Four Ways Through Collagen, and the Experiment Nobody Has Run

This is the mechanics of getting through tissue, and it opens in a skin chamber in 1990, where human and canine hookworm larvae crossed epidermis, basement membrane and dermis, and where the enzymes they used turned out to cut one of three connective-tissue substrates and not the other two. From there the two materials anything moving through a body meets - a fibrillar type I collagen scaffold and the thin type IV sheet under every epithelium, crossing which is the formal definition of an invasive cancer - and the four ways through, unequally strong: cutting, with the one tethered enzyme a gene-knockout series left standing; squeezing, with the measured limit at about ten percent of a nucleus's cross-section; pushing, filmed in a worm; being towed by a fibroblast, which no parasite is; plus a fifth thing that is not a route at all, eating the barrier for its proline. Then the architecture - aligned fibers, stiffness measured in patients, and the collagen signatures that carry hazard ratios - and the four facts the field learned the hard way: three failed phase III programs, a retracted Nature paper, a failed antibody, and the mouse in which deleting stromal collagen made pancreatic cancer worse. Then the one normal tissue licensed to invade, the placenta, which does it with the host's own enzymes, under a host brake, and stops; the parasite enzymology with its catalytic classes corrected on air; what is genuinely shared, which is three things; the arithmetic the verdict answers to; and the experiment somebody could fund, with the contamination failure mode that makes its controls the design.

Show notes

Start in 1990, in a chamber with a piece of skin in it. A modified in vitro skin chamber — in vitro meaning in glassware, outside any living animal — and two hookworm larvae put to it: Ancylostoma duodenale, one of the two hookworms of people, and Ancylostoma caninum, the hookworm of dogs that also gets into people. Both penetrated epidermis, basement membrane and dermis in similar ways: the outer cell layer of the skin, then the thin dense sheet of type IV collagen and laminin that sits under every epithelium, then the thick collagen-rich layer beneath (Hotez and colleagues 1990). Then the useful question, which is not how the worm moved but what it was cutting with.

The larvae carried protease activity — a protease being any enzyme that cuts protein — and o-phenanthroline inhibited it, which makes it metalloprotease activity: an enzyme cutting with a metal ion at its center, zinc in this family, where o-phenanthroline is a chelator that grips the metal, pulls it out and leaves the enzyme inert. Living larvae, not ground-up extracts, released it (Hotez and colleagues 1990); the alkaline pH optimum and the substrate-gel band masses sit on the hookworms' own Atlas page. The authors proposed two jobs for that enzyme, invasion and ecdysis — a larva shedding the cuticle it has outgrown, which a skin-penetrating hookworm does on its way in — and both are in the paper's own title. The first hint that a parasite's protease is not a dedicated invasion tool.

This part's thesis in one result: motile larvae were incubated with purified, radiolabeled connective-tissue molecules, one at a time. Both Ancylostoma species degraded human fibronectin — one of the glue proteins that fasten cells to the matrix — down to a 60,000-dalton intermediate. Neither could degrade solubilized bovine elastin, the rubbery protein that lets tissue recoil, or human laminin, the other glue protein and a main component of that thin sheet. Strongyloides stercoralis, the obligate skin-penetrating nematode with no route into a person except through skin, degraded all three. The authors wrote that this biochemical difference may explain some observed differences in invasiveness (Hotez and colleagues 1990). That worm has its own Atlas page.

The matrix a parasite can cut predicts how far it gets — not how much enzyme it carries, not how aggressive it looks on a slide, but which substrates its enzyme accepts and which it refuses. That is the opposite of the story the cancer literature told itself for twenty years: before 1990 a widely held belief held the matrix metalloproteinases made by cancer cells to be of critical importance to invasion and metastasis, and on it the industry built orally active inhibitors meant to block them across the board (Zucker and colleagues 2000); the trials come later in this part. Specificity, not breadth.

