estimate_trial_phase_transition_probability
ChatGPTEstimate the probability that a clinical trial advances to its next development phase (for example Phase 2 to Phase 3), using Gosset's trial phase-transition model (PTRS, probability of technical and regulatory success). Use when the user asks for a trial's probability of success, its chance of reaching the next phase, or what drives that estimate. This is one phase transition only, not the probability of eventual approval: a Phase 2 result of 0.44 means a 44% chance of reaching Phase 3. Do not multiply or extend it into an approval probability. Input: either nct_id (the trial is looked up in Gosset) or trial_params describing a trial. Returns: - phase, and transition_probability (the one-step probability). - source and confidence: "benchmark-v3"/"high" when 3 or fewer parameters are given (the base rate for the phase and disease); "sdk-api-v3"/"high" for an nct_id or a nearly complete description; "sdk-api-v3"/"low" for a partial description, with a warning. Adding avg_arm_size, targets, sponsors, has_designations or num_sites raises confidence. - warnings, e.g. a disease or modality name that could not be classified (ignored), or a value clamped to its valid range (avg_arm_size to 1-10000, negative counts to 0). A non-integer highest_phase such as 2.5 is rejected with an error. - param_importance: the signed influence of each input parameter on the estimate. - trial_params_used: the parameters given to the model. On failure it returns {"error": reason} and no score. Calls count toward the account's prediction allowance.
estimate_trial_time_to_completion
ChatGPTEstimate how many days remain until a clinical trial's primary completion, with an 80% prediction interval, using Gosset's trial duration model. Use when the user asks when a trial will finish or read out, or whether it is likely to run past its posted completion date. The estimate is conditioned on how long the trial has already run, and works for trials that are running, not yet started or completed. For a trial missing from the model's data, its public ClinicalTrials.gov record is fetched (about 1 second). Returns trial_status (running, not_started, completed or not_found), elapsed_days, expected_remaining_days, lower_bound_days and upper_bound_days (the 80% interval), expected_completion_date, sponsor_estimated_completion (the date the sponsor posted), actual_completion_date (for completed trials), model_version and warnings. Calls count toward the account's monthly prediction allowance.
fact_check_text
ChatGPTFact-check and edit a text the user provides, following the user's instructions. Runs asynchronously and returns a report_id. Use when the user asks to fact-check, verify or clean up a document or draft they supply. Each round: (1) the text is reviewed against the instructions and for accuracy, internal consistency, unsupported claims, clarity and structure; (2) factual claims are checked against web sources, a claim contradicted by a source is corrected to the sourced value, and a claim that cannot be checked is flagged; (3) the edits are applied. Rounds repeat up to max_rounds or until no further edits are found. Side effects and limits: creates a job stored in the caller's account, which can take many minutes. Resubmitting the same text and instructions from the same account returns the existing job. Returns {report_id, status, poll_with, ...}. Read the result with get_research_report(report_id): the edited text followed by a "## Review change log" listing the edits and any open findings.
find_clinical_trials
ChatGPTSearch Gosset's clinical trial database across ClinicalTrials.gov and international registries with a natural-language query. Use when the user asks for clinical trials by drug, target, modality, sponsor, disease, phase, status, design, dates or eligibility (age, sex, country), or for the details of a trial by NCT or registry-native ID. International searches are not limited to countries outside the US; omit geography restrictions unless requested. Registry coverage depends on available records. Each trial carries registry and trial_id, nullable nct_id, declared or protocol-based nct_xref with its source, secondary_ids and countries when available. Do not identify or deduplicate international records by null nct_id, titles or sponsor protocol codes alone. Use registry plus trial_id; retain cross-reference provenance. phase is available when highest_phase is absent. Other fields include official_title, brief_summary, status, phase, sponsors, start and completion dates, main_drug, targets, modalities, diseases, enrollment {count, type} (type ACTUAL or ESTIMATED; null when there is no ClinicalTrials.gov record), primary_outcomes [{measure, time_frame}] and eligibility criteria. results_view="brief" or "detailed" adds a results object with availability, stored result summary, publication links and clipping indicators. Requires the account's clinical-data permission; otherwise availability is access_restricted. At most 25 selected trial IDs are enriched per call. Protocol brief_summary and registered primary_outcomes are not observed results. Summaries are returned only when all of their source links are verified for this trial; numbers are not independently re-extracted or verified during the query. Use publication links for questions beyond the stored summary. Default searches do not open a UI. Results are cached for up to 7 days; refresh=True bypasses the cache. Example queries: "phase 3 trials for atopic dermatitis", "completed PD-1 inhibitor trials with results", "industry-sponsored breast cancer trials started after 2022", "recruiting lung cancer trials in the United States". Returns a list of trial records.
