Multimodal sequence model spanning proteins, coding DNA, and regulatory DNA for zero-shot fitness scoring and conditional sequence generation.
No providers recorded yet. Browse all providers
ProDMM is a multimodal sequence model that represents proteins, their coding DNA sequences (CDS), and the adjacent non-coding regulatory DNA (NCDS) in one shared vocabulary. Most biological language models pick a lane: protein models such as ESM-2 see amino acids and discard codon usage entirely, while genome models such as Evo see nucleotides but have no explicit notion of the protein a coding sequence produces. Yet the properties an engineer actually cares about — how much protein a construct expresses, how much product a pathway yields — depend on all three layers at once.
Developed at Shanghai Jiao Tong University with the Shanghai AI Laboratory and posted as a preprint in March 2025, ProDMM pairs a bidirectional encoder trained by masked language modeling with a conditional decoder for autoregressive generation. Input sequences carry modality tags (<aas>, <cds>, <ncds>), so a single model accepts a bare protein, a bare coding sequence, or a composite construct such as NCDS-CDS-NCDS. That framing lets one checkpoint score a variant, rank a pathway design, and write a promoter.
The ProDMM encoder is a 48-layer bidirectional transformer with a hidden size of 1,536, an intermediate size of 6,144, and 64 attention heads — roughly 1.36 billion parameters. It was pretrained with BERT-style masked language modeling (15% masking; 80% replaced with <mask>, 10% with a random token, 10% left intact) on 129 million CDS-protein-regulatory-element quadruples drawn from the Genome Taxonomy Database, spanning 23 bacterial phyla, over 260,000 gradient steps on 24 NVIDIA H100 GPUs across 40 days. The seq2seq model adds a 12-layer decoder attending to encoder representations through cross-attention, giving roughly 1.81 billion parameters, and was trained for 100,000 steps on 32 A100s.
On the prokaryotic subset of the ProteinGym deep mutational scanning benchmark, zero-shot mutant scoring from encoder perplexity outperforms ESM-2, GEMME, VESPA and TranceptEVE. CDS perplexity correlates with measured expression level at an average Spearman r of 0.8. For non-coding RNA fitness, promoter transcription intensity in E. coli, and naringenin yield from a four-gene pathway, ProDMM ranks candidates better than GenSLM, RNA-FM, DNABERT-2 and Evo. The generative side beats n-gram codon baselines and CD-GPT on reverse translation for both E. coli and human targets, and promoter-conditioned CDS generation works despite never appearing in pretraining.
The intended user is a metabolic or protein engineer choosing among constructs. Because scoring is perplexity-based, an entire promoter-CDS layout for a biosynthetic pathway can be ranked in silico against expected product titer before any DNA is ordered — the naringenin case study ranks promoter combinations against measured yield. The generative modes cover the complementary step: turning a designed protein into a coding sequence optimized for expression in a chosen host, or writing a regulatory element sized to the gene it drives. The encoder also supplies embeddings for supervised property prediction, including melting temperature and solubility.
ProDMM is an argument that protein-DNA co-design belongs in one model rather than a pipeline of specialists, and its pathway-level zero-shot results are the strongest evidence for that claim: no single-modality baseline sees both the promoters and the enzymes. Its limits follow from its corpus. Pretraining covers prokaryotic genomes from GTDB only, so eukaryotic performance rests on transfer rather than direct coverage, and the pretraining objective addresses interactions between a protein and its immediate flanking regions rather than long-range coupling between genes across a pathway — the gap the antiSMASH-tuned variant was built to narrow. The work remains a preprint awaiting peer review, the code is GPL-3.0, and the checkpoint published under the AI4Protein organization on Hugging Face is the antiSMASH-tuned encoder, gated behind manual access approval.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.