BindScreen

A screening campaign asks one question of a million molecules: which of these bind this target? A pair-based sequence model answers it a million times, running the protein and one molecule through the network together for every candidate. A CLIP-style dual encoder answers it once, embedding proteins and molecules separately so that ranking becomes a dot product — but CLIP's training recipe assumes each image has exactly one caption. Protein–ligand binding is not like that. A single kinase has thousands of known binders, and each of those binders hits several proteins, so a batch assembled the CLIP way is full of true binders labelled as negatives and pushed apart.

BindScreen, from Gabriel Bianchin de Oliveira and Fahad Saeed at Florida International University and posted to bioRxiv in August 2026, argues that the decisive fix is not a cleverer loss but how the batch is built. Its protein-centric batch construction assembles each batch around a set of proteins and pulls in all of their known binders, decoupling the number of proteins from the number of molecules per step. On top of that sits an asymmetric multi-positive InfoNCE objective that credits every true binder of a target in the batch as a positive instead of penalising it.

A factorial ablation separates the two contributions and lands on an unusual conclusion: the loss alone actually degrades performance under standard CLIP batching, the protein-centric batch alone recovers most of the gain, and the two together perform best. The effect is encoder-agnostic, holding across eight protein language models spanning four architectural families.

#Key Features

• Protein-centric batching: Batches are built around targets and their binder sets rather than sampled as one-to-one pairs, which both matches the structure of the task and decouples protein count from molecule count.
• Asymmetric multi-positive InfoNCE: The objective admits many positives per protein and treats the two directions of the contrast differently, reflecting that binding is many-to-many rather than symmetric.
• Library scoring as a matrix product: Proteins and molecules are encoded independently once, so ranking a library is a single cosine-similarity matrix rather than one forward pass per pair.
• Sequence-only by design: Inputs are amino acid sequences and SMILES strings, so targets with no experimental or predicted structure are screenable.
• Two released configurations: A frozen variant that trains only the projection heads, and a finetuning variant that additionally adapts the protein encoder through LoRA.

#Technical Details

The protein branch mean-pools ESM-2 3B (facebook/esm2_t36_3B_UR50D) to a 2560-dimensional vector; the molecule branch mean-pools MolDeBERTa, the same lab's SMILES encoder, to 768 dimensions. Each passes through a two-layer projection head (Linear → LayerNorm → GELU → Dropout → Linear) into a shared 512-dimensional L2-normalised space — 2.23M trained parameters in total. In the frozen configuration both encoders are fixed and only those heads are trained; in the finetuning configuration a rank-16 LoRA adapter (alpha 32, dropout 0.05) is applied to ESM-2's query, key and value projections while MolDeBERTa stays frozen. Heads are trained per benchmark rather than once for all of them.

Training uses a filtered ChEMBL 36 set split by protein sequence similarity: 1,492 proteins, 394,190 molecules and 583,960 positive pairs for training, with 326 proteins and 130,281 positive pairs held out for test. The LIT-PCBA evaluation covers 15 proteins, 404,586 molecules and 2,776,973 pairs. The preprint reports up to a 39% relative improvement in enrichment factor at 0.5% over the best sequence-based baseline, training cost cut by up to 32x, higher validation BEDROC after 86 hours than standard CLIP training reaches in 460, and roughly seven times fewer forward passes to screen LIT-PCBA than pair-based models — while remaining competitive with docking without using 3D structure.

#Applications

BindScreen suits early-stage hit finding against targets where structure is unavailable, low-confidence or conformationally awkward — orphan receptors, disordered regions, newly sequenced pathogen proteins. Because embeddings are computed once per protein and once per molecule, a group can cache a library's molecule embeddings and re-rank it against a new target for the cost of one encoder pass plus a matrix multiply, making repeated screens across a fixed in-house collection practical on a single GPU.

#Impact

BindScreen's contribution is a diagnosis more than a benchmark: it identifies batch construction, usually treated as plumbing, as the component that has to change when a symmetric contrastive recipe is transplanted onto an asymmetric biological relation, and shows the effect survives a change of encoder. The evaluation is retrospective and in silico, the model is sequence-only with no structural input, and the trained parameters are two small projection heads over externally pretrained backbones rather than a from-scratch model. Code, checkpoints and the benchmark splits are all public, though under a CC-BY-NC-ND 4.0 licence that bars commercial use and derivatives, and the preprint is still under review.