Towards scientific discovery with dictionary learning: Extracting biological concepts from microscopy foundation models

Dictionary learning (DL) has emerged as a powerful interpretability tool for large language models. By extracting known concepts (e.g., Golden-Gate Bridge) from human-interpretable data (e.g., text), sparse DL can elucidate a model's inner workings. In this work, we ask if DL can also be used to discover unknown concepts from less human-interpretable scientific data (e.g., cell images), ultimately enabling modern approaches to scientific discovery. As a first step, we use DL algorithms to study microscopy foundation models trained on multi-cell image data, where little prior knowledge exists regarding which high-level concepts should arise. We show that sparse dictionaries indeed extract biologically-meaningful concepts such as cell type and genetic perturbation type. We also propose a new DL algorithm, Iterative Codebook Feature Learning~(ICFL), and combine it with a pre-processing step that uses PCA whitening from a control dataset. In our experiments, we demonstrate that both ICFL and PCA improve the selectivity of extracted features compared to TopK sparse autoencoders.

ViTally Consistent: Scaling Biological Representation Learning for Cell Microscopy

Large-scale cell microscopy screens are used in drug discovery and molecular biology research to study the effects of millions of chemical and genetic perturbations on cells. To use these images in downstream analysis, we need models that can map each image into a feature space that represents diverse biological phenotypes consistently, in the sense that perturbations with similar biological effects have similar representations. In this work, we present the largest foundation model for cell microscopy data to date, a new 1.9 billion-parameter ViT-G/8 MAE trained on over 8 billion microscopy image crops. Compared to a previous published ViT-L/8 MAE, our new model achieves a 60% improvement in linear separability of genetic perturbations and obtains the best overall performance on whole-genome biological relationship recall and replicate consistency benchmarks. Beyond scaling, we developed two key methods that improve performance: (1) training on a curated and diverse dataset; and, (2) using biologically motivated linear probing tasks to search across each transformer block for the best candidate representation of whole-genome screens. We find that many self-supervised vision transformers, pretrained on either natural or microscopy images, yield significantly more biologically meaningful representations of microscopy images in their intermediate blocks than in their typically used final blocks. More broadly, our approach and results provide insights toward a general strategy for successfully building foundation models for large-scale biological data.

Benchmarking Transcriptomics Foundation Models for Perturbation Analysis : one PCA still rules them all

Understanding the relationships among genes, compounds, and their interactions in living organisms remains limited due to technological constraints and the complexity of biological data. Deep learning has shown promise in exploring these relationships using various data types. However, transcriptomics, which provides detailed insights into cellular states, is still underused due to its high noise levels and limited data availability. Recent advancements in transcriptomics sequencing provide new opportunities to uncover valuable insights, especially with the rise of many new foundation models for transcriptomics, yet no benchmark has been made to robustly evaluate the effectiveness of these rising models for perturbation analysis. This article presents a novel biologically motivated evaluation framework and a hierarchy of perturbation analysis tasks for comparing the performance of pretrained foundation models to each other and to more classical techniques of learning from transcriptomics data. We compile diverse public datasets from different sequencing techniques and cell lines to assess models performance. Our approach identifies scVI and PCA to be far better suited models for understanding biological perturbations in comparison to existing foundation models, especially in their application in real-world scenarios.

Masked Autoencoders for Microscopy are Scalable Learners of Cellular Biology

Featurizing microscopy images for use in biological research remains a significant challenge, especially for large-scale experiments spanning millions of images. This work explores the scaling properties of weakly supervised classifiers and self-supervised masked autoencoders (MAEs) when training with increasingly larger model backbones and microscopy datasets. Our results show that ViT-based MAEs outperform weakly supervised classifiers on a variety of tasks, achieving as much as a 11.5% relative improvement when recalling known biological relationships curated from public databases. Additionally, we develop a new channel-agnostic MAE architecture (CA-MAE) that allows for inputting images of different numbers and orders of channels at inference time. We demonstrate that CA-MAEs effectively generalize by inferring and evaluating on a microscopy image dataset (JUMP-CP) generated under different experimental conditions with a different channel structure than our pretraining data (RPI-93M). Our findings motivate continued research into scaling self-supervised learning on microscopy data in order to create powerful foundation models of cellular biology that have the potential to catalyze advancements in drug discovery and beyond.

Masked autoencoders are scalable learners of cellular morphology

Inferring biological relationships from cellular phenotypes in high-content microscopy screens provides significant opportunity and challenge in biological research. Prior results have shown that deep vision models can capture biological signal better than hand-crafted features. This work explores how weakly supervised and self-supervised deep learning approaches scale when training larger models on larger datasets. Our results show that both CNN- and ViT-based masked autoencoders significantly outperform weakly supervised models. At the high-end of our scale, a ViT-L/8 trained on over 3.5-billion unique crops sampled from 95-million microscopy images achieves relative improvements as high as 28% over our best weakly supervised models at inferring known biological relationships curated from public databases.