find_conference_abstracts
ChatGPTFind conference abstracts across meetings and years, as data cards. Use when the user asks which abstracts or presentations exist for a drug, company, disease, target or trial at any conference or across several, e.g. 'all dostarlimab abstracts', 'what was presented on NCT04484623', 'ASCO 2025 breast cancer abstracts', or for a list or table of presentations. The conference is optional. Each card has abstract_id, title, conference, year, edition, date, drug and drugs, developers, latest_phase, trial_ids, nct_ids and trial_links (each with trial_link_source reported_in_abstract or adjudicated), session type, abstract number, impact, annotation_level, url and an abstract excerpt. Returns the full matching total and next_offset for paging. Reads only Gosset's indexed conference export, the rows of the webapp Conferences tab. latest_phase is the drug's furthest phase anywhere, not the study phase. Most abstracts name no trial. Drug matches cover the primary drug and resolved arm interventions, so comparators named only in text are not matched; use search for free text. Names that do not resolve fail with a message and never widen the search. For an interactive view of one meeting, use show_conference_explorer.
find_disease_epidemiology
ChatGPTGet prevalence and incidence for a disease, with prevalent-case counts for the US, EU5, Japan and the seven major markets (7MM). Use when the user asks how many patients have a disease, or for its prevalence or incidence, for example to size a market. Starts from Gosset's disease epidemiology data (Orphanet, CDC PLACES, SEER) and checks it against current literature with web searches. When the database figure covers a broader disease than the one asked about, the estimate is adjusted and the reason is given. effort sets the amount of web research: - "low": one web search, about 5 seconds. - "medium" (default): several targeted searches (US and ex-US prevalence, diagnosis and treatment rates, incidence), about 1 minute. - "high": also starts an asynchronous research report (about 15 minutes) stored in the caller's account. The call returns at once with deep_research {report_id, status}; read the report with get_research_report. Returns {indication, prevalence, implied_prevalent_counts, db_layers, reconciliation}. prevalence holds the point prevalence per 100,000, a US prevalent-case estimate, its basis and sources; db_layers holds the unadjusted database values.
find_drug_list_prices
ChatGPTGet US list prices (wholesale acquisition cost, WAC) by year for a drug and for comparable drugs. Use when the user asks what a drug costs, or wants its list price next to other drugs' prices. Prices are per fill or per package, not per year, and the package size and time basis differ between drugs; convert with each drug's dosing before comparing annual costs. If no comparators are given, a default set of comparable drugs is used. Returns {asset: {name, sticker_price_latest_median, sticker_prices, revenue_forecast}, comps: [{name, latest_median, series}], found}.
find_drug_patents
ChatGPTGet patent and regulatory-exclusivity data for a drug: patents, composition-of-matter status, expiry dates and FDA Orange Book exclusivities. Use when the user asks when a drug's patents or exclusivities expire, or about its intellectual-property position. The data comes from two Gosset sources, annotated patent records and FDA Orange Book records; layer_conflict is true when they disagree. Returns {aliases, annotations, verified, designations, layer_conflict, found}.
