De novo generation of functional terpene synthases using TpsGPT

AI summary

Overview

Research area: Machine learning for protein design — specifically, generative protein language models applied to enzyme engineering (terpene synthases).

Technical level: Intermediate. The paper is readable for someone familiar with basic machine learning and molecular biology concepts, but assumes comfort with terms like perplexity, fine-tuning, pLDDT, and enzyme classification numbers.

Scope: The paper describes TpsGPT, a fine-tuned protein language model that generates novel terpene synthase enzyme sequences in silico, seven of which passed multi-stage computational filters and two of which showed enzymatic activity in wet-lab yeast experiments.

What This Paper Is About

Terpene synthases (TPS) are enzymes that build the carbon scaffolds of terpenes, the largest and most diverse class of natural products, which includes medicines like the anticancer drug Taxol. Designing new TPS enzymes is slow and expensive because conventional methods like directed evolution largely explore sequences that stay close to naturally occurring proteins. This paper asks whether fine-tuning a protein language model on a curated TPS-specific dataset, followed by strict filtering, can generate genuinely new and functional TPS enzymes that are evolutionarily distant from anything in the training data.

Key Contributions

1. A generative pipeline for TPS design: TpsGPT, built by fine-tuning the distilled ProtGPT2 Tiny model (38.9 million parameters) on a curated dataset of 79k TPS sequences mined from UniProt (seeded from 1125 experimentally characterized TPS enzymes).

2. A seven-metric filtering framework: A multi-stage pipeline combining sequence filters (perplexity, maximum sequence identity), function filters (EnzymeExplorer, CLEAN, InterPro), and structure filters (ESMFold pLDDT, Foldseek TM-score) that reduced 28k generated sequences to seven putative enzymes.

3. Experimental confirmation of activity: Heterologous expression of the generated genes in Saccharomyces cerevisiae strain JWY501 with LC-MS detection confirmed enzymatic activity in two of the seven candidates (TpsGPT1 and TpsGPT2).

4. A cost-efficiency demonstration: The full pipeline was run on a single NVIDIA L4 tensor core GPU for less than $200 in GPU cost, showing that class-specific fine-tuning of a small distilled model is accessible without large-scale compute.

Main Findings

• Generation scale and quality: TpsGPT produced 28k sequences; more than 40% of the top 2,800 sequences ranked by perplexity achieved pLDDT scores of at least 70, indicating stable predicted structures.

• Function filter yield: Of the top 2,800 sequences, 77 passed the EnzymeExplorer TPS detection threshold of 0.7 or higher.

• Seven candidates survived all filters: These had sequence filters, function filters, and structure filters satisfied, with maximum sequence identities to the training set ranging from 49.67% (TpsGPT1) to 60.08% (TpsGPT4), and Foldseek TM-scores to their top training-set structural matches between 0.6 and 0.9.

• CLEAN and InterPro agreement: CLEAN assigned all seven sequences to TPS EC classes, including Germacrene D Synthase (4.2.3.75) for TpsGPT1, Squalene Synthase (2.5.1.21) for TpsGPT2, TpsGPT4, and TpsGPT6, Cucurbitadienol Synthase (5.4.99.33) for TpsGPT3, Beta-amyrin Synthase (5.4.99.39) for TpsGPT5, and Cycloartenol Synthase (5.4.99.8) for TpsGPT7. InterPro detected at least one TPS-specific domain in every sequence.

• Two enzymes experimentally validated: LC-MS analysis of yeast expressing TpsGPT1 and TpsGPT2 showed extracted ion chromatograms at the mass of C20H36O2 consistent with diterpene-like products (for example, sclareol, CHEBI:9053), compared against a black control trace.

• A stated caveat: The authors note that enzymatic activity was confirmed in only two of the seven sequences, and that the presence of oxygen in the product chemical formula means these cannot yet be confirmed as canonical TPS enzymes.

• Hyperparameter tuning results: Learning rate 1e-4 produced training loss 6.1 and evaluation loss 7.5, chosen because higher rates (e.g., 1e-3, training loss 4.2, evaluation loss 7.4) increased overfitting; 4,000 max steps converged at training loss 4.94 and evaluation loss 7.32.

Methodology in Plain English

The authors started with 1125 experimentally validated TPS sequences and used them as a seed to mine 79k homologous TPS sequences from UniProt, applying filters such as retaining only sequences 300–1100 amino acids long, requiring characteristic catalytic motifs (DDXXD, NSE/DTE, DXDD), and discarding sequences more than 80% identical to known isoprenyl diphosphate synthases. They used SpanSeq to split the data at 30% sequence identity into roughly 63k training sequences and roughly 16k validation sequences (an 80/20 split), which prevents the model from simply memorizing near-duplicates.