The tumor side, in one sentence. The enzyme that lets a cell cross a collagen barrier is not poured out at large; it is tethered in the cell's own membrane — MT1-MMP, also called MMP-14, a zinc-dependent matrix-cutting enzyme kept on a short leash at the surface — and the gene-knockout series that found it by elimination also found that fibroblasts, the connective-tissue cells that make and maintain collagen, and cancer cells express an indistinguishable collagen-cutting activity in that thin shell against the cell (Sabeh and colleagues 2004). A tumor cell leaving a breast duct and a worm larva crossing the dermis both get through the same wall with an enzyme held at their own surface. One mechanism, or two that look alike?

Before anyone answers it, the worm side has to be kept honest, because a parasite's invasion enzyme can be very unlike itself. The blood fluke Schistosoma mansoni gets through human skin with cercarial elastase, named for the elastin it degrades, carried by the cercaria, the free-swimming larva that leaves a snail and finds a person in water. It is not one protein: genome analysis shows a greatly expanded family of cercarial elastase gene isoforms — slightly different versions of the same enzyme from duplicated copies of one gene — an expansion unique to S. mansoni, with modeling predicting differences in what each copy accepts and the enzyme switched on before the parasite leaves its snail (Ingram and colleagues 2012). One worm, one job, a whole family of slightly different tools, switched on early; the schistosoma Atlas page carries the life cycle.

And nobody should carry invasion equals enzymes into the rest of this part either, because well-studied parasites get through tissue cutting nothing at all. Toxoplasma gondii glides: the force comes from the glideosome, a motor complex in which myosin A walks along short actin filaments and drags surface adhesins, the parasite's grip proteins, backward, so the parasite goes forward (Farhab and Yuan 2025). A review, not a new experiment, on a single-celled parasite in culture; Toxoplasma has its own Atlas page.

Then shape, which is stranger. Plasmodium berghei sporozoites — the stage a mosquito injects into skin, the rodent malaria species — released into arrays of silicone pillars built as blood-vessel surrogates, preferentially migrated around pillars whose curvature matched their own crescent shape, which the authors call structural tropism and read as how a sporozoite finds a capillary (Muthinja and colleagues 2017). A rodent malaria parasite in a microfabricated device, navigating by shape rather than chemistry. Neither of those two cuts a path: one walks on its adhesins, the other feels for a curve. The malaria parasite has its own Atlas page.

So here is the question for this part. Do a cancer cell and a parasite larva get through tissue the same way? Two refusals go up front, so nobody wonders which way my thumb is on the scale. This part will not say that parasite and tumor matrix-cutting enzymes are the same enzymes; they are not, the catalytic chemistry is where that claim fails, and each class gets named as we come to it. And it will not say that softness is a property parasites and tumor cells share, because on one of those sides the measurement has never been made: no deformability figure, no elastic modulus, for any parasitic helminth larva.

Four Ways Through Collagen, and Four Facts the Field Learned the Hard Way

This part turns on two materials anything moving through a body meets: a fibrillar scaffold of type I collagen and the basement membrane, the thin type IV sheet under every epithelium (Hohenester and Yurchenco 2013). Crossing that sheet is the formal definition of an invasive cancer. Collagen is a family of twenty-eight proteins (Ricard-Blum 2011); that chemistry sits on the Atlas page for the collagen types. Four ways through, unequally strong, plus a fifth that is not one at all.

Cutting has the cleanest mechanism. Fibroblasts from gene-targeted mice, each with one gene switched off, found the enzyme by elimination: not plasminogen, not the MMP-2 and TIMP-2 axis, TIMP-2 being the body's own brake on these enzymes, not MMP-9, MMP-13, MMP-8 or MMP-3. What was left was MT1-MMP, tethered rather than released, cutting only in the thin shell against the cell's surface. Deleting or suppressing it abolished collagenolysis and invasion in vitro and in vivo, in a living animal, and the same paper found fibroblasts and cancer cells proteolytically indistinguishable (Sabeh and colleagues 2004). That is what moved the field off MMP-2 and MMP-9, the two enzymes every older review names; the MT1-MMP Atlas page carries the exclusion series.

