QuIP: 2-Bit Quantization of Large Language Models With Guarantees.

This work studies post-training parameter quantization in large language models (LLMs). We introduce quantization with incoherence processing (QuIP), a new method based on the insight that quantization benefits from weight and Hessian matrices, i.e.

Online Calibrated and Conformal Prediction Improves Bayesian Optimization.

Accurate uncertainty estimates are important in sequential model-based decision-making tasks such as Bayesian optimization. However, these estimates can be imperfect if the data violates assumptions made by the model (e.g.

Cross-species modeling of plant genomes at single nucleotide resolution using a pre-trained DNA language model.

Interpreting function and fitness effects in diverse plant genomes requires transferable models. Language models (LMs) pre-trained on large-scale biological sequences can learn evolutionary conservation and offer cross-species prediction better than supervised models through fine-tuning limited labeled data. We introduce PlantCaduceus, a plant DNA LM based on the Caduceus and Mamba architectures, pre-trained on a curated dataset of 16 Angiosperm genomes.

A machine-compiled database of genome-wide association studies.

Tens of thousands of genotype-phenotype associations have been discovered to date, yet not all of them are easily accessible to scientists. Here, we describe GWASkb, a machine-compiled knowledge base of genetic associations collected from the scientific literature using automated information extraction algorithms. Our information extraction system helps curators by automatically collecting over 6,000 associations from open-access publications with an estimated recall of 60-80% and with an estimated precision of 78-94% (measured relative to existing manually curated knowledge bases).

A guide to deep learning in healthcare.

Here we present deep-learning techniques for healthcare, centering our discussion on deep learning in computer vision, natural language processing, reinforcement learning, and generalized methods. We describe how these computational techniques can impact a few key areas of medicine and explore how to build end-to-end systems. Our discussion of computer vision focuses largely on medical imaging, and we describe the application of natural language processing to domains such as electronic health record data.

Fast Metagenomic Binning via Hashing and Bayesian Clustering.

We introduce GATTACA, a framework for fast unsupervised binning of metagenomic contigs. Similar to recent approaches, GATTACA clusters contigs based on their coverage profiles across a large cohort of metagenomic samples; however, unlike previous methods that rely on read mapping, GATTACA quickly estimates these profiles from kmer counts stored in a compact index. This approach can result in over an order of magnitude speedup, while matching the accuracy of earlier methods on synthetic and real data benchmarks.

Genome assembly from synthetic long read clouds.

Motivation: Despite rapid progress in sequencing technology, assembling de novo the genomes of new species as well as reconstructing complex metagenomes remains major technological challenges. New synthetic long read (SLR) technologies promise significant advances towards these goals; however, their applicability is limited by high sequencing requirements and the inability of current assembly paradigms to cope with combinations of short and long reads.

Results: Here, we introduce Architect, a new de novo scaffolder aimed at SLR technologies.

Synthetic long-read sequencing reveals intraspecies diversity in the human microbiome.

Identifying bacterial strains in metagenome and microbiome samples using computational analyses of short-read sequences remains a difficult problem. Here, we present an analysis of a human gut microbiome using TruSeq synthetic long reads combined with computational tools for metagenomic long-read assembly, variant calling and haplotyping (Nanoscope and Lens). Our analysis identifies 178 bacterial species, of which 51 were not found using shotgun reads alone.

Probabilistic single-individual haplotyping.

Motivation: Accurate haplotyping-determining from which parent particular portions of the genome are inherited-is still mostly an unresolved problem in genomics. This problem has only recently started to become tractable, thanks to the development of new long read sequencing technologies. Here, we introduce ProbHap, a haplotyping algorithm targeted at such technologies.

Whole-genome haplotyping using long reads and statistical methods.

The rapid growth of sequencing technologies has greatly contributed to our understanding of human genetics. Yet, despite this growth, mainstream technologies have not been fully able to resolve the diploid nature of the human genome. Here we describe statistically aided, long-read haplotyping (SLRH), a rapid, accurate method that uses a statistical algorithm to take advantage of the partially phased information contained in long genomic fragments analyzed by short-read sequencing.