find_drugs
ChatGPTSearch Gosset's database of drugs and drug candidates by name, target, disease, developer, modality, development phase or route of administration, using a natural-language query. Use when the user asks which drugs exist for a target, disease or mechanism, which drugs a company is developing, or for facts about a named drug (developer, target, modality, phase, indications, approval status). Behavior and limits: - The search returns broad matches: loosely related drugs can appear, so filter the returned records to what the user asked for. - The query filters only by name, disease, target, developer, modality, phase and route of administration. Other constraints (dates, geography, biomarkers, dosing) are not applied; filter the results for those. - effort="low" (default) queries the database directly and returns in about a second. effort="medium" or "high" adds semantic search, LLM reranking and a web search and takes 30 to 200 seconds; use them only when a low-effort search returns nothing relevant for a drug that should exist, such as a very new compound or an unusual spelling. - Records without an id are drugs found in publications or on the web that are not yet in the database. - Results are cached for up to 7 days; refresh=True bypasses the cache. Example queries: "pembrolizumab", "PD-1 inhibitors", "Merck antibody", "approved breast cancer", "phase 3 oral obesity drugs". Returns {items, count, query}: the matching drug records (names, developers, targets, modalities, diseases, phase), the number returned, and the parsed query. Each record's target_evidence says how well each target is established for THAT drug: {target, status, confidence, served, top_evidence [{title, url, locator, support}]}. status is verified (a reviewed primary source for this exact compound, e.g. a measured IC50 in a named table), limited_support (only weak sources), conflicting (sources disagree), contradicted (a primary source refutes it), unsupported (sources were reviewed but none establishes it) or no_verified_source (taken from annotation only, unreviewed; the target is still listed). served=false marks a reviewed target that is not in targets, e.g. one removed as contradicted, kept so you can say why. Cite top_evidence with its locator when you state a target as established; say a no_verified_source or unsupported target is unconfirmed rather than presenting it as fact. interaction_evidence carries the same evidence_status per mechanistic interaction.
find_pharma_companies
ChatGPTSearch Gosset's database of biotech and pharmaceutical companies by name, category, therapeutic area, modality, target, development stage or location, using a natural-language query. Use when the user asks about a company's focus, pipeline or status, or for companies that match criteria such as "clinical-stage gene therapy companies in Boston". effort="low" (default) searches the company database and returns in about a second. effort="medium" or "high" adds a web search and LLM reranking and takes 30 to 200 seconds; use them only when a low-effort search returns nothing relevant for a company that should exist. Example queries: "Genentech", "Boston biotech companies", "PD-1 antibody developers", "clinical stage gene therapy". Returns {items, count, query}: the matching company records (therapeutic focus, pipeline, financial status), the number returned, and the parsed query.
find_pharma_deals
ChatGPTSearch Gosset's database of biopharma transactions: licensing agreements, acquisitions and mergers, collaborations and similar deals. Use when the user asks about deals involving a company, drug, target, modality or therapeutic area, or for precedent transactions and their terms (upfront payment, total value, date). effort="medium" (default) combines the database search with a web search and LLM ranking to match deals to drugs and diseases. effort="low" searches the database only; it is faster and can miss deals for narrow queries. effort="high" runs a wider search and takes longer. Results are cached for up to 7 days; refresh=True bypasses the cache. Example queries: "Merck licensing deals", "PD-1 acquisitions over 1 billion", "Roche acquisitions 2023", "anti-TL1A antibody deals". Returns {items, count, query}: the matching deals (parties, deal type, date, values, assets, source url), the number returned, and the parsed query. Each deal's source_verified says whether the linked source itself reports the deal: true, false (the link does not report it; cite it as unverified) or null (not checked).
find_pharma_events
ChatGPTSearch Gosset's timeline of dated pharma and biotech events, such as approvals, FDA designations, discontinuations, deals, financings and trial readouts. Event types: drug approvals, FDA designations (Breakthrough Therapy, Fast Track, Orphan Drug, Priority Review, RMAT), program discontinuations, licensing and acquisition deals, company financings (venture rounds, IPOs), clinical trial milestones and result readouts, and news. Use when the user asks what happened over a period for a drug, company, target, modality or disease, or for events of one type, such as Breakthrough Therapy designations granted this year. The query can name entities, an event type and a date window ("last week", "in 2024", "between 2025-01-01 and 2025-06-30"). Results are cached for up to 7 days; refresh=True bypasses the cache. Each event has event_type, date, title, summary, url, importance and the linked drugs, companies, diseases, targets and modalities, plus type-specific fields such as designation_type or amount_usd. Returns a list of events, newest first.