Squeezing came from outside cancer. Human T cells crossed three-dimensional collagen with no detectable collagenolysis, unaffected by a protease inhibitor cocktail, deforming through gaps that already existed (Wolf and colleagues 2003). Ten years later the same group found its limit: protease-independent migration fell linearly with pore size and stopped when the nucleus was squeezed to about ten percent of its cross-section — seven square micrometers for tumor cells, four for T cells, two for neutrophils (Wolf and colleagues 2013).

The two combine into digest-on-demand. Change matrix pore size, or lamin A expression, which sets nuclear stiffness, and you change how much MT1-MMP collagenolysis a cancer cell performs (Infante and colleagues 2018). The cell cuts only as much as its nucleus cannot get through. No parasite equivalent has been described.

Name the structure that does the cutting, because a fabricated paper circulates under its name. An invadopodium is a short protrusion a cell drives into the matrix with its cutting enzymes concentrated at the tip. The 2006 paper usually cited for invadopodia degrading three-dimensional collagen does not exist: that title returns nothing, the first author is Vira Artym rather than Anna, and the real 2006 paper is in Cancer Research, on a gelatin matrix, finding that cortactin organizes the structure while MT1-MMP does the degrading, so blocking the enzyme left invadopodia that formed and could not cut (Artym and colleagues 2006). The fibrillar-collagen result came nine years later: dense fibrillar collagen is itself a potent inducer of invadopodia, through integrin alpha-2-beta-1 — an integrin being one of the clamps a cell grips matrix with — and kindlin-2, and it induced them in primary human fibroblasts as well as carcinoma lines (Artym and colleagues 2015). That kills a slogan: invadopodia are neither the moment a tumor turns lethal nor the property of aggressive cancers alone. One more correction: contact guidance in three dimensions is Kim and colleagues 2021, not Park 2024 — breast cancer cells polarized more strongly as fiber directional coherence rose, with adhesion and contractility shifting cells between rounded and elongated shapes rather than remodeling the matrix.

Pushing is third, and the clearest picture is in a worm: live imaging of anchor-cell invasion in Caenorhabditis elegans showed actin-based invadopodia breaching the basement membrane first, then the invasive protrusion physically displacing the sheet rather than relying on proteolysis alone to widen the opening (Hagedorn and colleagues 2013). Being towed is fourth: cancer-associated fibroblasts pull on cancer cells through a mismatched junction, N-cadherin against E-cadherin — the cadherins being how the cells of a sheet hold each other — reinforced under load, and impairing it blocked collective invasion (Labernadie and colleagues 2017). Mechanical, no enzyme, and no parasite counterpart, because no parasite is dragged through tissue by a host cell holding onto it.

Then the fifth thing. Collagen is a proline reservoir, and pancreatic cancer cells take up collagen fragments, survive nutrient limitation on them, and need proline oxidase, PRODH1, to proliferate in vitro and in vivo (Olivares and colleagues 2017). The barrier the tumor cuts is also the meal it eats — with its caveat: the in vivo proof that tumors drink and digest surrounding protein tracks albumin, not collagen (Davidson and colleagues 2017), and nobody has traced labeled collagen proline into tumor biomass in a living animal.

Now the architecture. Multiphoton microscopy with second harmonic generation, a laser method that lights up collagen with no stain, on intact unfixed mouse mammary tissue, defined tumor-associated collagen signatures one to three; the third is straightened fibers aligned perpendicular to the tumor boundary. Invasion ran predominantly along aligned fibers, and the signatures showed in pre-palpable tumors — descriptive imaging, no effect sizes, no p-values (Provenzano and colleagues 2006). Two years later the same group made it causal: more stromal collagen — stroma being the supporting tissue around a tumor, as against the cancer cells — roughly tripled tumor formation, p below 0.00001, and about tripled lung metastasis in a mouse (Provenzano and colleagues 2008).

Then the humans. In 196 breast cancers the third signature carried hazard ratios between 3.0 and 3.9 — a hazard ratio of three meaning events arrive three times as fast — for disease-specific and disease-free survival, independent of grade, size, receptor status, HER2 status and node status, HER2 being a growth receptor whose level sets treatment (Conklin and colleagues 2011); that abstract reports no confidence intervals, an unusual gap for a prognostic study. A decade later, signatures in 995 patients reproduced the prognostic signal with strong disease-free-survival areas under the curve — the single number for how well a score separates the patients who will have an event from those who will not — and hazard ratios, these ones reported with confidence intervals (Xi and colleagues 2021).

