Industries / Health / Healthcare / Medical labs
Microbiology
Finding which bacteria, fungi or viruses are causing an infection, and which drugs will still kill them.
7
Milestones achieved, from 1884 to 2025.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1884
A stain sorts bacteria
Hans Christian Gram, a Danish doctor, worked out a staining method in 1884 that splits bacteria by how their cell walls hold a violet dye after an alcohol rinse. Together with shape, it sorts them into Gram-positive and Gram-negative cocci (round cells) and rods. A 2009 lab manual still calls it the cornerstone of bacterial identification.
Credit
Hans Christian Gram
2013
A laser names the germ
The US Food and Drug Administration (FDA) allowed bioMérieux to sell the VITEK MS, which the FDA called the first mass spectrometer system for automated identification of bacteria and yeasts that cause serious illness. The method, MALDI-TOF, fires a laser at a smear of the organism and matches the pattern of particle masses to a database. In the study the FDA relied on, covering 7,068 microorganisms, it named the group or family correctly 93.6% of the time and the species 87.5% of the time. It can start as soon as growth is visible, usually within 18 to 24 hours, where older methods could take up to five days.
Credit
bioMérieux; US FDA
2016
Software reads culture plates
On 6 October 2016 the FDA granted Clever Culture Systems' APAS Compact a De Novo classification, the route for a new kind of device, and created a new device type for automated reading of colonies on culture plates. The machine photographs urine culture plates after 18 hours and sorts them as negative, review or positive; a trained microbiologist must check every plate it does not call negative. At three sites, against a blinded panel of three microbiologists, it called 99.0% of cultures with growth positive on blood agar and 99.5% on MacConkey agar.
Credit
Clever Culture Systems AG; US FDA
2017
A floor for flu tests
In 2009 the Centers for Disease Control and Prevention (CDC) tested three rapid antigen tests, the kind widely used in doctors' office labs, on 45 samples with the new pandemic H1N1 flu virus. They detected it in 40%, 49% and 69% of samples, and the CDC warned that a negative result does not rule out infection. In January 2017 the FDA set minimum performance for these tests: at least 80% agreement with a molecular test for flu A and B, and yearly testing against the strains circulating that season, with the CDC supplying a strain panel.
LessonA test checked against last year's virus has to be checked again when the virus changes.
Credit
CDC; US FDA
2017
Drug results from the bottle
On 23 February 2017 the FDA granted a De Novo classification to Accelerate Diagnostics' PhenoTest BC kit. It names the germ and tests which antibiotics work directly on a blood culture bottle that has turned positive, without first growing colonies on a plate overnight. Plates are still needed for germs off its list and for mixed infections. Outside labs now measure such systems: in a 2026 study of 83 patient samples at Penn State College of Medicine, a later system, Q-linea's ASTar, agreed with the lab's usual method 94.6% of the time, with 0.5% very major errors (resistant germs called susceptible).
Credit
Accelerate Diagnostics; E. M. Garrett, A. M. Bobenchik and colleagues
2022
Spectra shared with results
Caroline Weis, A. Egli, K. Borgwardt and colleagues published DRIAMS, a free database of more than 300,000 MALDI-TOF spectra linked to more than 750,000 antibiotic test results from four Swiss labs. Their models predicted resistance from the spectrum alone, which an outside team says would give results a day before the usual test. They scored 0.80 for Staphylococcus aureus and 0.74 for E. coli and Klebsiella pneumoniae, on a scale where 0.5 is chance and 1 is perfect. An outside test at the University of Münster found similar scores on local data, but poor results when training and test data came from different places or times, and drops of 0.10 to 0.25 within 18 months.
LessonA model trained on one lab's spectra must be retrained on local data, and retrained regularly.
Credit
C. Weis, A. Egli, K. Borgwardt and colleagues; University of Münster (outside test)
2025
Cutoffs reach the machines
Breakpoints are the cutoffs that decide whether a germ is reported susceptible or resistant to a drug. From 2010 the Clinical and Laboratory Standards Institute (CLSI) revised many of them, partly because patients with "susceptible" germs were failing treatment. Since 2006 the FDA had required its own recognized breakpoints on cleared test machines, and delays in recognizing the new ones left many US labs using cutoffs more than 10 years out of date. A 2016 law made the FDA list its breakpoints on a website, launched in December 2017, and in January 2025 the FDA recognized CLSI's main standard as a whole.
LessonA corrected cutoff protects patients only once it reaches the machines and software labs use.
Credit
CLSI; US FDA; R. M. Humphries and P. J. Simner, who documented the lag
Watch for
Drug results in two hours
The company PhAST says its blood culture panel for Gram-negative bacteria, cleared by the FDA in June 2026, gives results for 15 antibiotics in a mean of 99 minutes for Enterobacterales and Acinetobacter and 141 minutes for Pseudomonas aeruginosa, and calls it the fastest FDA-cleared system of its kind. The study it cites, at four sites, reports 97.3% agreement with the reference method.