Deconstructing word embedding algorithms

Word embeddings are reliable feature representations of words used to obtain high quality results for various NLP applications. Uncontextualized word embeddings are used in many NLP tasks today, especially in resource-limited settings where high memory capacity and GPUs are not available. Given the historical success of word embeddings in NLP, we propose a retrospective on some of the most well-known word embedding algorithms. In this work, we deconstruct Word2vec, GloVe, and others, into a common form, unveiling some of the common conditions that seem to be required for making performant word embeddings. We believe that the theoretical findings in this paper can provide a basis for more informed development of future models.

Learning Efficient Task-Specific Meta-Embeddings with Word Prisms

Word embeddings are trained to predict word cooccurrence statistics, which leads them to possess different lexical properties (syntactic, semantic, etc.) depending on the notion of context defined at training time. These properties manifest when querying the embedding space for the most similar vectors, and when used at the input layer of deep neural networks trained to solve downstream NLP problems. Meta-embeddings combine multiple sets of differently trained word embeddings, and have been shown to successfully improve intrinsic and extrinsic performance over equivalent models which use just one set of source embeddings. We introduce word prisms: a simple and efficient meta-embedding method that learns to combine source embeddings according to the task at hand. Word prisms learn orthogonal transformations to linearly combine the input source embeddings, which allows them to be very efficient at inference time. We evaluate word prisms in comparison to other meta-embedding methods on six extrinsic evaluations and observe that word prisms offer improvements in performance on all tasks.

Deconstructing and reconstructing word embedding algorithms

Uncontextualized word embeddings are reliable feature representations of words used to obtain high quality results for various NLP applications. Given the historical success of word embeddings in NLP, we propose a retrospective on some of the most well-known word embedding algorithms. In this work, we deconstruct Word2vec, GloVe, and others, into a common form, unveiling some of the necessary and sufficient conditions required for making performant word embeddings. We find that each algorithm: (1) fits vector-covector dot products to approximate pointwise mutual information (PMI); and, (2) modulates the loss gradient to balance weak and strong signals. We demonstrate that these two algorithmic features are sufficient conditions to construct a novel word embedding algorithm, Hilbert-MLE. We find that its embeddings obtain equivalent or better performance against other algorithms across 17 intrinsic and extrinsic datasets.

Word Embedding Algorithms as Generalized Low Rank Models and their Canonical Form

Word embedding algorithms produce very reliable feature representations of words that are used by neural network models across a constantly growing multitude of NLP tasks. As such, it is imperative for NLP practitioners to understand how their word representations are produced, and why they are so impactful. The present work presents the Simple Embedder framework, generalizing the state-of-the-art existing word embedding algorithms (including Word2vec (SGNS) and GloVe) under the umbrella of generalized low rank models. We derive that both of these algorithms attempt to produce embedding inner products that approximate pointwise mutual information (PMI) statistics in the corpus. Once cast as Simple Embedders, comparison of these models reveals that these successful embedders all resemble a straightforward maximum likelihood estimate (MLE) of the PMI parametrized by the inner product (between embeddings). This MLE induces our proposed novel word embedding model, Hilbert-MLE, as the canonical representative of the Simple Embedder framework. We empirically compare these algorithms with evaluations on 17 different datasets. Hilbert-MLE consistently observes second-best performance on every extrinsic evaluation (news classification, sentiment analysis, POS-tagging, and supersense tagging), while the first-best model depends varying on the task. Moreover, Hilbert-MLE consistently observes the least variance in results with respect to the random initialization of the weights in bidirectional LSTMs. Our empirical results demonstrate that Hilbert-MLE is a very consistent word embedding algorithm that can be reliably integrated into existing NLP systems to obtain high-quality results.

Clustering-Oriented Representation Learning with Attractive-Repulsive Loss

The standard loss function used to train neural network classifiers, categorical cross-entropy (CCE), seeks to maximize accuracy on the training data; building useful representations is not a necessary byproduct of this objective. In this work, we propose clustering-oriented representation learning (COREL) as an alternative to CCE in the context of a generalized attractive-repulsive loss framework. COREL has the consequence of building latent representations that collectively exhibit the quality of natural clustering within the latent space of the final hidden layer, according to a predefined similarity function. Despite being simple to implement, COREL variants outperform or perform equivalently to CCE in a variety of scenarios, including image and news article classification using both feed-forward and convolutional neural networks. Analysis of the latent spaces created with different similarity functions facilitates insights on the different use cases COREL variants can satisfy, where the Cosine-COREL variant makes a consistently clusterable latent space, while Gaussian-COREL consistently obtains better classification accuracy than CCE.