find_pharma_news
ChatGPTSearch Gosset's stored feed of pharma and biotech news articles and company press releases. Use when the user asks for recent news or press releases about a company, drug, disease or target, or for news of one kind, such as FDA decisions or acquisitions. Only articles already stored in Gosset are searched; the web is not searched at request time. Results are cached for up to 7 days; refresh=True bypasses the cache. Each article has title, date, source, url, companies, drugs, diseases, targets, categories (for example M&A, FDA, Clinical, Conference) and article_type. Example queries: "Merck press releases about pembrolizumab", "FDA approvals 2024", "atopic dermatitis news". Returns a list of articles, newest first.
find_similar_trials
ChatGPTFind completed historical trials that resemble a given trial, measured on the features of Gosset's trial phase-transition model, with a similarity score and the model features of each trial. Use when the user asks which past trials are comparable to a trial, or wants the historical evidence behind a phase-transition probability from estimate_trial_phase_transition_probability. Give exactly one of nct_id or trial_params. Optional filters: disease_class_ids, phase, and limit (at most 100). Only trials completed on or before the given trial's cutoff year are candidates, so trials that have not started can also be used. Similarity = 1 / (1 + weighted Euclidean distance), weighted by the model's feature importances. Returns {trials: [{nct_id, similarity, distance, trial_data, feature_values}], feature_weights, candidate_count, cutoff_year, model_path}.
find_trial_results
ChatGPTGet the numeric results of clinical trials as a long table: one row per trial, arm, endpoint and week, with the value, its unit, dispersion, number analysed, the difference versus the comparator arm, and the source. Use when the user asks for efficacy or safety numbers reported by trials, such as body-weight change by drug and week, responder rates, or adverse-event rates per arm. Use find_clinical_trials to find trials by their design or status instead. Coverage today is obesity: the results posted on ClinicalTrials.gov for 538 trials, plus verified values from the trials' publications and conference reports. By default each trial, arm, endpoint and week has one row: the verified publication (or conference) value when there is one, otherwise the ClinicalTrials.gov posted primary analysis. source_type says which ("publication", "conference" or "ctgov_posted"). Posted values are often observed means, which can differ from the estimand a publication reports (STEP 1: -15.6% posted, -14.9% published; the default row is the published -14.9%). estimand is set only when the source names one; rows that name none carry estimand_status "unknown" and flags ["estimand_unknown"]. When verified sources disagree on one arm's value, each reading is returned with conflicting=true and conflicting_sources listing every source. all_analyses=true returns every analysis from every source. comparator_difference is test minus comparator, from the analysis reported for the same arm; where a publication value replaced the ClinicalTrials.gov posting, the posting's own between-arm analysis is in posted_comparison (source_type "ctgov_posted"). Every row has source_url and a locator to the exact place in the source. Examples: - % body-weight change by drug by week for semaglutide vs tirzepatide: endpoint_key="bw_pct_change", drug=["semaglutide", "tirzepatide"] - 72-week weight loss in SURMOUNT-1: nct_id="NCT04184622", endpoint_key="bw_pct_change", min_week=72, max_week=72 - share of patients losing at least 15% of body weight on phase 3 trials: endpoint_key="bw_responders_15pct", phase=3 - nausea rates for liraglutide: drug="liraglutide", endpoint_key="ae_nausea" Returns {items: [{nct_id, trial, phase, arm, drug, dose, endpoint_key, endpoint, week, value, unit, dispersion, n, estimand, estimand_status, analysis_label, primary, flags, conflicting, conflicting_sources, comparator_arm, comparator_difference, comparator_difference_ci, comparator_p_value, posted_comparison, source_type, verification_status, source_url, locator, ...}], total_matched, returned, truncated, filters, notes}.
get_drug_evidence
ChatGPTRead one drug's attributed description, mechanism and preclinical evidence. Use when the user selects an asset from a landscape or asks for the underlying evidence for a returned drug id. Requires the connected account's drug-search permission. Reads the current database record and its primary-source links. It does not infer clinical results or fill missing fields. The landscape's disease-scoped phase remains authoritative for that landscape. Returns the asset identity, available mechanism and preclinical fields, and source attribution.
