Research map / Probabilistic and scientific modeling / Molecules and proteins

Lifting Biomolecular Data Acquisition

NeurIPS 2026 · Probabilistic and scientific modeling › Molecules and proteins · approach: Generative protein structure design · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes LIFT, which pools multiple antibody/cell-therapy sequences per cell at high MOI and learns a neural network molecule-activity map via functional L1-regularized (neural LASSO) training, with theory and wet-lab validation on MAGE-A4 binding.

Problem

Learning sequence-activity maps from wet-lab screens is bottlenecked by the number of individually tested molecules; each measurement typically tests only one sequence.

Approach

Lift compressed screens from finite vectors to functions: test random Poisson mixtures of molecules per measurement (high MOI, droplet overloading) and fit a neural network under a functional L1 prior via MAP, so the additive mixture observations deconvolute into per-molecule activity.

Key results

• In simulation, LIFT at high MOI improves PR AUC from about 0.3 to 0.8 while conventional training degrades as MOI increases.
• In wet-lab MAGE-A4 binding experiments, LIFT at the highest MOI increases PR AUC by about 50% over conventional training on low-MOI data (some recall values nearly 100%).
• Ablating L1 regularization drops PR AUC from 0.092 to 0.006 in the wet-lab experiment.

Surprising findings

• Conventional training degrades as MOI increases, so high MOI only helps when paired with the functional L1 training algorithm.
• The LeaVS likelihood correction approximately acts as a functional L1 regularizer under Poisson noise.

Limitations

• Assumes additive activity across molecules in a mixture; violated by interactions such as drug-drug effects.
• Assumes sparsity of activity; violated when outcomes are graded.
• Excessively high MOI can cause toxicity and cell death, capping gains.

Experimental setup

Datasets: synthetic antibody CDRH3 simulations (variational synthesis model) in-house TCRm antibody scFv CAR-T screen against pHLA targets

Benchmarks: synthetic antibody binding simulation MAGE-A4 pHLA binding wet-lab held-out cells

Metrics: PR AUC precision-recall curve

Baselines: conventional (x,y) supervised training conventional training at low MOI LIFT without L1 regularization

Builds on: compressed sensing compressed screens LASSO variational synthesis lifting trick

Closest papers by meaning

• Mixture of neural fields for heterogeneous reconstruction in cryo-EM (NeurIPS 2024 · similarity 0.83)
• Efficient Biological Data Acquisition through Inference Set Design (ICLR 2025 · similarity 0.83)
• Learning Molecular Representation in a Cell (ICLR 2025 · similarity 0.82)
• How Molecules Impact Cells: Unlocking Contrastive PhenoMolecular Retrieval (NeurIPS 2024 · similarity 0.82)
• Monroe: A Molecular Foundation Model for In-context Probabilistic Inference (NeurIPS 2026 · similarity 0.81)
• A Systematic Evaluation of Co-folding Model Representations for Small-Molecule Learning (NeurIPS 2026 · similarity 0.81)