Stiffness is measured in patients now, not only gels: magnetic resonance elastography found tumor tissue stiffer than normal in 107 colorectal cancer patients (Wang and colleagues 2025); and in mice, cross-linking stiffened matrix and induced invasion of an oncogene-initiated epithelium, while less cross-linking by lysyl oxidase, the enzyme that welds collagen fibers, impeded malignancy (Levental and colleagues 2009).

Now four hard facts the field learned the hard way. One: broad matrix metalloproteinase inhibition failed three phase III programs begun in 1997 and 1998 — marimastat, prinomastat and BAY 12-9566 — and two sponsors shut their programs down (Zucker and colleagues 2000). One of them, the largest randomized pancreatic-cancer study run to that point, put 414 patients with unresectable pancreatic cancer on three marimastat doses or gemcitabine: no significant survival difference, p equals .19, though gemcitabine significantly outlived the 5 and 10 milligram arms at p below .003 and only the 25 milligram arm matched it, with musculoskeletal toxicity in 44 percent on marimastat (Bramhall and colleagues 2001). The post-mortem named three errors: most tumor MMPs come from stromal cells rather than carcinoma cells, the cancer cells induce them there, and some MMP activity generates inhibitors of blood-vessel growth, so blocking everything blocks the brakes too.

Two: the 2006 Nature paper that launched lysyl-oxidase-driven stiffening, whose categorical claim was that inhibiting LOX eliminates metastasis in mice, was retracted in March 2020, fourteen years later (Erler and colleagues 2006; Erler and colleagues 2020). The reason is not recoverable from the PubMed record, which states only that a retraction has been published; anyone reporting why has gone beyond it. Three: the drug failed too. Simtuzumab, an antibody against lysyl oxidase-like 2, added nothing to gemcitabine in 240 patients with metastatic pancreatic adenocarcinoma in a randomized phase II trial: adjusted progression-free-survival hazard ratios of 1.09 and 1.13 at p equals .73 and .61 for the two doses, and overall survival no better either, 0.83 and 1.07 at p equals .28 and .69 (Benson and colleagues 2017). The lysyl-oxidase Atlas page carries the retraction and the simtuzumab null on its face.

Four, and this resets the framing. Delete type I collagen in the alpha-smooth-muscle-actin-positive myofibroblasts — the contractile, collagen-depositing version of a fibroblast — that make most of a pancreatic tumor's collagen, in a dual-recombinase mouse model of spontaneous pancreatic cancer, and precursor lesions and carcinoma emerge faster and survival shortens, through Cxcl5 upregulated via SOX9, recruiting myeloid-derived suppressor cells and suppressing CD8 T cells (Chen and colleagues 2021). For two decades that dense stroma was the enemy to be stripped away. Somebody stripped it away, and the cancer got worse. The human data agree: in colorectal cancer the absence of a stromal reaction at the invasion front marks the worse outcome, pooled hazard ratio 1.77 for cancer-specific survival, interval 1.55 to 2.01, across seven retrospective cohorts and 5,932 patients, overall survival 1.53, interval 1.39 to 1.68, the authors asking for prospective validation (Reitsam and colleagues 2026) — though what is scored there is direct tumor-adipocyte contact, not collagen.

One last reversal, in the receptor. DDR1 is a collagen receptor with a tyrosine kinase domain, and the kinase is the half a drug developer would reach for. The function that aligns the fibers and excludes T cells needs the extracellular collagen-binding domain, not the kinase: an untethered external domain alone rescued knockout tumor growth, and antibodies against it disrupted alignment and inhibited growth (Sun and colleagues 2021). A kinase inhibitor would have been aimed at the wrong half of the molecule — this show's reading; the paper reports which domain is required and says nothing about anyone's drug program. And every signature study is retrospective; the line about one in five patients who might have been undertreated is a model's inference, not a trial result.

