Conference Agenda

Many problems in science can be phrased as `under what conditions does something of interest happen?'. In cancer research, we have therapies that work well for some patients but are very nasty for others. Wouldn't it be great if we can automatically find those subgroups of people with exceptional survival characteristics? In materials science, we are after highly effective photo-voltaic panels. Wouldn't it be great if we can automatically characterize the atomistic properties of the most effective molecules in an automated manner?

Each of the above requires solutions that outside of standard machine learning. Rather than optimally predicting a target value, we are specifically interested learning about actionable patterns from the data. In this talk I will discuss recent work that allows us to do exactly that. In particular, I will give a high-level introduction SyFlow (Xu et al. 2024) for learning exceptional subgroups and SySurv for learning survival subgroups. Both are inherently explainable, highly scalable, and very flexible.

Motivation: RNA-binding proteins (RBPs) regulate key post-transcriptional processes, yet many remain undiscovered because RNA-binding activity often occurs outside canonical RNA-binding domains (RBDs), including within intrinsically disordered regions (IDRs) or through protein complexes. Computational methods can help identify novel RBPs, but approaches relying solely on sequence-derived features or ignoring the cellular interaction context are limited in capturing the complexity of RNA-binding behavior. To date, no framework rigorously integrates both sequence information and protein interaction context for RBP prediction.
Results: We introduce RIBEX, a multimodal framework that combines protein language model (pLM) embeddings with protein interactome topology to improve RBP prediction and interpretation. Specifically, we integrate sequence representations with graph-derived positional encodings (PE) from the human STRING protein–protein interaction (PPI) network. PE are computed using Personalized PageRank, reduced with principal component analysis, and fused with pooled sequence embeddings through FiLM conditioning, while Low-Rank Adaptation (LoRA) enables parameter-efficient task adaptation.
Across both an annotation-based benchmark and experimental RNA Interactome Capture (RIC) dataset, positional encodings consistently improve predictive performance, indicating that interactome topology provides complementary information beyond sequence features alone. LoRA adaptation of ESM2-650M further yields larger gains than simply scaling frozen backbone size. RIBEX outperforms state-of-the-art methods such as RBP-TSTL and HydRA, particularly on challenging subsets including proteins lacking canonical RBDs and those enriched in intrinsically disordered regions. For interpretability, we combine sequence-level computational alanine scanning with network-level positional-encoding ablation and inverse-PCA mapping, recovering known domain-centered signals, highlighting IDR-associated contributions, and identifying functional interactome communities linked to RBP predictions.

While models like RFdiffusion excel at generating protein backbones, their "black box" nature currently restricts design to a process of stochastic sampling rather than precise engineering. To bridge this gap, we introduce FoldSAE, a framework that adapts Sparse Autoencoders (SAEs) to decompose RFdiffusion’s dense activations into interpretable, monosemantic features. We demonstrate that these unsupervised features capture fundamental physical properties, including secondary structure formation and solvent-accessible surface area (SASA). Leveraging these insights, we implement a steering mechanism that enables targeted modulation of backbone folding and surface exposure during the denoising process, both in de novo and anlogue generation. Our work pioneers a new framework for making RFdiffusion more interpretable, demonstrating how understanding internal features can be directly translated into precise control over the protein design process.

Modern bioinformatics increasingly relies on representation learning to extract mechanistic insights from high-dimensional biological data. In single-cell biology, representation learning models integrate multiple modalities into unified latent spaces that are assumed to capture underlying molecular and disease mechanisms. However, strong performance on downstream predictive tasks does not guarantee that latent variables encode biologically meaningful factors. To address the need for explainability in biological representation learning, we propose a framework to inspect, evaluate, and interpret multi-modal latent representations. The framework adapts feature attribution methods, including gradient-based approaches (saliency maps, integrated gradients) and SHAP-based methods (DeepLIFT SHAP, Gradient SHAP), to quantify the contribution of genes, pathways, and other molecular features to individual latent dimensions. This enables the identification of pathway and gene activation patterns both globally and within specific tissue contexts. We apply the framework to Multimodal Integration with Modality-agnostic Autoencoders (MIMA), a state-of-the-art multi-omic integration model developed and published from our lab, using prostate cancer multi-modal datasets. Our analysis identifies latent dimensions associated with prostate cancer-related genes and pathways and reveals interactions between gene expression and lipid concentrations. We further quantify the relative contribution of each modality to these latent factors. In addition, we incorporate gene embeddings derived from large language models to disentangle latent dimensions into networks of interacting gene-function clusters, enabling more interpretable biological inspection. The proposed framework is model-agnostic and applicable to a broad class of biological representation learning models, with a focus on genomics and multi-omics integration. By combining feature attribution methods, advances in multi-modal representation learning, and large biological knowledge bases, this approach provides a practical pathway toward trustworthy and interpretable AI systems capable of extracting biologically meaningful and clinically relevant insights from high-dimensional omics data.

Biological foundation models provide rich representations of genes, proteins, molecules, cells, images, and text, but using these embeddings reproducibly in perturbation modeling remains difficult. We present embpy, a unified Python framework for resolving biological identifiers, generating embeddings from 130+ foundation models, storing them in provenance-tracked embedding stores, and integrating them with AnnData workflows. Building on this infrastructure, we develop a perturbation world model for single-cell transcriptomics. The model represents transcriptomic context as stacked cellular state tokens, encodes genetic perturbations as embpy action embeddings, and learns GPT-style autoregressive dynamics to predict perturbed cellular states in gene-expression space. We evaluate this combined stack on Nadig and Replogle Perturb-seq datasets across within-dataset, fine-tuning, and zero-shot transfer settings, comparing multiple action embeddings against classical baselines with per-perturbation metrics. Together, embpy and the world model provide a reproducible foundation for studying how pretrained biological representations can support generalizable prediction of cellular perturbation responses.