NewOrigin

Ask a sequence-only protein language model why a substitution three residues from a catalytic pocket destroys an enzyme, and the only evidence it holds is the company that position keeps in an alignment. The geometry that swings the side chain against the substrate, the reaction it takes part in, and the selective pressure that conserved it are each a different kind of observation, and most protein modeling stacks recover them from different models, leaving a human to reconcile the three before deciding what to build.

NewOrigin is MoleculeMind's attempt to collapse that stack into one pretrained model. The company describes it as holding four views of a protein at once — sequence, three-dimensional structure, function, and evolutionary constraint — together with the interactions a protein participates in, and as being trained for molecular design rather than for a single prediction task, which is what it credits for the model's reuse across tasks and industries without task-specific retraining. From that one checkpoint MoleculeMind runs sequence generation, structure prediction, function prediction, and de novo design. NewOrigin was unveiled at the World Artificial Intelligence Conference on 7 July 2023 by computational biologist Jinbo Xu, who founded the company in 2022 and whose 2016 RaptorX-Contact work was among the first demonstrations that deep learning could sharply improve protein contact and structure prediction.

Joint modeling across modalities is not unique to it — ESM-3 likewise generates and reasons across sequence, structure, and function — so what separates NewOrigin is less the modality set than the explicit evolutionary track and the orientation toward design rather than prediction. It sits underneath MoleculeOS, MoleculeMind's commercial protein design platform, alongside the all-atom complex predictor MMFold and the generative design model MMDesign.

#Key Features

• Four-way representation: sequence, structure, function, and evolutionary constraint are modeled together rather than chained through separate task-specific networks.
• Design-oriented pretraining: MoleculeMind reports training the model toward molecular design instead of one prediction objective, and reusing it across tasks and industries without refitting.
• One model across the workflow: sequence generation, structure prediction, function prediction, and de novo design are served from the same model.
• Small candidate lists: the company reports needing only a few dozen generated molecules to find a usable candidate, in place of large-scale wet-lab screening.
• Conversational access: the model was launched with a natural-language interaction mode aimed at biologists without a machine learning background; in current descriptions that conversational layer belongs to MoleculeOS rather than to the model.

#Technical Details

NewOrigin's architecture is undisclosed. There is no preprint, paper, technical report, or model card for it, and no published tokenizer, training-data source, compute budget, or evaluation protocol. The quantitative claims that exist come from MoleculeMind through Chinese press coverage: a scale on the order of tens of billions of parameters, reported in July 2024, and training over hundreds of billions of multimodal data points, reported at launch. No benchmark result has been published for the base model itself — the figures MoleculeMind circulates, including FoldBench antibody–antigen accuracy and per-target design hit rates, belong to MMFold and MMDesign. The model has been described as the company's current base model continuously from July 2023 through 2026 with no public version numbering.

#Applications

MoleculeMind packages NewOrigin into five named solution areas: antibody affinity optimization, protein stability under extreme temperature, pH, and solvent conditions, enzyme activity optimization, enzyme–substrate docking, and de novo protein design. It reports applying the model in innovative drug development, materials, food, chemicals, and agriculture, and cites a redesigned green fluorescent protein that retains function at under half the residue count of the natural protein. Access is through MoleculeOS, which the company opened to industry on 2 July 2026 and delivers as an API, software, and services; there is no other route to the model.

#Impact

NewOrigin is the model layer beneath a shipping commercial platform rather than a research artifact, and that shapes both its reach and what can be said about it. MoleculeMind raised a Series A of over $100 million in June 2026 and positions the platform as research infrastructure for the bio-economy, with the model appearing in Chinese state media coverage from its launch onward. The countervailing fact is that NewOrigin is entirely closed: no weights, no code, no license, no paper, and no model card, which puts every technical and performance claim about it in the company's voice with no independent evaluation available. This is not uniform across MoleculeMind's work — it co-published SurfFlow with Stanford and issued a technical report covering MMFold and MMDesign — which makes the absence of any comparable document for the base model the more notable. Researchers can use NewOrigin through MoleculeOS, but they cannot reproduce it, audit its training data, or benchmark it against alternatives.