get_drug_landscape
ChatGPTRetrieve a drug landscape with an interactive company map and asset table. Use when the user asks to browse a landscape by target, disease, modality or company, filter its assets, or open a visual pipeline. For a plain list, count, table or data export without an interactive view, use find_drugs instead. Make one successful call for the requested scope. Keep the default limit unless the user requests a specific asset count. The returned Gosset panel already provides the interactive landscape: use it and give a brief summary, unless the user requests a separate format. Do not repeat a successful search with a smaller limit merely to summarize it or because the chat preview is shortened. A new user-requested scope/count or an actual tool error can justify another call. Do not create a duplicate explorer or invent placeholder assets for records absent from the model's preview. Uses the database-only find_drugs search and the connected account's search permissions and limits. Assets are ordered by descending development phase before limits are applied: Approved, Phase 4, Phase 3, Phase 2, Phase 1, Phase 0, Preclinical, Unknown, Discontinued. Counts describe the returned rows; reaching a limit does not establish complete market coverage. Disease searches retain and sort by the phase in that disease. Returns structured assets, phase counts, parsed filters and any truncation, with an optional interactive view. Tools remain usable in clients without UI.
get_research_report
ChatGPTGet the status of one research report and, once it is finished, its output. Use when the user wants to check on or read a report started with submit_research_report, given its report_id (returned by submit_research_report or list_research_reports). Only reports submitted by the caller's own account can be read. Status is one of pending, processing, qc (waiting for human review), done, error, paused or quota_exceeded. While the status is pending or processing, check again later; reports usually take 10 to 45 minutes. Returns {report_id, status, final_output, created_at, error}. final_output is markdown with links to the report files and is empty until the status is done.
list_research_reports
ChatGPTList the research reports the caller's account has submitted, newest first. Use when the user asks which reports they have, or needs the report_id of an earlier report to read it with get_research_report. Only the caller's own reports are returned. Optional filters: status, and a text match on the prompt. Returns {reports: [{report_id, prompt, status, agent_type, created_at, completed_at}], count}.
show_conference_explorer
ChatGPTExplore conference presentations, abstracts and reported findings for a selected research space. Use when the user requests conference findings, such as 'show me new obesity results from EASD': pass conference='EASD', disease='obesity' and omit year. Retrieves the latest indexed edition of that exact meeting through the current year. Pass only user-requested filters, carrying forward an explicit research scope for follow-ups. No preceding landscape or results call is needed. Includes drug and non-drug research, clinical, preclinical, observational and mechanistic abstracts, authors, affiliations, sessions, oral/poster formats, reported numeric findings where stored, and links to abstracts, PDFs, posters and recordings where available. Does not restrict to trials or require measured endpoints. Asset stage is not presentation study phase. New means recent indexed presentations, not verified first disclosure. Returns compact cards, the full matching count and whole-cohort search/filter controls; full abstracts load when selected. Use presentation_id with the same scope to inspect an original record before extracting endpoints. Do not build comparisons from card excerpts or treat missing stored numbers as missing results. Drug-role and company-role filters distinguish a main asset from comparator mentions and an asset developer from author affiliations. Ambiguous ontology names fail closed with actionable choices. Reads only Gosset's indexed conference export. Does not crawl sources, extract missing endpoints, infer trial matches, normalize incompatible findings or certify stored summaries. Returns a conference widget, source records and selection/limit coverage. Keep accompanying prose brief; do not recreate its table. For a deeper quantitative comparison, inspect the original abstract and use show_result_explorer with source-bound measurements.
show_deal_explorer
ChatGPTDisplay supplied pharmaceutical transactions and terms in an interactive explorer. Use when the user requests transactions, ownership or disclosed financial terms. First retrieve records with find_pharma_deals. Supply data.deals with parties, assets, dates, types, original terms/currencies, rights and sources. Keep upfront, financing, paid or potential milestones, acquisition and reported values distinct. Portfolio amounts remain portfolio amounts. Returns supplied records and a deals UI resource. Does not search or verify transactions. Plain deal lists do not require this tool.