Now the parasite side, enzyme classes corrected on air. Ac-MTP-1 is an astacin-like metalloprotease, a zinc family unrelated to the human MMPs, secreted by activated hookworm infective larvae. Recombinant Ac-MTP-1, expressed in insect cells, digested gelatin, collagen, laminin and fibronectin; antiserum localized the enzyme to secretory granules and the channels opening at the larval surface, and cut the recombinant enzyme's collagen digestion by 85 percent and larval migration through tissue in vitro by 70 to 75 percent, against 5 to 10 percent for preimmune serum, while chelators cut skin penetration by 43 to 61 percent (Williamson and colleagues 2006). That is the closest structural parallel to MT1-MMP — enzyme released at the organism's own surface, cutting locally — and still a different protein family.

Schistosome skin invasion is the richest case and needs three corrections. First the catalytic class, named once as a set, because the whole correction is that they differ: a serine protease cuts using a serine residue, a metalloprotease using a zinc ion at its center, an aspartyl protease using a pair of aspartate residues, usually at acid pH, a cysteine protease using a sulfur-bearing cysteine, and the astacin-like enzymes are a branch of the metalloproteases named after the first one found, in a crayfish. The schistosome's is a chymotrypsin-family serine protease, cutting with a serine residue and not a zinc ion — the most important correction here, because the line about parasites using MMPs the way tumors do leans on it. The 1985 purification in the Journal of Biological Chemistry gives 30,000 daltons, isoelectric point 8, pH optimum 9, calcium dependence 2 millimolar (McKerrow and colleagues 1985). The enzyme sits in vesicles in the acetabular gland cells at the larva's head end, ruptured inside host skin beside degraded matrix (Fishelson and colleagues 1992); the schistosoma Atlas page carries the life cycle.

Second: the cercaria does not climb through dermal collagen. McKerrow's own 1985 purification in the Biochemical Journal reported a type-specific collagenase active against the basement-membrane collagens IV and VIII but not the interstitial collagens I, III and V (McKerrow and colleagues 1985, Biochemical Journal), and dermal collagen is overwhelmingly I and III. It opens the sheet and cuts the glue proteins but not the dermal rope. Keep what McKerrow wrote himself: 'The substrate specificity of this enzyme resembles that of the proteolytic enzymes which facilitate tissue invasion by inflammatory cells and tumor cells.' The man who ran the experiment drew that parallel forty years ago. A 2025 systematic review of cercariae encountering humans, searched to May 2024, separates the same two jobs: permeation through the acetabular glands as the cercaria becomes a schistosomulum, the young worm it is once through the skin, from the head-gland enzymes driving the epidermal-to-dermal transition (Panzner, Utzinger and Keiser 2025). Third: it is species-specific. The serine proteases dominating S. mansoni secretions were absent from S. japonicum, whose glands carried forty-fold greater cathepsin B-like activity (Dvorák and colleagues 2008), and a bird schistosome does the job with a cathepsin B — the cathepsins being protein-cutting enzymes that normally work inside cells at acid pH and that parasites secrete instead — TrCB2, 77 percent similar to S. mansoni cathepsin B2 (Dolecková and colleagues 2009). Deeper than isoforms: the acetabular-gland enzyme of Schistosomatium douthitti, an agent of swimmer's itch, is a metalloprotease of 50,000 daltons with no significant primary-structure homology to the 30,000-dalton S. mansoni serine protease (Amiri and colleagues 1988). And Necator americanus larvae secrete every mechanistic class, but when live larvae were set against hamster skin, penetration was significantly inhibited only by pepstatin A, which blocks aspartyl proteinases (Brown and colleagues 1999).

