AMix-1

AMix-1 is a protein sequence foundation model built on Bayesian Flow Networks (BFNs), a generative framework that refines a continuous belief state over amino acid distributions through iterative Bayesian updates rather than modeling discrete tokens directly. Its central argument is methodological: rather than a single checkpoint with benchmark numbers attached, AMix-1 is presented as a four-stage pathway borrowed from large language model practice — characterize a pretraining scaling law, identify when capabilities emerge, expose an in-context learning interface, then scale further at test time.

The model was developed by Shanghai AI Laboratory with the Generative Symbolic Intelligence Lab (GenSI) and the Institute for AI Industry Research (AIR) at Tsinghua University, alongside collaborators at Fudan, Tianjin, Georgia Tech, BUPT, the University of Chinese Academy of Sciences, and City University of Hong Kong. It first appeared as a preprint in July 2025, with weights for the 650M and 1.7B checkpoints released under Apache 2.0. It is the sequence-only predecessor to AMix-2, the same lineage's later protein–text model.

What distinguishes AMix-1 in use is that protein design requires no fine-tuning. A multiple sequence alignment is compressed into a position-wise amino acid frequency profile, and that profile becomes the prompt — a dense, family-level analogue of few-shot exemplars in a language model. The frozen model then samples sequences that inherit the alignment's evolutionary constraints.

#Key Features

• Bayesian Flow Network generative process: Generation proceeds as a sender–receiver message-passing loop in which noisy observations iteratively sharpen a belief distribution over each residue, giving a tunable noise parameter that trades novelty against fidelity to the input.
• Predictive pretraining scaling law: Validation cross-entropy is fit as a function of compute, model size, and token count across six model scales, making the loss of a larger run predictable before it is trained.
• Emergent structural understanding: Trained purely on sequences, the model's predicted-structure metrics stay flat and then rise sharply once cross-entropy falls below a threshold, an emergence effect visible only at intermediate noise levels.
• MSA profiles as prompts: Structure- and function-guided design are unified into one training-free interface, with inference cost independent of alignment depth because any MSA compresses to a fixed-size profile.
• EvoAMix-1 test-time scaling: A propose–verify–update loop uses the model as proposer and any plug-in scorer as verifier, refining the prompt rather than the weights across rounds of in silico directed evolution.

#Technical Details

The AMix-1 series uses an encoder-only Transformer backbone with rotary position embeddings, scaled across six configurations — 8M, 35M, 150M, 350M, 650M, and 1.7B parameters (6 to 48 layers, hidden dimensions 320 to 1680). Pretraining uses the EvoDiff-preprocessed UniRef50 split (EvoDiff supplied the data pipeline): 41,546,293 training sequences and 82,929 validation sequences, with sequences longer than 1,024 residues randomly cropped. Models were trained with AdamW for up to 1M steps in BF16; the 650M and 1.7B runs used 128 A800 GPUs across 16 nodes.

Emergence was measured on 20 CAMEO proteins using sequence recovery, and pLDDT and TM-score from ESMFold. Scaling and emergence hold at moderate corruption levels and vanish at extreme ones. In-context design case studies report TM-scores of 0.92 and 0.95 against reference structures at 58% and 74% sequence identity, and over 80% of generations conditioned on a beta-pinene synthase alignment retain the correct EC number under CLEAN. In wet-lab validation, 40 candidate AmeR repressor variants generated by the 650M model (each with ten or fewer mutations) reached a fold repression of 169.6 against 3.3 for wild type — roughly a 50-fold activity gain, and about 77% above the best previously reported variant. Across six in silico directed-evolution benchmarks, EvoAMix-1 produced the best final design in five and improved fastest in five, outperforming ALDE, EVOLVEpro, and MLDE.

#Applications

AMix-1 targets protein engineers who want to improve a natural protein rather than design one from scratch: enzyme activity and specificity engineering, optimal-temperature and optimal-pH tuning, transcriptional regulator improvement, and generating foldable variants of orphan proteins with few known homologs. Because EvoAMix-1 treats the verifier as a swappable module, the same loop runs against a cheap in silico predictor or against real assay readouts, which is what makes it usable inside a lab-in-the-loop campaign.

#Impact

AMix-1 is the first work to characterize scaling laws and emergent ability for a Bayesian Flow Network generative model, and it supplies an unusually complete open release — code, six-scale training methodology, and Apache 2.0 checkpoints — for a protein design model with wet-lab evidence behind it. Its limitations are stated plainly by the authors: the model is sequence-only and ignores the growing body of structural data, and the test-time scaling algorithm was validated against computational metrics rather than experimental assays. The preprint has been revised through four versions and has not been peer reviewed. Its successor AMix-2 extends the lineage toward joint protein–language modeling.