They then fine-tuned ProtGPT2 Tiny — a distilled version of the 738-million-parameter ProtGPT2 with 38.9 million parameters — using the HuggingFace run_clm.py script on a single NVIDIA L4 GPU, with block size 512 tokens, batch size 64, gradient accumulation of 8 (an effective batch of 512), and 4,000 training steps.

From the model they generated 28k sequences, then applied filters in stages: keep the top 10% by perplexity (2,800 sequences), keep only those with at most 60% identity to training sequences, require an EnzymeExplorer TPS score of at least 0.7, require a CLEAN-predicted EC number mapping to a terpenoid pathway in BRENDA, require an InterPro TPS-related domain, require an ESMFold pLDDT of at least 70, and require a Foldseek TM-score between 0.6 and 0.9 against the best training-set structural match. Seven sequences survived. Their genes were expressed in a yeast strain engineered to overproduce the diterpene precursor geranylgeranyl pyrophosphate, and the resulting metabolites were analyzed by LC-MS.

Why This Matters

Impact on research: The work tests whether a small, fine-tuned protein language model can push into underexplored regions of protein sequence space rather than producing variants closely related to natural proteins. If the approach generalizes, it offers a cheap computational route to enzyme candidates for families with few characterized members.

Real-world applications:

• Producing terpene scaffolds for flavors, fragrances, and pharmaceutical intermediates through microbial biosynthesis rather than multi-step chemical synthesis.
• Supporting synthetic biology routes to high-value diterpenes such as Taxol, whose chemical synthesis is described as structurally complex and resource-intensive.
• Enabling enzyme discovery for underrepresented protein families where few experimental sequences exist.
• Reducing the cost of early-stage enzyme screening by using in silico filtering before committing to wet-lab experiments.

Industry relevance: The under-$200 GPU cost and the use of a 38.9-million-parameter distilled model lower the barrier to entry for industrial and academic labs that lack large compute budgets. The pipeline's emphasis on evolutionary distance is relevant to intellectual property and novelty considerations in enzyme development.

Future Directions

• Characterizing the catalytic mechanisms of TpsGPT1 and TpsGPT2 in more detail, since their oxygen-containing products currently prevent confirmation as canonical TPS enzymes.
• Validating the remaining five generated sequences that passed computational filters but have not yet shown activity.
• Conditioning TPS generation on terpene subclasses using curated datasets, so the model can target specific terpene products rather than general TPS-like output.
• Generalizing the methodology to other protein families, with lysozymes named explicitly as an example, to explore functional diversity beyond terpene synthases.
• Refining the in silico pipeline based on the results of ongoing experimental validation.

Target Audience

This paper is most useful for computational biologists and protein engineers working on generative protein design, machine learning researchers interested in domain-specific fine-tuning of language models, and synthetic biologists or metabolic engineers looking for enzyme candidates for terpene biosynthesis. It is also relevant to researchers in enzyme discovery who want a concrete example of how in silico filtering metrics (perplexity, pLDDT, TM-score, EnzymeExplorer, CLEAN, InterPro) can be combined into a practical screening pipeline, and to those evaluating the tradeoffs of using small distilled models instead of large pretrained ones.

Authors’ abstract

Terpene synthases (TPS) are a key family of enzymes responsible for generating the diverse terpene scaffolds that underpin many natural products, including front-line anticancer drugs such as Taxol. However, de novo TPS design through directed evolution is costly and slow. We introduce TpsGPT, a generative model for scalable TPS protein design, built by fine-tuning the protein language model ProtGPT2 on 79k TPS sequences mined from UniProt. TpsGPT generated de novo enzyme candidates in silico and we evaluated them using multiple validation metrics, including EnzymeExplorer classification, ESMFold structural confidence (pLDDT), sequence diversity, CLEAN classification, InterPro domain detection, and Foldseek structure alignment. From an initial pool of 28k generated sequences, we identified seven putative TPS enzymes that satisfied all validation criteria. Experimental validation confirmed TPS enzymatic activity in at least two of these sequences. Our results show that fine-tuning of a protein language model on a carefully curated, enzyme-class-specific dataset, combined with rigorous filtering, can enable the de novo generation of functional, evolutionarily distant enzymes.