragment linking rather than de novo generation.

FragAtlas-62M addresses that gap with a chemical language model pretrained on the complete fragment subset of the ZINC-22 tangible-compound database — 62,015,589 molecules from an April 2025 snapshot, defined by heavy-atom-count bins H08 through H16 and LogP bins M500 through P240. It was developed by Alexander Ho, Sukyeong Lee, and Francis T. F. Tsai at Baylor College of Medicine and posted as an arXiv preprint in September 2025.

The design is deliberately modest in scale. Where chemical language models such as MoLFormer-XL pursue billion-molecule pretraining for property prediction, and target-conditioned generators such as SAFE-T add biological context, FragAtlas-62M is a plain causal language model over fragment SMILES whose value comes from exhaustive coverage of a well-defined chemical subspace rather than from architectural novelty. Evaluation is distributional — validity, descriptor and fingerprint agreement, novelty and rediscovery rates — rather than downstream task benchmarking.

#Key Features

• Complete ZINC-22 fragment coverage: Training used every molecule in the ZINC-22 2D fragments subset rather than a sample, so the model's prior spans the commercially tangible fragment space that screening groups actually order from.
• High chemical validity: Of 62,015,589 sampled SMILES, 61,951,924 (99.90%) passed RDKit sanitization, leaving 42,597,827 unique canonical structures after deduplication.
• Balanced rediscovery and novelty: Generated molecules recover 53.55% of the ZINC fragment set while 22.04% of unique canonical outputs are structures absent from it.
• Distribution-matched chemistry: Across 12 molecular descriptors — QED, LogP, molecular weight, TPSA, hydrogen-bond donors and acceptors, heavy-atom and ring counts, rotatable bonds, synthetic accessibility, and natural-product likeness — effect sizes between generated and training distributions are negligible (all |d| < 0.2).
• Consumer-hardware training and inference: Sampling sustains more than 1,000 molecules per second on a single RTX 4090, and both training and generation fit on commodity GPUs.

#Technical Details

The model is a GPT-2 decoder implemented with HuggingFace Transformers on PyTorch: 6 layers, 12 attention heads, 768-dimensional embeddings, a 128-token context window, and 42.7M parameters. Inputs are SMILES strings encoded with a 42-symbol character-level tokenizer; the release corpus is roughly 2 billion tokens with an average string length of 31.54 ± 5.85 characters. Exact duplicate strings were removed, but SMILES were deliberately left un-canonicalized during training to retain the augmentation benefit of non-canonical representations, with canonicalization applied only at evaluation. Training ran for 5 epochs on a 99:1 train/validation split with a 5e-5 initial learning rate and a linear schedule with 10% warmup.

Validation compared generated and reference molecules using Morgan (ECFP4, radius 2, 2048-bit), MACCS (166-bit), and topological (2048-bit) fingerprints. Intergroup-to-intragroup Tanimoto distance ratios between novel and rediscovered sets sit near unity for all three fingerprint types (0.978–1.027), and t-SNE projections show the two sets occupying overlapping regions of chemotype space, indicating the novel outputs are not drifting into a separate, less realistic corner of chemical space. The released artifacts are the model weights, distributed on Hugging Face under an MIT license, and the preprocessed ZINC-22 fragment map with ZINC identifiers, hosted on the ZINC file server.

#Applications

The most direct use is enumerating fragment libraries: sampling from the fixed checkpoint yields large sets of valid, property-matched fragments that can seed growing, linking, or merging campaigns without re-training. Because outputs carry ZINC lineage for the rediscovered majority, screening groups can separate immediately purchasable fragments from novel proposals that would require synthesis. The small parameter count also makes the checkpoint a practical starting point for fine-tuning or conditioning on target-specific or property-driven objectives, which matters for academic groups without large GPU allocations.

#Impact

FragAtlas-62M is a preprint awaiting peer review, and its contribution is a well-characterized open baseline for a niche that lacked one: a fragment-scale generative model with published validity, coverage, and novelty statistics on a defined corpus. The authors are explicit about the boundaries. The model does not represent stereochemistry, 3D geometry, or fragment-to-fragment connectivity rules, so it cannot drive structure-based assembly or automated lead construction on its own, and its evaluation is entirely in silico — no generated fragment was tested experimentally, and no downstream affinity or hit-rate benchmark is reported. Realizing target-specific performance requires fine-tuning or coupling to docking, scoring, or synthesis-aware modules.