Read what matters.
Tailored for you.

Duit scans thousands of papers from your subscribed journals and surfaces the ones matching your research.

300+
Journal Feeds
1000+
New Papers Daily
Currently
In Beta
Researchers from
Cornell University

Just du it.

(讀, dú — to read, to study)

Features

Built for researchers

AI Relevance Scoring

Each paper gets a 0–10 relevance score based on your research profile, keywords, and reading history.

300+ Journal Feeds

Subscribe to RSS feeds from Nature, Science, Cell, bioRxiv, and hundreds more. Add any feed you want.

Social Feed

See what researchers you follow are saving and recommending. Papers ranked by relevance, recency, and signal from your network — every pick tied to a face.

Reading Companion

Turn any paper into a structured read — a tl;dr, key findings, section summaries, and AI-explained figures. Ask it to focus on your question.

Email Digests

A personalized digest of new papers ranked by relevance, on your schedule — daily, twice weekly, or weekly — filtered to the score threshold you set.

Podcast

Turn selected papers into a two-host audio podcast with a readable transcript. Set the depth per paper and listen on the go.

Zotero Integration

Connect your Zotero library to automatically build your research profile. Export papers back with one click.

Explore

Map any paper's neighborhood — references, citing work, and related papers — as an interactive citation graph from OpenAlex and Semantic Scholar.

Live demo
Watch your feed rank itself
Pick a research field and every paper re-scores and re-orders on the spot — the way Duit ranks your feed from your profile. Try it right here.
app.duitapp.ai/papers Live
I mostly work on…
Sample of 10 papers · scores illustrative. The real Duit scores every paper across 300+ feeds from your reading history.
Paper Details
Everything at a Glance
Click any paper to see paper details, AI relevance score, journal info, and quick actions.

Guided Directed Evolution of Enzymes with a Protein Language Model

Published: Jun 11 Journal: Briefings in Bioinformatics DOI: 10.1093/bib/bbag279

Relevancy Score: 9.0

(Introduces a deep-learning framework for targeted protein evolution — matches your methodology and keywords.)

Abstract

Directed evolution has transformed protein engineering, yet exhaustive screening of sequence space remains costly and slow. Here we introduce EvoPLM, a deep-learning framework that couples a protein language model with an active-learning loop to prioritize high-fitness variants for targeted evolution. Trained on millions of unlabeled sequences, the model captures evolutionary constraints and predicts the functional effect of mutations without task-specific labels. Across three enzyme families, EvoPLM-guided campaigns reached target activity in roughly half the experimental rounds required by conventional saturation mutagenesis. Our results show that language-model priors can substantially reduce the wet-lab burden of engineering novel biocatalysts.

Social Feed
Discover Through Your Network
See what researchers you follow are saving and recommending, ranked by relevance and what's trending in your field.
Social Feed — ranked by relevance, recency & your network
Trending on Bluesky
AI & ML models researchers are talking about this week
GLM-5.2
text-generation
121 1.5k 11.9k
Rio-3.5-Open-397B
image-text-to-text
88 973 6.2k
8.7
Gut commensal short-chain fatty acids tune regulatory T-cell set point during neonatal colonization
Cell A researcher you follow · 1d ago

Colonization by commensal Bacteroides species raised luminal butyrate and propionate, biasing naive CD4+ T cells toward a regulatory fate and lowering inflammatory tone — a metabolite-level lever on the neonatal immune set point.

8.1
A deep generative model designs thermostable protein binders from backbone geometry alone
Nature A researcher you follow · 3d ago

Conditioning only on target backbone coordinates, the model proposed minibinders that folded as designed in 41% of tested sequences, several retaining nanomolar affinity after a heat challenge at 75 °C.

Explore
Map a Paper's Neighborhood
Turn any paper into an interactive citation graph — its references, the work citing it, and topically related papers. Drag a node, filter, or switch to the list.
Accurate predictions of disordered protein ensembles with STARLING1 seed · 25 related
Seed Similar References Cited by Topics Click a node · drag to move
Reading Companion
Read Any Paper in Minutes
Drop a PDF and get a structured read — tl;dr, key findings, section summaries, and AI-explained figures, focused on your question.