And the hardest claim in this run needs its wording repaired, because of a paper from this year. Leishmanolysin is the most predominant protease in Schistosoma japonicum cercariae, and the authors' own bioinformatic analysis places their leishmanolysin-like peptidase isoform 1, SjLLPi1, within the M8 matrix metalloprotease family; immunofluorescence puts it in the acetabular glands and their ducts of the cercarial head. Then the blocking experiment, in mice: cercariae treated with a monoclonal antibody against SjLLPi1 gave a marked reduction in skin-invading parasite numbers as early as 30 minutes after infection, significantly lower numbers at 7 days, and significantly lower worm and egg burden at 42 days (Chen and colleagues 2025, PLoS Pathogens) — a loss-of-function skin-invasion result in a living host, in the very species just used to show how species-specific the enzymology is. So state the chemistry as a family: none of these enzymes belongs to the family MT1-MMP belongs to — the hookworm's astacin, S. mansoni's chymotrypsin-family serine protease, Necator's aspartyl proteinase, and S. japonicum's leishmanolysin, which its own authors do call a matrix metalloprotease, a family label, not a human MMP. Three parasites through skin, three catalytic chemistries, a fourth that borrows the name, and the tumor's enzyme in none of them.

The protozoan counterparts are thinner. Leishmania overexpressing the surface metalloprotease gp63 migrated better through matrix, degrading collagen IV and fibronectin — a gain-of-function report, not a knockout (McGwire and colleagues 2003). Entamoeba histolytica yields three collagen-binding proteins with type I collagenolytic activity, the 30-kilodalton receptor assignment resting on the authors' own words, indirect evidence, in a 1992 paper with no permanent digital identifier and unreplicated since (Rosales-Encina and colleagues 1992). Two honesty notes. The strongest Fasciola hepatica evidence is a surrogate rather than a collagen assay: newly excysted juveniles prefer substrates with proline two residues back from the cleaved bond, the P2 position, shared in that subfamily only with the adult fluke's CL2 and with vertebrate cathepsin K, the mammalian collagenolytic cathepsin (Cancela and colleagues 2008). Trypanosoma cruzi collagen degradation is usually passed along on a review assertion (Watanabe Costa and colleagues 2016), which is not evidence; the primary enzyme-and-substrate paper does exist, and this part owes a correction for having once said it did not, which it pays at the end.

Last, the asymmetry of rigor. Against SmCB1, the Schistosoma mansoni gut cathepsin B that digests host blood protein, peptidomimetic vinyl sulfones produced the most potent inhibitors reported, subnanomolar, with two co-crystal structures and selectivity over the human ortholog (Jílková and colleagues 2021). In cancer, three phase III protease-inhibitor programs failed outright. Why uneven: SmCB1 is a digestive gut enzyme with a sparable human counterpart, not a collagen-cutting invasion enzyme shared with the host. The parasitologists still got the better drug out of the same idea. One thing is still owed to the tumor side before the ledger: the half of the architecture that has moved since 2009.

What Stiffness Does, and the Two Newest Results That Cannot Be Quoted

The stiffness half has moved since 2009, and two primary papers from the last year carry it forward. They fail in opposite directions: one reports a number too large to trust, the other none at all.

Start with the instrument, stiffness measured rather than built. Atomic force microscopy presses a very fine probe into a sample and reports how hard the sample pushes back, so the number comes out of the tissue itself. Pressed into human cutaneous, oral and lung squamous cell carcinoma — the cancer of the flat surface cells that line skin, mouth and airway — the tumor tissue came back stiffer than the adjacent normal tissue. Then the mechanism, in cells on gels of adjustable stiffness: stiff matrix activated Piezo1, a channel that opens when the membrane around it is stretched, which promoted proliferation and invasion through the Hippo pathway, the growth-restraining circuit whose output protein YAP enters the nucleus and drives growth genes once the restraint comes off; and activating Piezo1 induced TGF-beta-1, which pushes fibroblasts into the myofibroblast state, which lays down more collagen, which stiffens the matrix further (Jiang and colleagues 2025). The loop closes on itself — the mechanism the 2009 cross-linking experiment already in this part predicted (Levental and colleagues 2009), now with a named receptor and a feedback arm.

The same paper has a clinical arm, and it contains a number that should make a listener suspicious. In a cutaneous squamous cell carcinoma cohort of 53 patients, high Piezo1 expression tracked poor differentiation at p equals 0.0034 and recurrence at p equals 0.047, and disease-free survival carried a hazard ratio reported as 181.03 at p equals 0.01 (Jiang and colleagues 2025). Ten would already be a very strong signal; 181 is not an enormous effect, it is an unstable estimate — the kind a survival model throws out when the two groups separate almost completely, with next to nothing left to count in one of them. The tell is the p value beside it: a p of 0.01 is ordinary evidence, and ordinary evidence does not come with a 181-fold point estimate attached, and the abstract reports no confidence interval. The biology may well be right. The number should not be quoted, by this show or anyone else.

