| name | literature-review |
| description | Conduct deep, structured literature research for Ersilia drug discovery work. Use this skill whenever a user asks to find, review, or summarise scientific papers about a disease, biological target, compound class, or AI/ML method in the context of drug discovery — especially for neglected tropical diseases (NTDs), antimicrobial resistance (AMR), or global health. Triggers include target names (e.g. Mtb, PfDHFR, TryR, InhA), disease names (e.g. malaria, leishmaniasis, tuberculosis, Chagas, schistosomiasis), model or method types (e.g. GNN, QSAR, ChemBERTa), and requests like "find papers on", "what does the literature say about", "summarise research on", "literature review for", or "what's been published about". Always use this skill for Ersilia-context literature work, even if the request is phrased casually.
|
Literature Research for Ersilia
Conduct rigorous, comprehensive literature research for Ersilia's drug discovery mission
— focused on neglected tropical diseases (NTDs), antimicrobial resistance (AMR), and
AI/ML-driven drug discovery for global health.
The research question may be a disease, a protein target, a compound class, an ML method,
or a combination. The output is a well-cited synthesis with 20–40 curated papers, a
curation marker on each entry (🤖, 🌍, 🗃️, 💻, ⭐), and a written markdown report.
Inputs
The user will invoke this skill with:
- A research topic (required): disease, target, method, or compound class.
- (optional) A date range, e.g.
--from 2020-01-01 --to 2025-12-31. Default: last 5 years.
Foundational / landmark papers (original target structure, seminal resistance mechanism,
field-defining method) are exempt from this window — include them regardless of age and
mark them as seminal.
- (optional)
--depth focused to cap at 8–15 papers and run faster (default is comprehensive: 20–40 papers).
- (optional)
--out <path> to override the default output file location.
If the topic is ambiguous, ask one focused question before searching. Never invent missing inputs.
Reference files
Read these before starting any search:
references/hub-incorporation-criteria.md — empirical priors for 🤖 (small-molecule-input gate, subtask taxonomy, venue priors). Read before triage.references/lmic-countries.md — World Bank low/lower-middle income countries (the 🌍 rule).references/apis.md — PubMed, Europe PMC, bioRxiv/ChemRxiv endpoint formats. Read before any API call.
Workflow
Step 1 — Parse the query
Identify:
- Primary entity: disease (M. tuberculosis, malaria), target protein (PfDHFR, TryR, InhA),
compound class (AMPs, antimalarials, nitroaromatics), or ML method (GNN, ChemBERTa, REINVENT).
- Research angle: biology/mechanism, drug discovery assays, AI/ML models, clinical findings,
open datasets, ADMET.
- Scope signals: LMIC-first framing? A comparison of methods? A target-specific deep dive?
A biology/mechanism focus vs. a computational-methods focus?
Map the entity to Ersilia disease priorities (see reference files); note Hub subtasks where
relevant, but treat them as a tag, not the organising principle of the review.
Step 2 — Multi-source search
Run searches across all four sources. Use web_search as the primary tool; supplement with
API calls (bash subprocess) where the domains are accessible — see references/apis.md for
exact endpoint formats.
| Source | web_search prefix | Best for |
|---|
| Nature / Nature family | site:nature.com | High-impact biology, chemistry, AI — search first |
| Science / Science family | site:science.org | High-impact biology, chemistry, AI — search first |
| Cell / Cell Press | site:cell.com | High-impact biology, disease mechanisms — search first |
| PubMed / MEDLINE | site:pubmed.ncbi.nlm.nih.gov | Peer-reviewed biology, pharmacology, clinical |
| Europe PMC | site:europepmc.org | Open-access full text, preprints + journals |
| PLOS | site:journals.plos.org | NTDs (PLOS Negl. Trop. Dis.), pathogens, global-health medicine |
| bioRxiv | site:biorxiv.org | AI/ML preprints, computational biology |
| ChemRxiv | site:chemrxiv.org | Chemistry and ADMET preprints |
| arXiv | site:arxiv.org | ML-method preprints (cs.LG) — new architectures not on bioRxiv |
Search Nature, Science, and Cell first — landmark papers here set the framing for the rest of the review.