show_drug_landscape
ChatGPTShow a drug landscape with an interactive company map and asset table. Use when the user asks to browse a landscape by target, disease, modality or company, filter its assets, or open a visual pipeline. For a plain list, count, table or data export without an interactive view, use find_drugs instead. Make one successful call for the requested scope. Keep the default limit unless the user requests a specific asset count. The returned Gosset panel already provides the interactive landscape: use it and give a brief summary, unless the user requests a separate format. Do not repeat a successful search with a smaller limit merely to summarize it or because the chat preview is shortened. A new user-requested scope/count or an actual tool error can justify another call. Do not create a duplicate explorer or invent placeholder assets for records absent from the model's preview. Use a short entity query, retaining only user-requested constraints. For TL1A, use "TL1A drugs", without adding clinical/preclinical, inhibitor or antibody filters. Unknown-phase and discontinued records remain eligible unless explicitly excluded. Uses the database-only find_drugs search and the connected account's search permissions and limits. Assets are ordered by descending development phase before limits are applied: Approved, Phase 4, Phase 3, Phase 2, Phase 1, Phase 0, Preclinical, Unknown, Discontinued. Counts describe the returned rows; reaching a limit does not establish complete market coverage. Disease searches retain and sort by the phase in that disease. Returns structured assets, phase counts, parsed filters and any truncation, with an optional interactive view. Tools remain usable in clients without UI.
show_result_explorer
ChatGPTDisplay supplied measured clinical or preclinical results in an interactive explorer. Use when the user requests efficacy, safety or preclinical comparisons. Retrieve summaries or inspect published results first; prepare data.clinical and data.preclinical according to the research-explorer evidence contract. Retain dose, endpoint, numeric value, original units, timepoint, analysis, source URL and locator. Protocol outcomes are not reported findings. Returns supplied records and a results UI resource. Does not fetch publications, extract results or independently verify evidence.
show_trial_explorer
ChatGPTDisplay supplied clinical study records in an interactive trial explorer. Use when the user requests trial design, recruitment, dates or international registrations. First retrieve records with find_clinical_trials. Supply data.trials with native registry IDs, dates, phases, assets, design and source links. Returns the supplied records and a trial UI resource. Does not search, forecast dates, infer registry matches or verify findings. Plain trial lists do not require this tool.
submit_feedback
ChatGPTSend user-approved feedback to Gosset, store it and notify its team by email. Use when the user wants to report incorrect or missing data, a problem with the interface, or a feature suggestion to Gosset. Before calling, show the exact feedback and any context, then ask: "May I send this feedback, the context shown above and your Gosset account email to Gosset? It will be stored by Gosset and emailed to lukasz@gosset.ai." Wait for the user's affirmative answer. Complaining about a result alone is not consent. Set consent=true only after that answer; if they decline, do not call. With consent=false the tool returns consent_required and stores/sends nothing. Only include the feedback and context the user agreed to share. Do not automatically include conversation history, private files, credentials or other tool results. The sender's email comes from their authenticated Gosset account; the recipient is fixed to lukasz@gosset.ai. Requires sign-in. Allows up to 20 new feedback submissions per 24 hours, separately from search/report allowances. Reuse request_id for retries. Email failures leave the feedback saved and are reported; an uncertain email attempt is not automatically repeated. Returns status (consent_required or saved), feedback_id after saving, notification_status and a message. An accepted notification means the email provider accepted it, not that inbox delivery was verified.
submit_research_report
ChatGPTStart a long-form research report on a pharmaceutical or biotech question. Runs asynchronously and returns a report_id. Use when the user asks for a written report that draws on many sources: an evaluation of one named drug, the standard of care for a disease, a competitive landscape for a target, modality or indication, a head-to-head comparison of drugs in an indication, a disease brief, a deal analysis, or an open research question. Report types, chosen from the prompt unless agent_type is set: investment-assessment, standard-of-care, competitive-landscape, head-to-head, disease-brief, deal-analysis, deep-research. Side effects and limits: - Creates a report job stored in the caller's Gosset account and counts toward the account's weekly report quota. - Submitting the same prompt again from the same account returns the existing report (with "deduped": true) and does not start a new one. - A report takes about 10 to 45 minutes. The report agents search the public web while writing. Returns {report_id, status, message}. Read the result with get_research_report(report_id); list earlier reports with list_research_reports.