The architecture story moved too, toward something usable. Hematoxylin and eosin are the two routine dyes every pathology laboratory uses: the blue-and-pink slide. Nuclear features defined relative to the tumor-associated collagen signatures were extracted from those ordinary images in 941 invasive breast cancers and compressed into one score by least absolute shrinkage and selection operator regression, which shrinks all but a few candidate measurements to zero. The score was an independent prognostic factor on multivariable Cox regression — the survival analysis that asks whether a measurement still predicts outcome after the known predictors are accounted for — and the full model stratified patients better than the clinical model alone (Li and colleagues 2025). But the abstract reports no hazard ratios, no areas under the curve, no confidence intervals and no p values, so there is no effect size to quote; and one of its three corresponding authors is at a commercial pathology-imaging company, the other two at a university photonics laboratory, on a paper validating an image-derived score. The thing that would make this deployable — just the slide the hospital already cut — arrives without its numbers.

Then why the enzyme drugs failed, in the field's own words rather than mine. A 2026 review of lysyl oxidase-like 4 describes it cross-linking collagen and elastin, raising matrix stiffness and aberrant mechanotransduction — the conversion of a mechanical force into a chemical signal inside a cell — and in tumors reinforcing the matrix barrier and keeping T cells out; then it states where the therapy stands: clinical development remains limited by functional heterogeneity, the lack of robust biomarkers, and compensatory mechanisms within the lysyl oxidase family (Tian and Li 2026). Block one member and the others cover for it — a mechanism for the simtuzumab null this part already carries, where an antibody against lysyl oxidase-like 2 moved neither progression-free nor overall survival in 240 patients (Benson and colleagues 2017). One parasite reaches that same enzyme family, not by carrying one of its own: Cryptosporidium parvum delivers its own RNA transcript into the nuclei of infected human intestinal epithelial cells, where it is recruited to the promoters of the host's cadherin 3 and lysyl oxidase-like 4 genes and suppresses both, the suppression attenuated by knocking the parasite transcript down (Ming and colleagues 2018). In vitro, in human cell lines — and the direction is suppression: evidence against parasites stiffening matrix, not for it.

The duality is now stated outright in print, its affiliation in the open. A 2025 narrative review titles itself on the duality of collagens in metastasis, but its abstract names only the enabling half: collagens mediating tumor cell adhesion, migration, invasion, survival, immune evasion and therapeutic resistance, in a matrix far from a passive scaffold. What the second half of that duality is, the indexed record does not say (Carnazza and colleagues 2025). Its first author is at a commercial nutraceutical company; that does not make the review wrong, but a listener should know it. And the restraining half is not the review's evidence — it is the mouse already in this part, where deleting type I collagen from the myofibroblasts of a spontaneous pancreatic cancer made the lesions come faster and shortened survival (Chen and colleagues 2021).

One last check, and it is the one that stops the phrase invasive matrix program from meaning cancer. In rheumatoid arthritis the fibroblast-like synoviocytes — the fibroblast-type cells that line a joint and make its lubricating layer — take on what the authors themselves call an aggressive, 'tumor-like phenotype', their phrase and their spelling. The collagen-binding integrin alpha-11-beta-1 was strongly expressed in rheumatoid synovium and at the invasion sites where joint-lining cells meet cartilage; take the gene for it, Itga11, out of mice that carry a human tumor-necrosis-factor transgene and get arthritis on their own, and cartilage degradation, bone erosion and synoviocyte attachment all fell significantly, by clinical score, by micro-computed tomography — a high-resolution X-ray scan that shows where bone has been eaten away — and on the slide (De Giuseppe and colleagues 2025). The caveats sit with it: no effect sizes, confidence intervals or p values in the record, the destruction experiment is in mice, and the authors advance this integrin as a therapeutic target, an interest worth naming; the comparison to cancer-associated fibroblasts is theirs, not mine. So a non-malignant disease runs a collagen-receptor-driven invasion program: receptor, adhesion, proteolysis, tissue destroyed. The second time in this part that invasion turns out not to belong to cancer; the first was the gene-targeted fibroblasts at the top of it (Sabeh and colleagues 2004).

