📢🧬Looking for fast and reliable ways to predict binding affinities? Meet BALM!👋
BALM predicts binding affinity from protein sequences and ligand SMILES using pretrained protein and ligand language models.
📄 doi.org/10.1101/2024.11.01.6…
🔗 github.com/meyresearch/BALM
🧵1/9
Antibody–Antigen Affinity Prediction with Chain-Aware Protein Language Modeling
1. The paper introduces AbAffinity, a sequence-only affinity predictor that keeps antibody heavy chain, light chain, and antigen as three separate streams (instead of early concatenation), aiming to preserve chain-specific paratope signals and antigen epitope signals during modeling.
2. Core modeling idea: use frozen ESM-2 (650M) residue embeddings, then train only a compact “interaction module” that performs heavy-chain CDR-focused pooling, heavy–light self-attention, adaptive heavy/light fusion gating, and gated cross-attention to model partner-specific antibody–antigen compatibility.
3. A notable biological inductive bias is heavy-chain CDR-focused pooling (CDR-H1/H2/H3) rather than full-chain pooling; this concentrates the representation on likely paratope residues while keeping the antigen pooled globally because the epitope is unknown at prediction time.
4. Architecture details: heavy and light chain are treated as a 2-token “antibody sequence” for self-attention; a learned softmax fusion gate produces an antibody embedding; two stacked gated cross-attention blocks refine an antibody-conditioned antigen embedding; cosine similarity in a shared latent space is mapped to predicted pKd.
5. Main benchmark (SAAINT-DB, 2,575 complexes from 1,271 PDBs; includes 1,913 paired antibodies and 662 nanobodies): AbAffinity reports strong random 10-fold CV performance (Pearson r 0.86, Spearman 0.84, RMSE 0.69) and improved robustness on antigen-cold splits (Pearson r 0.57, Spearman 0.55), indicating partial generalization to unseen antigens.
6. External generalization: across SAbDab, AB-Bind, SKEMPI 2.0, and a held-out docking/affinity benchmark, AbAffinity consistently outperforms the compared sequence baseline MVSF-AB on correlation and (where reported) RMSE, supporting transfer beyond the primary training distribution.
7. Affinity maturation setting: on AbBiBench landscapes, zero-shot transfer is target-dependent (can be weak or even negative), but few-shot fine-tuning with only 10–30% labeled variants substantially improves within-target ranking (e.g., on 1n8z reaching ~0.69 Pearson and ~0.70 Spearman at 30% labels).
8. Mechanistic ablations/interventions support the design choices: (i) preserving heavy vs light chain identity beats collapsing them; (ii) CDR-focused pooling improves over full-chain mean pooling; (iii) closing the learned gate sharply degrades performance, and shuffling antigens across antibodies causes major drops—evidence the model relies on correct partner pairing rather than antibody-only priors.
9. Interpretability: Integrated Gradients attributions (with differentiable pooling) localize importance to antibody CDRs and known antigen epitope residues in structurally validated complexes (e.g., D1.3–lysozyme), and recover more known paratope residues than a fused two-stream baseline, especially improving light-chain attribution sharpness.
💻Code: github.com/harshitsinghsnu/A…
📜Paper: biorxiv.org/content/10.64898…#ComputationalBiology#Bioinformatics#Antibodies#ProteinLanguageModels#DeepLearning#TherapeuticDiscovery#AffinityPrediction#ExplainableAI#ProteinEngineering
We are releasing Carbon: a crazy fast DNA model
Carbon is 275x faster than the next best model. So fast you can process the whole human genome on a single GPU in <2 days.
Here are the tricks we used:
When modelling DNA sequences a lot of the performance comes down to tokenizing the sequences in a smart way. BPE tokenizer struggle because there are no whitespaces and character (called base in DNA) level tokenizers waste a lot of compute on too many tokens.
Carbon is built with a unique tokenizer: we split sequences in chunks of 6 bases, but during both training and inference we can work with single base resolution. That's similar to having word tokens but resolving them at the character level. All possible thanks to the DNA tokens unique structure.
The architecture combined with the tokenizer makes the model 275x faster than the previous SoTA (Evo2) at this size.
We built an interactive demo so you can explore how the model can generate DNA sequences, investigate the structure of genes, predict the effect of mutations, generate and fold proteins and even reconstruct parts of the tree of life.
huggingface.co/spaces/Huggin…
We just published our work on an explainable active learning framework for ligand–protein binding affinity prediction in Digital Discovery.
🔗 pubs.rsc.org/en/content/arti…
Here’s a quick breakdown of what we did and why 👇
Apply for the AITHYRA-CeMM International PhD Program!
15-20 fully funded PhD fellowships available in Vienna in AI/ML and Life Sciences
Deadline for applications: 10 September 2025
apply.cemm.at/
New Anthropic Research: “Inverse Scaling in Test-Time Compute”
We found cases where longer reasoning leads to lower accuracy.
Our findings suggest that naïve scaling of test-time compute may inadvertently reinforce problematic reasoning patterns.
🧵
🙌 Special shoutout to all my co-authors, @aryopg@Xi_Yang_Ian, @jjuarez_jimenez, Ben Suutari, Álvaro
Morrás for all their hard work. Big thanks to @ppxasjsm for her incredible support!
🧵9/9
🔍 Comparing with Docking
BALM achieves strong predictive accuracy on the LP-PDBBind test set, showing promising results compared to docking methods like AutoDock Vina and rDock, especially in rescoring of binding poses for different target families.
🧵8/9
⚡ Quick Adaptation with Few-shot Tuning
With few-shot tuning on USP7 and Mpro, BALM adapts quickly—updating <1% of parameters (~152.5M). Unlike Gaussian Processes (GP), BALM performs reliably across targets without sensitivity to kernel/fingerprint adjustments.
🧵7/9
🚀 Efficient Zero-shot
BALM’s zero-shot predictions on unseen targets USP7 and Mpro (~2K ligands) complete in 1.5 minutes on an NVIDIA A100 (or 5-10 mins on CPU) at ~0.15-0.4 seconds per prediction. Zero-shot is not always reliable; works well for USP7 but not for Mpro.
🧵6/9
🔍📉 BindingDB performance
BALM learns better than the baseline on challenging data splits involving unseen drugs, scaffolds and targets on the BindingDB (25K Kd interactions with ~1.1K targets and 9.2K ligands after cleaning the assay limits). PEFT enhances performance.
🧵4/9
💻 🧪 BALM learns affinities by optimising cosine similarity in a shared embedding space. Parameter-efficient fine-tuning (PEFT) helps to adapt these embeddings for binding by achieving a balance between accuracy and computational efficiency.
🧵3/9
💡Why BALM? While structure-based methods have limitations, they are widely used compared to recent deep learning (DL) models due to generalisability issues. To make DL models reliable for screening, we address challenges at model, data and evaluation levels.
🧵2/9