Minimum search coverage:
Run at least 4 query variants across the sources (~30 searches total). Mix — and keep the
balance, do not let the computational variants crowd out the biology/clinical layer that
anchors a literature review:
- Biology / mechanism:
<disease/target> mechanism resistance pathogenesis review
- Narrow:
<target/disease> <specific subtask> drug discovery
- Broad:
<disease family> AI machine learning compound
- Method-facing:
<method type> ADMET activity prediction <disease>
Run every variant against the general sources (Nature, Science, Cell, PubMed, Europe PMC). Route
the specialized sources to where they pay off: PLOS for variants 1–2 (NTD biology / drug
discovery), arXiv for variants 3–4 (ML methods), bioRxiv/ChemRxiv for the computational
and chemistry variants.
Example queries for PfDHFR:
site:pubmed.ncbi.nlm.nih.gov PfDHFR inhibitor malaria drug discovery
site:europepmc.org PfDHFR antifolate resistance Plasmodium falciparum
site:biorxiv.org DHFR malaria machine learning activity prediction
site:chemrxiv.org dihydrofolate reductase antimalarial QSAR 2023
site:pubmed.ncbi.nlm.nih.gov antifolate Plasmodium drug resistance mechanism
site:europepmc.org malaria drug discovery AI deep learning SMILES
...
After web_search, use web_fetch on individual paper pages to retrieve full abstracts
when snippets are too short. Also use web_fetch on Europe PMC / PubMed landing pages
to get complete metadata (authors, journal, year, DOI).
Target raw pool: 40–80 raw results before screening.
Citation snowballing. Database keyword search alone misses foundational work. After the
initial searches, take the 2–3 strongest reviews or landmark papers in the pool and scan their
reference lists for works that are repeatedly cited but not yet in your pool — fetch the review
page (or its references via Crossref is-referenced-by-count / the reference list) and add the
recurring citations. This backward search is the main way seminal older papers enter the review.
Step 3 — Verify metadata via Crossref
Before composing any entry, verify the first-author surname and publication date via
Crossref:
curl -s "https://api.crossref.org/works/<doi>"
Use message.author[0].family and message.published["date-parts"][0] as canonical
values. When only the title is known:
curl -s "https://api.crossref.org/works?query.title=<title>&filter=container-title:<journal>"
If lookup fails, omit the author rather than guessing. Never fabricate author names or dates.
Canonical URL — once the DOI is confirmed via Crossref, set the entry URL to
https://doi.org/<doi>. This resolves to the publisher's landing page and is the only
reliable permanent link. For preprints without a DOI, use the direct biorxiv/chemrxiv
paper page URL (e.g. https://www.biorxiv.org/content/<id>). Never use search-result
pages, PubMed search URLs, or any URL that does not resolve directly to the paper itself.
Step 4 — Screen and tier
Deduplicate first. The same work often appears multiple times in the raw pool — most
commonly a preprint (bioRxiv/ChemRxiv) and its later journal version. Collapse these to a
single entry, preferring the peer-reviewed published version (and its DOI); keep the preprint
only if no published version exists. Also merge near-identical titles from different sources.
Then apply filters in order:
Filter A — scope. An item must fit at least one of three buckets:
- Antibiotic / antimicrobial / AMR drug discovery — TB, NTD antibacterials, AMP / peptide
antibiotic work, AMR surveillance with an ML hook.
- Global health / LMIC drug discovery / open-science capacity-building — NTDs (malaria,
leishmaniasis, HAT, schistosomiasis, Chagas), Africa/LMIC-led work, public datasets or open
infrastructure releases.
- General-purpose AI methods for drug discovery — featurizers, ADMET and toxicity
predictors, generative chemistry, CPI/docking surrogates, synthesis planning, multi-task
chemistry foundation models, open chemistry datasets, methodology reviews.
- Foundational biology / mechanism / epidemiology of a priority pathogen or target —
target structure and druggability, resistance mechanisms, host–pathogen biology, disease
burden and epidemiology. These anchor the review even without a drug-discovery or ML hook,
provided the pathogen or target is an Ersilia priority.
Borderline cancer / cardiology / diabetes papers do not qualify unless they introduce a
general-purpose tool with clear applicability to Ersilia targets, or concern a priority
pathogen/target under bucket 4.
Filter B — Hub-incorporability tag. Re-read references/hub-incorporation-criteria.md.
