Accelerated Learning on Large Scale Screens using Generative Library Models

TL;DR

Biological machine learning is often data-limited, which LeaVS addresses by co-designing data generation and learning for large-scale sequence screens. By prioritizing measurements of active sequences ($q=1$) and incorporating a marginal likelihood term that leverages a generative library model $p(x)$, LeaVS achieves consistent estimation of $p(y\mid x)$ and dramatic information gains. The approach is validated through theoretical analysis (sparse activity and posterior concentration) and empirical demonstrations on synthetic data, TCR screening, and large-scale antibody experiments, where it yields superior accuracy, calibration, and enrichment of functional sequences. This co-design strategy enables orders-of-magnitude improvements in data efficiency, with implications for accelerating antibody discovery and therapeutic sequence design.

Abstract

Biological machine learning is often bottlenecked by a lack of scaled data. One promising route to relieving data bottlenecks is through high throughput screens, which can experimentally test the activity of $10^6-10^{12}$ protein sequences in parallel. In this article, we introduce algorithms to optimize high throughput screens for data creation and model training. We focus on the large scale regime, where dataset sizes are limited by the cost of measurement and sequencing. We show that when active sequences are rare, we maximize information gain if we only collect positive examples of active sequences, i.e. $x$ with $y>0$. We can correct for the missing negative examples using a generative model of the library, producing a consistent and efficient estimate of the true $p(y | x)$. We demonstrate this approach in simulation and on a large scale screen of antibodies. Overall, co-design of experiments and inference lets us accelerate learning dramatically.

Figures (9)

• Figure 1: Overview of LeaVS. a. In a large scale screen, sequences $x$ are delivered into cells for testing activity $y$. Each cell receives a sample $x \sim \mathrm{p}(x)$. b. We can measure $(x,y)$ for some cells, but only observe $y$ for the rest. We can choose the fraction $q$ of measurements to allocate to positive and negative examples. c. If $\mathrm{p}(y\mid x)$ is sparse and we collect a representative sample from the library ($q \approx \mathrm{p}(y=1)$), most $x$ will not be very informative. Allocating all measurements to positive examples ($q=1$) yields more information about $\mathrm{p}(y\mid x)$.
• Figure 2: Performance on synthetic data.
• Figure 3: Performance on TCR data.
• Figure 4: Predicting binding of TCRm scFv CAR therapeutic candidates to MAGE-A4, a challenging oncology target.
• Figure 5: Accuracy on synthetic data (\ref{['fig:accuracy_zoomed']}) with y-axis extended to zero.

Theorems & Definitions (11)

• Proposition 1: Consistent for any $q$
• Proposition 2: Inconsistent without $y$ data
• Theorem 1: Bernstein-von Mises
• Proposition 3
• Lemma 1: Nonparametric identification
• proof
• Proposition 4: $\theta_0$ is optimal
• proof
• Proposition 5: $\tilde{\theta}$ is optimal
• proof