Ailin0103/CodonBERT-Cambray-expression-regression

CodonBERT for Cambray protein-expression regression

Pretrained lhallee/CodonBERT, specialized by full supervised fine-tuning on synthetic E. coli sequences for continuous protein-expression regression. This card documents the existing fine-tuned model.

Try it in your browser

Open the live protein-expression demo. Paste a DNA/RNA coding sequence and click Predict expression. The result is shown on a simple 1–100 expression bar; no installation is required.

Training data

Cambray et al. (2018), Supplementary Data 15: 242,269 usable sequences; 193,815 training, 24,227 validation, and 24,227 test examples. Input: gs.sequence; target: clean.lin.prot.mean. Split seed 42, 80/10/10. Raw training data are not redistributed with the model.

DNA is uppercased, stripped, converted to RNA, and split into non-overlapping codons. Invalid bases and non-frame-aligned sequences are excluded. Maximum length is 64 tokens including special tokens. Training uses learning rate 1e-5, three epochs, batch size 16, and training-set target standardization.

Evaluation

Previously reported held-out test metrics, on the original expression scale:

RMSE	R²	Pearson	Spearman
14.0871	0.7186	0.8554	0.8557

No new experiments or test evaluations were performed to prepare this release.

Intended use and limitations

Research and exploratory expression prediction on related coding-sequence distributions. Training covers a specific synthetic E. coli library. External generalization has not been established. Predictions are not experimentally validated expression measurements for arbitrary organisms or sequence distributions. Confirm upstream model and dataset license terms before redistribution; this card does not assert a new license for upstream materials.

Inference

import json
from huggingface_hub import hf_hub_download
import torch
from transformers import AutoTokenizer, BertForSequenceClassification

model_dir = 'Ailin0103/CodonBERT-Cambray-expression-regression'
tokenizer = AutoTokenizer.from_pretrained(model_dir)
model = BertForSequenceClassification.from_pretrained(model_dir).eval()
scaler = json.loads(open(hf_hub_download(model_dir, 'target_scaler.json')).read())
sequence = 'ATGGCC'  # Replace with an appropriate related coding sequence.
rna = sequence.upper().strip().replace('T', 'U')
if not rna or len(rna) % 3 or set(rna) - set('AUGC') or len(rna) // 3 > 62:
    raise ValueError('Expected a valid, frame-aligned coding sequence of <=62 codons')
formatted = ' '.join(rna[i:i+3] for i in range(0, len(rna), 3))
inputs = tokenizer(formatted, return_tensors='pt', truncation=False)
with torch.no_grad():
    standardized_prediction = model(**inputs).logits.item()
expression = standardized_prediction * scaler['target_std'] + scaler['target_mean']
print(expression)

Scaling is reversed as expression = prediction_z * training_std + training_mean. The standard deviation uses pandas' sample convention (ddof=1). Actual values are stored in target_scaler.json; never estimate them from inference examples.

Reference

Cambray G, Guimaraes JC, Arkin AP. Evaluation of 244,000 synthetic sequences reveals design principles to optimize translation in Escherichia coli. Nature Biotechnology (2018).

Upstream license

The base model carries academic/non-commercial restrictions. The upstream IP License Notice is preserved in LICENSE; this release does not grant additional rights. Source: https://huggingface.co/lhallee/CodonBERT