Apply the 🤖 marker using its exact rules (small-molecule-input gate, six subtask taxonomy,
open code/weights requirement). This is a useful relevance signal for items that pass scope,
not a curation gate — a paper does not need to be Hub-incorporable to belong in the review.
Tier each survivor:
| Tier | Criteria | Include? |
|---|
| 1 | Directly addresses the query entity with experimental or model validation data (an open model/dataset release is one such case, not a privileged one) | Always |
| 2 | Reviews, methodology papers, or datasets with strong relevance to Ersilia's work | Include if space (after Tier 1) |
| 3 | Tangential, low quality, or duplicate | Skip |
Target final set: 20–40 papers (or 8–15 in --depth focused mode).
Equity check. Apply the LMIC lens last:
- Apply 🌍 per the rule in
references/lmic-countries.md (first or last author at an LMIC
institution).
- Give 🌍 items a ranking bonus when near tie-breaks.
- If a paper addresses LMIC pathogens but has no LMIC authorship, note it under
"Known gaps" rather than promoting it.
Coverage self-check. Before writing, verify the final set isn't lopsided. Ask:
- Are both layers represented — foundational biology/mechanism and drug-discovery/methods?
- Are the major sub-areas of the topic covered, or is one angle over-represented?
- Are the main research groups / labs working on this target or disease present?
- Is anything obviously missing (a known landmark paper, a key method, a recent breakthrough)?
If a gap shows up, run targeted follow-up searches to fill it before proceeding. Note any gap
you cannot fill under "Known gaps" rather than papering over it.
Step 5 — Synthesise the report
Structure the synthesis around these sections (adapt to the specific query):
- Overview — High-level picture, key open questions, why this topic matters for
global health and Ersilia specifically.
- Disease / Target Biology — Mechanism of action, druggability, known resistance
issues. Cite landmark papers.
- Drug Discovery Approaches — Assays, screening campaigns, known scaffolds / lead
series.
- AI/ML Methods — Models applied to this target or disease, key datasets, benchmark
results. Cover the methods landscape broadly; note where a tool is Hub-relevant (🤖) but
don't restrict the section to Hub candidates. Mention any notable open dataset (🗃️) inline
here rather than in a dedicated section.
- Research Gaps — What is missing, understudied, or contradicted in the literature.
Call out LMIC authorship gaps explicitly.
Write in clear scientific prose. Every factual claim carries an inline citation:
(Author et al., Year, PMID/DOI).
Surface consensus vs. disagreement. A good review does not just list findings — it states
where the literature agrees and where it conflicts. Where you find contested mechanisms,
inconsistent activity/potency data, or failed replications, say so explicitly and cite the
competing papers side by side, rather than presenting one side as settled.
Step 6 — Compose each curated entry
For every paper in the final set, compose a one-line entry in this format:
[Author et al., *Venue*, YYYY](url) {ribbon} — **Title.** TL;DR + why it matters for Ersilia.
Curation markers (ribbon, fixed display order ⭐🌍🤖🗃️💻):
| Marker | When to apply |
|---|
| ⭐ | Very-high-impact venue: Nature, Science, Cell, PNAS, NMI, NEJM, Lancet, JACS, Angewandte Chemie, Nature/Science/Cell family |
| 🌍 | First or last author at an LMIC institution (see references/lmic-countries.md) |
| 🤖 | Hub-incorporable model — small-molecule input, one of six subtasks, openly available (see references/hub-incorporation-criteria.md) |
| 🗃️ | Releases a notable open dataset relevant to Ersilia targets (bioactivity/ADMET). Secondary signal — apply when clear, don't hunt for it |
| 💻 | Abstract or paper page explicitly names a public repo URL — do not infer |
Entry rules:
TL;DR — 1–2 sentences, plain language, written fresh. Never paste the abstract verbatim.Why it matters for Ersilia — required, one sentence, specific. Name the Hub subtask,
NTD pipeline, partner institution, or open-science release.Author — verified via Crossref (see Step 3). Omit if lookup fails.- Venue — journal or preprint server, verified.
url — must be https://doi.org/<doi> for published papers, or the direct preprint page
for preprints. Never link to a search result page, a PubMed search URL, or any intermediary
that is not the paper itself. If no verified URL exists, omit the hyperlink entirely rather
than linking to the wrong page.- Within each section, sort by relevance to the query, then by venue tier, then by recency.
