PLAPT

A virtual screen usually starts with less information than a scoring function wants. The target is known as an amino acid sequence, the candidates arrive as SMILES strings, and there is no co-crystal structure, no docked pose, and no annotated binding pocket. Structure-based affinity predictors need all three, which puts a docking pipeline between a compound library and its first ranking. PLAPT takes the opposite premise: that a protein language model has already encoded enough about a chain's fold and its cavities for a small regressor to read affinity off the embedding directly.

The construction is deliberately minimal. A protein sequence passes through ProtBERT, one of the ProtTrans protein language models, and a ligand SMILES string through ChemBERTa. Both encoders are frozen, contributing a 1024-dimensional and a 768-dimensional pooler vector that are concatenated into a 1792-dimensional feature. A branching neural network then refines the protein and ligand halves in separate linear streams before fusing them down to one scalar — the normalized negative base-10 logarithm of the dissociation constant. Nothing in the backbones is fine-tuned, so the only weights PLAPT itself learns are the roughly 1.5 million in that head.

Released in February 2024 as a bioRxiv preprint, PLAPT came out of the Wolfram Emerging Leaders Program, written by Tyler Rose of Wolfram Research, Nicolò Monti of ASC27, and two secondary-school students, Navvye Anand and Tianyu Shen. The contribution is an argument about how cheaply a competitive affinity predictor can be assembled from encoders somebody else pretrained, not a new backbone.

#Key Features

• Two strings in, one number out: Inference needs only the target's amino acid sequence and the ligand's canonical SMILES, so a library is scored straight from a FASTA file and a compound table.
• Frozen backbones, tiny head: Both encoders run under inference mode and are never updated, which is what lets the prediction module train in about three minutes on a single consumer GPU.
• Branching prediction module: Protein and ligand features each pass through their own 512-node layer before merging, then through batch normalization, dropout, and two further layers.
• Embedding cache for screening: Encoder outputs are memoized to disk keyed by sequence and SMILES, so a fixed target is embedded once and every additional candidate costs only the ligand pass plus a sub-millisecond head evaluation.
• Distributed as a fixed artifact: The trained head ships in the repository as a 5.9 MB ONNX file behind a command-line tool, a Python API, and a Docker image, with no retraining step in the way.

#Technical Details

The head was trained on 100,000 protein-ligand pairs drawn from Glaser et al.'s binding affinity dataset, a 1.9-million-pair aggregation of BindingDB, PDBbind-CN, BioLIP, and Binding MOAD, split 90/10 into training and validation. Protein sequences are truncated or padded to 3,200 tokens and SMILES to 278; roughly 0.2% of sequences exceed the protein limit. Training used Wolfram Language's NetTrain with mean squared error, the Adam optimizer, a learning rate of 0.001, batch size 256, and 60 rounds, finishing in about three minutes on an RTX 4060 Ti after a twelve-hour embedding pass on an RTX 2060. On the Test2016_290 benchmark PLAPT reports a Pearson R of 0.845, RMSE 1.196, and MAE 0.906, against 0.843 / 1.200 / 0.966 for CAPLA; on CSAR-HiQ_36 it reports R 0.731, RMSE 1.349, and MAE 1.157. Validation R was 0.683 against 0.886 on the training split, a gap the authors attribute to overfitting.

#Applications

The natural use is triage: score a target against thousands of purchasable compounds before committing to docking or free energy calculations, which is what the caching design and the score_candidates entry point are built for. Because it is small, permissively licensed, and fast enough to sit inside a training loop, it has also been adopted as a reward signal for generative chemistry — DrugGen uses PLAPT as a frozen affinity oracle in its reinforcement learning objective — and as a screening filter in target-prioritization work on druggable cancer-driving proteins.

#Impact

PLAPT demonstrates that most of the representational work in sequence-based affinity prediction can be delegated to existing protein and chemical language models, and it is reused precisely because it is cheap to run and ships as a working checkpoint rather than a training recipe. The caveats are substantial. The preprint was never peer-reviewed and its benchmark comparisons are self-reported; the authors note that 177 of the 290 Test2016 pairs also appear in a competing model's training data, and the drop from training to validation correlation points at overfitting on a 100,000-pair subset of a far larger corpus. Using 1D sequence alone, the model cannot distinguish binding sites within a target or reason about a pose, so it ranks candidates rather than explaining them. Later structure-aware predictors such as PBCNet2.0 treat it as a sequence-based baseline.