BioM3

An engineer who wants a new SH3 domain capable of rescuing osmosensing in yeast must translate that intent into whatever conditioning signal a generative model happens to accept: a control tag, a backbone geometry, a curated alignment. The knowledge that actually specifies the design — "key transmembrane SH3 domain protein in osmotic sensing involving Cdc42p/MAP kinase interactions" — already exists in the free-text annotation fields of SwissProt and Pfam, in a form no sequence model consumes. BioM3, the Biological Multi-Modal Model, makes that annotation the conditioning signal itself.

Developed at the University of Chicago by Nikša Praljak, Hugh Yeh, Miranda Moore, Michael Socolich, Rama Ranganathan and Andrew Ferguson, and released as a bioRxiv preprint in November 2024, BioM3 is a three-stage framework. PenCL aligns protein sequences with their text annotations in a shared 512-dimensional embedding space. The Facilitator maps a text embedding into the region where protein embeddings actually live. ProteoScribe, an order-agnostic autoregressive diffusion decoder, then writes sequences conditioned on that facilitated embedding.

Its lineage runs through ProtST, which first aligned proteins with biomedical text, and ProteinDT, which first used such a space to steer generation. BioM3 scales that recipe to 45 million text-protein pairs, adds a family-level contrastive term treating Pfam homology as a training signal, and swaps in a text-conditioned version of the order-agnostic diffusion process introduced by EvoDiff. What separates it from that lineage is the evidence: prompt-designed sequences were synthesized and assayed in living cells.

#Key Features

• Free-text conditioning: Generation is steered by an ordinary written description of lineage, function and localization, so specifying a design requires no controlled tag vocabulary.
• Homology as a training signal: A protein family contrastive loss attracts Pfam homologs and repels non-homologs, embedding evolutionary structure directly into the joint space.
• Order-agnostic generation and in-painting: The diffusion decoder reveals residues in arbitrary order, so a known functional motif can be held fixed while the surrounding scaffold is generated.
• Retrieval-based zero-shot annotation: Because sequences and text share one space, an uncharacterized sequence is annotated by retrieving its nearest text prompt, with no fine-tuning.
• Wet-lab validated designs: Prompt-written sequences were gene-synthesized and tested in a yeast selection assay and in biochemical binding measurements.

#Technical Details

PenCL pairs ESM-2 at 650M parameters with PubMedBERT-full at 100M as the biomedical text encoder, each followed by a projection head into the 512-dimensional joint space. Four objectives are minimized together: a global InfoNCE contrastive loss over text-protein pairs, the protein family contrastive loss, and masked-language-model losses on both backbones, which keep training rather than staying frozen. The corpus is 45 million pairs curated from SwissProt and Pfam, each prompt concatenating annotation fields such as protein name, function, subcellular location and lineage. The Facilitator is a small autoencoder trained with a maximum mean discrepancy loss. ProteoScribe is a 16-block, 16-head transformer with embedding dimension 512, a 1,024-token context and roughly 90M parameters.

On zero-shot text retrieval, PenCL reaches 0.531 accuracy on enzyme reaction classification against 0.300 for ProtST, and 0.402 on subcellular localization against 0.420. On remote homology retrieval it leads BLASTp, ProtT5, ESM-2 and ProtST at every level, reaching 0.977 top-1 accuracy at the family level. Sequences generated from held-in prompts folded to TM-scores above 0.922 and RMSD below 2.32 Å against the reference, sharing as little as 53.59% identity with the closest BLASTp hit.

#Applications

The immediate use is candidate generation for protein engineering when the specification is functional rather than structural: an enzyme with a given catalytic activity, a domain binding a named partner, a library seeded around a described phenotype for directed evolution. In-painting extends this to grafting, embedding a fixed catalytic site or binding loop into fresh scaffolds under a prompt describing the desired context. The retrieval side of the joint embedding serves a different audience — annotating orphan sequences and detecting remote homologs that sequence-identity search misses. A hosted pipeline runs all three stages behind an access allowlist.

#Impact

BioM3 supplies the first reported in vivo and in vitro validation of proteins designed from natural-language prompts. ProteoScribe was fine-tuned on 25,030 SH3 domain sequences, then prompted five ways to produce 984 designs, assayed for their ability to rescue osmosensing in Saccharomyces cerevisiae under high-osmolarity challenge. Prompt 5 — a short, informal description unlike the formal training annotations — yielded 25 of 317 sequences with wild-type-level function, and one surviving Prompt 4 design returned no BLAST hits at all, at 28% identity to wild-type Sho1 SH3. That a casually worded prompt still produced functional protein is the result that matters most. The caveats are equally concrete: success came only after family-specific fine-tuning on thousands of SH3 sequences, validation covers a single domain family, and the three stages must be assembled by hand from separately downloaded checkpoints. The work remains a preprint.