README.md

---
language:
  - en
license: mit
tags:
  - biology
  - protein
  - esm2
  - plant
  - viridiplantae
  - masked-language-modeling
  - domain-adaptation
base_model: facebook/esm2_t30_150M_UR50D
datasets:
  - uniprot-trembl-viridiplantae
pipeline_tag: fill-mask
---

# PlantPLM-150M

<img src="Plant_PLM_logo.png" alt="Alt Text" width="800">

**ESM-2 150M parameter model continued-pretrained on Viridiplantae (plant) protein sequences.**

This is a domain-adapted version of [`facebook/esm2_t30_150M_UR50D`](https://huggingface.co/facebook/esm2_t30_150M_UR50D), fine-tuned on a non-redundant subset of UniProt TrEMBL plant-kingdom proteins.

Part of the **[Plant-PLM](https://huggingface.co/collections/dipayan26/plant-plm)** - ESM-2 models at 8M, 35M, 150M, and 650M parameters, each adapted on plant protein data.

---

## Model Description

| Property | Value |
|---|---|
| Base model | `facebook/esm2_t30_150M_UR50D` |
| Architecture | ESM-2 · 30 layers · hidden=640 · heads=20 · FFN=2560 |
| Position embeddings | Rotary (RoPE) |
| Vocabulary | 33 tokens (20 standard + rare amino acids + special tokens) |
| Parameters | 148M (full-parameter continued pretraining) |
| Training objective | Masked Language Modeling (MLM, 15% masking) |

---

## Training Data

Unlike the 8M and 35M variants (trained on the raw, redundant plant TrEMBL corpus), this model was trained on a **redundancy-reduced ("nr50") corpus**: the raw Viridiplantae corpus was clustered with MMseqs2 `easy-linclust` (50% identity / 80% coverage, mirroring ESM-2's own training-data construction) and one representative sequence per cluster was kept.

| Property | Value |
|---|---|
| Source | UniProt TrEMBL — Viridiplantae (plant kingdom) subset, MMseqs2-deduplicated (50% ID / 80% cov) |
| Sequences | **4,372,758** (down from 19,938,415 raw, −78%) |
| Avg sequence length | 279 AA · median 199 AA |
| Token budget | **~1.11 billion** amino acid tokens (≈ 1 full epoch over the nr50 corpus) |

---

## Training Details

| Hyperparameter | Value |
|---|---|
| Training steps | 90,000 optimizer steps (1 epoch over nr50) |
| Batch size | 48 sequences (12 per micro-batch × 4 gradient accumulation steps) |
| Optimizer | AdamW · β=(0.9, 0.98) · ε=1e-8 · weight_decay=0.01 |
| Learning rate | 1e-5 |
| LR schedule | Linear warmup (1,000 steps) → linear decay |
| Gradient clipping | 1.0 |
| Precision | 16-bit mixed |
| Gradient checkpointing | Enabled |
| Hardware | 1× NVIDIA RTX 3060 (12 GB) |

**Final metrics (validation set, 5% holdout):**

| Metric | Value |
|---|---|
| `val/mlm_loss` | 2.185 |
| `val/perplexity` | 8.98 |
| `val/masked_token_acc` | 34.3% |

---

## Usage

```python
from transformers import EsmForMaskedLM, EsmTokenizer
import torch

model = EsmForMaskedLM.from_pretrained("dipayan26/PlantPLM-150M")
tokenizer = EsmTokenizer.from_pretrained("dipayan26/PlantPLM-150M")

# --- Masked token prediction ---
sequence = "MSPQTETKASVGFKAGVKDYKLTYYTPEYETK"
inputs = tokenizer(sequence, return_tensors="pt")

# mask one position
inputs["input_ids"][0, 5] = tokenizer.mask_token_id

with torch.no_grad():
    logits = model(**inputs).logits

masked_pos = (inputs["input_ids"] == tokenizer.mask_token_id).nonzero()[0, 1]
top5 = logits[0, masked_pos].topk(5)
print(tokenizer.convert_ids_to_tokens(top5.indices.tolist()))

# --- Sequence embedding ([CLS] token) ---
inputs = tokenizer(sequence, return_tensors="pt")
with torch.no_grad():
    hidden = model.esm(**inputs).last_hidden_state
cls_embedding = hidden[0, 0, :]   # shape: [640]
print("Embedding shape:", cls_embedding.shape)
```

---

## Intended Use

- **Plant protein function prediction** — GO term annotation, subcellular localization, signal peptide detection
- **Plant-specific protein embeddings** — clustering, retrieval, similarity search
- **Transfer learning starting point** — fine-tune on small labeled plant protein datasets

## Out-of-scope Use

- Non-plant organisms — the model has been shifted toward Viridiplantae statistics; use the original `facebook/esm2_t30_150M_UR50D` for general protein tasks
- Structural prediction — not trained for structure; use ESMFold for that

---

## Limitations

- No downstream benchmark evaluation has been run on this checkpoint yet


---

## Citation

If you use this model, please cite:

```bibtex
@misc{sarkar2026plantplm,
  author       = {Sarkar, Dipayan},
  title        = {PlantPLM: Domain-Adaptive Pretraining of ESM-2 on Viridiplantae Proteins},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/dipayan26/PlantPLM-150M}},
}
```