Achieved whena lab with no tie to the company publishes time to result and error rates from routine use.
Claimed by
PhAST, September 2026
AI reads the antibiogram
An antibiogram is the table of a germ's results against each drug, which experts read to infer how it resists them. A team in Bogotá built a custom GPT model that named the resistance mechanism in 95.7% of 210 Gram-negative antibiograms, against 91.9% for a version built on Gemini, judged against expert reading of one hospital's past results.
Achieved whenthe model is tested on another center's results, as its authors say it must be.
Claimed by
J. Sprockel Díaz, S. Sotomonte and A. Buitrago, July 2026
Open
01
A dated FDA cutoff table
A machine-readable table of every breakpoint the FDA has recognized, added or set aside since its breakpoint website opened on 13 December 2017, each row dated and linked to the FDA page or Federal Register notice it came from. Add a small checker that reads a lab's exported breakpoint table and flags each cutoff that no longer matches current CLSI, EUCAST (the European committee) or FDA values.
The briefCredit · unclaimed02
An antibiogram reading test
A public test of about 1,000 sets of MICs (the lowest drug concentration that stops a germ growing) from the CDC & FDA Antimicrobial Resistance Isolate Bank, each paired with the resistance genes CDC found by sequencing. The model or rules engine sees only the species and the MICs, and must name the likely mechanism, for example an ESBL enzyme or a KPC, NDM or OXA-48-like carbapenemase. Score hosted AI models, open-weight models and rules software against the genes.
The briefCredit · unclaimed
Open 01 · the brief
A dated FDA cutoff table
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A machine-readable table of every breakpoint the FDA has recognized, added or set aside since its breakpoint website opened on 13 December 2017, each row dated and linked to the FDA page or Federal Register notice it came from. Add a small checker that reads a lab's exported breakpoint table and flags each cutoff that no longer matches current CLSI, EUCAST (the European committee) or FDA values.
Why it is still open
The FDA's breakpoint page, current as of 17 June 2026, lists its exceptions drug by drug in web pages, and no download file was found. Humphries and Simner wrote in 2025 that the date the FDA recognized a breakpoint has to be traced through a "Notice of Updates" page or through Federal Register changes. The open interpretation tools, the AMR package for R and WHONET's AMRIE engine, carry CLSI and EUCAST tables but not the FDA's exceptions.
Built with what exists
the Federal Register API; the FDA breakpoint pages; the AMR package's breakpoint table (EUCAST and CLSI, 2011 to 2026) to compare against; an AI model such as Claude, Gemini or the open-weight Mistral Large 4 (see the Note "A trillion open parameters") to turn pages into rows, with a script that checks every extracted value against the source text. AI models that review long regulatory records are the advance that makes this a small job (see the Note "Seven more flagged").
First step
Build the history for one drug group, such as the carbapenems, by hand from the FDA pages and Federal Register notices, to serve as the answer key for the AI extraction.
Done looks like
every FDA breakpoint change from December 2017 to date in one file, with date and source link. A clinical microbiologist checks 100 random rows and finds at least 99 correct. The checker flags every planted out-of-date cutoff in 20 made-up lab tables built from old CLSI editions, with no false flags. The table is updated within a week of each FDA change for six months.
Open 02 · the brief
An antibiogram reading test
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test of about 1,000 sets of MICs (the lowest drug concentration that stops a germ growing) from the CDC & FDA Antimicrobial Resistance Isolate Bank, each paired with the resistance genes CDC found by sequencing. The model or rules engine sees only the species and the MICs, and must name the likely mechanism, for example an ESBL enzyme or a KPC, NDM or OXA-48-like carbapenemase. Score hosted AI models, open-weight models and rules software against the genes.
Why it is still open
The published test of a model built for this job used 210 antibiograms from one hospital's records, and its authors call for outside validation. The bank's pages give reference MICs and genes for each isolate, and an October 2026 project has gathered them into one table for 1,087 isolates, but to check rapid blood culture panels, not to score how AI or rules software reads results.
Built with what exists
the isolate tables in the crosswalk project above; the AMR package for R and WHONET's AMRIE as rules-based baselines; hosted models such as Claude or Gemini and open-weight models (see the Note "A trillion open parameters"); a model that returns a probability for each fixed answer (see the Note "A thousand questions, one pass").
First step
Take the 95 KPC, 77 NDM and 38 OXA-48-like isolates listed in the crosswalk, hide the gene columns, and score one hosted model and the AMR package on them.
Done looks like
a frozen test set with an answer key from CDC's gene calls. At least five AI models and two rules engines scored on mechanism accuracy and on the worst miss, calling a carbapenemase producer free of one. Two clinical microbiologists agree with the key on at least 95 of 100 sampled items. Data, prompts and scores are public and rerun after each major model release.