- Only apply a marker when the criterion is load-bearing. Absent is better than wrong.
Step 7 — Write and surface output
Markdown file — write to:
/mnt/user-data/outputs/literature_<topic>_<YYYYMMDD>.md
(Use the topic slug and today's date.)
Use the output template:
# Literature Research: [Topic]
*Generated: [Date] | Sources: Nature, Science, Cell, PubMed, Europe PMC, PLOS, bioRxiv, ChemRxiv, arXiv | Papers: N*
---
## Overview
[2–3 paragraphs]
## Disease / Target Biology
[With citations]
## Drug Discovery Approaches
[With citations]
## AI/ML Methods
[With citations — broad methods coverage; tag Hub-relevant tools 🤖; mention 🗃️ datasets inline]
## Research Gaps
[Including LMIC authorship gaps]
---
## Curated Entry List
### Tier 1 — Core papers
- [Author et al., *Venue*, YYYY](url) ⭐🌍🤖 — **Title.** TL;DR. Why it matters for Ersilia.
### Tier 2 — Supporting papers
- [Author et al., *Venue*, YYYY](url) — **Title.** TL;DR. Why it matters for Ersilia.
---
## Known Gaps
- [LMIC authorship gaps, missing experimental data, understudied angles]
## Search Log
| Source | Query | Results retrieved |
|---|---|---|
| Nature | ... | N |
| Science | ... | N |
| Cell | ... | N |
| PubMed | ... | N |
| Europe PMC | ... | N |
| PLOS | ... | N |
| bioRxiv | ... | N |
| ChemRxiv | ... | N |
| arXiv | ... | N |
Call present_files immediately after writing to surface the download link.
In-chat summary — after presenting the file, write a concise in-chat block with:
- A one-paragraph synthesis (the "so what" for Ersilia) — the key findings of the review.
- The 2–3 most important papers (any tier) with one line each on why.
- The most important gap, in one sentence.
- Secondary callouts, only if present: Hub candidates (🤖) and LMIC-led papers (🌍), each as
a short bulleted list.
Ersilia context
Priority diseases / organisms
- Mycobacterium tuberculosis, M. abscessus (TB, AMR)
- Plasmodium falciparum, P. vivax (malaria)
- Leishmania spp. (leishmaniasis)
- Trypanosoma cruzi (Chagas), T. brucei (sleeping sickness / HAT)
- Schistosoma mansoni (schistosomiasis)
- ESKAPE pathogens, GLASS AMR priority list
Hub subtask taxonomy (from references/hub-incorporation-criteria.md)
- Activity prediction (41 % of Ready models — highest prior)
- Featurization (25 %)
- Property calculation / prediction (20 %)
- Similarity search (6 %)
- Generation (5 %)
- Projection (3 %)
Relevant AI/ML model types
- QSAR (RF, XGBoost, SVM)
- Graph neural networks (GNN, MPNN, AttentiveFP)
- Transformers (ChemBERTa, MolBERT, Uni-Mol)
- Generative models (VAE, diffusion, REINVENT)
- ADMET predictors, bioactivity classifiers, docking surrogates
Key databases to mention where relevant
ChEMBL · BindingDB · ZINC · PubChem · Open Targets · MMV/DNDi compound sets
Framing priorities
- Open-source tools and reproducible workflows
- Low-data / few-shot regimes
- Resource-limited setting applicability
- Models available in or compatible with the Ersilia Model Hub
Things to avoid
- Do not include items you cannot link to a primary source.
- Do not paste abstracts verbatim — write fresh TL;DRs.
- Do not invent or guess DOIs, authors, or dates. If uncertain, omit and note why.
- Do not promote work about LMICs without LMIC authorship into the equity section.
- Do not apply 🤖 to models with non-small-molecule primary inputs (protein, RNA, peptide,
image, pocket-tensor). Surface them as context items instead.
- Do not apply 💻 without an explicit public repo URL from the paper.
- Do not pad to hit the 20–40 target — if the week or topic is sparse, say so.
- Do not link to search-result pages, PubMed search URLs, or any URL that is not the paper
itself. Every link must resolve directly to the paper's landing page — use
https://doi.org/<doi>
for published work, or the direct preprint URL for preprints. If no verified direct URL is
available, omit the link rather than using a proxy or search page.
