MDHGPT

Generative protein models are usually trained to reproduce their training sequences amino acid by amino acid. That objective cannot distinguish a substitution that preserves function from one that destroys it — both incur identical loss — and the problem compounds for enzymes, where catalytic function depends on a small number of residues held in precise geometric relation across a long sequence.

MDHGPT, from the Machine Biology Group at the University of Pennsylvania, generates malate dehydrogenases under a scheme designed to supply the missing signal. It is fine-tuned from ProGen2-large, the 2.7-billion-parameter autoregressive protein transformer, and trained against a second objective alongside cross-entropy: the distance between the generated sequence and its training template in the latent space of a frozen ESM-2 encoder. Because that encoder clusters functionally similar sequences together, the distance acts as an approximate function checker obtained without any labeled activity data.

It shares its recipe with AMPGPT, which applies the same method to short antimicrobial peptides. The enzyme case is where the method's evidence is strongest: with the latent objective, 96.8% of generations are classified as malate dehydrogenases; without it, none are.

#Key Features

• Function-carrying generation at enzyme length: The model produces full-length catalytic proteins rather than the 1–100 residue peptides these generative schemes are usually demonstrated on.
• Latent-space likelihood as a second objective: The total loss adds a mean squared error term between the encoder embeddings of the generated and template sequences, weighted by a hyperparameter swept over 1, 10, and 100.
• Policy gradient bridges the non-differentiable gap: Because the latent error cannot be backpropagated through discrete sampling, the generator is treated as a policy and the encoder as the environment, reducing in practice to reweighting each sample's ordinary gradient by one plus the scaled latent error.
• Reward baseline variants: EMA variants subtract an exponential moving average of the historical reward, a standard variance-reduction trick that yields the strongest results on this enzyme class.

#Technical Details

The generator is initialized from ProGen2-large and the frozen validator is the 33-layer ESM-2 encoder. Training used 16,706 enzyme sequences for 30 epochs at batch size 4, with AdamW at a 1e-5 peak learning rate and warm-up over the first 5% of steps. Generation is seeded by sampling terminal decapeptides from the training frequency distribution — a longer seed than the peptide sibling uses, matched to the longer target.

The headline evaluation is an ablation. Judged by a purpose-built enzyme classifier, 96.8% of generations are called malate dehydrogenases, against 0% for the same architecture trained without the latent objective and 0% for four ProtGPT checkpoints fine-tuned on sequence clusters. A result that moves from zero to near-total is unusually clean for a generative ablation, and it is the paper's central evidence that the latent term carries functional information rather than merely regularizing.

#Applications

The direct use is enzyme engineering: generating novel malate dehydrogenase variants for screening, in a setting where the wild-type family is well characterized but the accessible sequence space around it is not. The broader significance is scope — showing that a recipe demonstrated on short peptides carries to full-length, structurally constrained catalytic proteins, which is the step most peptide-scale generative methods never take. The training scheme itself is architecture-agnostic: any generator that explicitly models training-data likelihood, including variational autoencoders, flow models, and diffusion models, can adopt the same reweighting term.

#Impact

This is the result that makes the case for the method. Where the peptide sibling's gains are visible under one external scorer and absent under another, here the ablation is unambiguous, and it establishes that a frozen protein language model can supply a functional training signal for generation without labeled activity data.

The caveats are real. Validation is entirely computational — nothing was synthesized or assayed, and the authors name wet-lab confirmation as future work — so the 96.8% figure is a statement about a classifier's judgment, not about catalysis. The evaluation covers a single enzyme family. And no checkpoint has been released, with no public code accompanying the preprint, which remains unreviewed and licensed CC BY-NC, so the recipe is one others would have to rebuild themselves.