Metagenome-informed metaproteomics of the human gut microbiome, host, and dietary exposome

59 pages
TL;DR

Pairing shotgun metagenomics with high-resolution metaproteomics, the authors measure which microbial, host, and dietary proteins are actually expressed in the human gut — not just which genes are present. Across a healthy adult cohort, expressed function frequently diverged from genomic potential, and dietary proteins were traceable well enough to define a measurable "dietary exposome."

  • Metaproteomic activity correlated only weakly with metagenomic abundance for many taxa — gene presence poorly predicted expressed function.
  • A core set of ~2,100 microbial protein groups was expressed in most individuals despite wide taxonomic variation.
  • Dietary proteins from plant and animal sources were detectable and source-traceable, defining a quantifiable dietary exposome.
  • Host-derived proteins (mucosal and immune) co-varied with specific bacterial families.
  • Short-chain fatty acid biosynthesis was among the most consistently expressed microbial functions.

Background

Metagenomic surveys have mapped the genes present in the gut microbiome, but genomic potential is not the same as biological activity. Which proteins a community actually expresses determines how it metabolizes the diet and interacts with the host.

Metaproteomics closes this gap by measuring expressed proteins directly. Here the authors combine per-sample metagenomes with mass-spectrometry proteomics to resolve microbial, host, and dietary contributions within a single stool sample.

Study design & workflow

Stool samples from a cohort of healthy adults underwent shotgun metagenomic sequencing to build a sample-specific protein sequence database. Peptides were then measured by LC-MS/MS and matched against that database, with the false-discovery rate controlled at 1%.

Spectra were assigned a microbial, host, or dietary origin, and protein groups were quantified by spectral counts to compare expressed abundance against metagenomic gene abundance.

Expressed function vs genomic potential

Over 2,100 microbial protein groups were reproducibly detected, alongside several hundred host proteins and a smaller but consistent set of dietary proteins. For many taxa, protein-level activity ranked differently from metagenomic abundance.

Carbohydrate metabolism and short-chain fatty acid pathways dominated the expressed microbial proteome, while host signatures were enriched for mucosal defense and digestive enzymes.

Implications

Because expressed function diverges from gene content, composition-only microbiome studies may miss the activity that matters for host health. Metaproteomics offers a direct functional readout that integrates microbe, host, and diet in one assay.

The traceable dietary exposome also points toward using stool proteomics to reconstruct recent dietary intake objectively, without relying on self-report.

Figure 3. Expressed protein groups by source across the cohort.

AI-explained

Figure 7. Metaproteomic activity vs metagenomic abundance per taxon.

AI-explained
Table 1. Detected protein groups by source.
SourceProtein groups% of spectra
Microbial2,14861%
Host64229%
Dietary12110%
Podcast
Your Papers, As a Podcast
Turn selected papers into a two-host audio podcast with a readable transcript. Set the depth per paper and listen on the go.
Podcasts
ProMaya: a hierarchical universal deep-learning framework; How AI is reengineering protein engineering
3 papers · 8:03 · Apr 10
  • ProMaya: a hierarchical universal deep-learning framework for protein fitness
  • How AI is reengineering protein engineering
  • Benchmarking generative models for enzyme design
0:00 / 8:03
Host AToday we're breaking down ProMaya — a hierarchical deep-learning framework for predicting protein fitness.
Host BThe neat trick is sharing signal across related proteins instead of training each one from scratch.
Host APaper two zooms out: how AI is reengineering protein engineering end to end.
Host BIt reframes design as search — models propose sequences, and the lab validates the promising ones.
Host AThen the benchmark paper asks whether generative models really beat classical enzyme design.
Host BSometimes they do — but data quality still decides who wins.
Host AA good reminder: the model is only as sharp as the assay behind it.
Single-cell atlases at scale; Foundation models for genomics
2 papers · 5:41 · Apr 8

Start reading smarter

Join researchers who use Duit to stay on top of their field without the noise.

Currently in internal testing — join the waitlist to get early access.

Join Waitlist Open Duit
Private beta

Get early access

Tell us where to reach you and we'll let you in as spots open up.

We'll only use this to invite you to the beta. No spam, no sharing — ever.