The One Tissue That Is Licensed to Invade

There is one normal human tissue that does all of this on purpose. Every time a pregnancy implants, cells from the outer layer of the embryo cross a basement membrane and keep going into the wall of the uterus. Those cells are trophoblast, the first differentiated cells of the embryo, the ones that build the placenta rather than the baby. Comparing them to a tumor is not this show's idea: the canonical review opens by noting the striking similarities between the proliferative, migratory and invasive properties of placental and cancer cells (Ferretti and colleagues 2007). The comparison is older than that review, and older than the index I can search. The placenta and the trophoblast have their own Atlas page.

The primary result is from 1991 and still the cleanest thing here. In culture, only early-gestation human cytotrophoblasts — the unfused, still-dividing trophoblast cell — invaded a basement-membrane-like substrate. The invasive cells made both metalloproteinases and urokinase-type plasminogen activator, an enzyme that switches on a second, broader protein-cutting system the body keeps for clot breakdown. Metalloproteinase inhibitors, and a function-perturbing antibody specific for the 92-kilodalton type IV collagen-degrading metalloproteinase — the enzyme now called MMP-9 — completely inhibited invasion, while inhibitors of the plasminogen activator system had only a 20 to 40 percent effect (Librach and colleagues 1991). The paper's own words for what it watched are transient and tumor-like, and developmentally regulated. The enzyme class the chatbot transcripts wanted for parasites is real here, in us, and runs on a schedule. One line for consistency: that substrate is the sheet, not the rope, and MMP-9 was excluded earlier for crossing fibrillar collagen.

The parts list is in that same review, its citation corrected in the same breath. Ferretti and colleagues catalog what trophoblast and cancer cells share: epidermal growth factor and its receptor; hepatocyte growth factor and its receptor Met; vascular endothelial growth factor with its receptors, all as autocrine and paracrine loops, autocrine meaning the cell makes the signal it answers itself and paracrine one its neighbors answer; increased integrins; downregulation of E-cadherin; MMP-9 expression and activation, which the review calls a prerequisite; and PI3-kinase and AKT, a relay carrying growth and survival instructions inward from the cell surface (Ferretti and colleagues 2007, e-published October 2006). Now the correction: the transcripts cited this paper under an invented title in an invented journal and credited it with a finding it does not contain — there is no PD-1 or PD-L1 immune tolerance anywhere in it.

The checkpoint half is real, and belongs to a different paper. A checkpoint is one of the brakes on an immune response, the class of brake that antibody drugs release in cancer treatment. B7-H1, the molecule now called PD-L1 — a surface protein that switches off an approaching T cell — is abundant as message — messenger RNA — in term placenta, with cytotrophoblasts a source; as protein it is expressed by the syncytiotrophoblast, the fused outer layer, and by extravillous cytotrophoblasts, both juxtaposed to maternal blood, while placental stromal cells lack it. Expression was low in first-trimester placenta compared with second and third, p below 0.05, and was raised in cultured cytotrophoblasts by interferon-gamma, an immune signaling protein, or by epidermal growth factor (Petroff and colleagues 2003). The second author is Lieping Chen, who identified B7-H1, a conflict worth naming, though the 2003 record declares none.

The transcripts' citation for the blood-vessel signal is the clearest forgery here, and the real paper's finding cuts against the use it was put to. Every bibliographic element was wrong: the first author's initials, an author who is not on it, the journal, and a receptor that does not appear in it. The paper is Sharkey and colleagues 1993, in the Journal of Reproduction and Fertility. Message for vascular endothelial growth factor appeared in villous and extravillous trophoblast and in fetal and maternal macrophages, the maternal ones in the decidua, the lining of the uterus as it is rebuilt for a pregnancy; but the strongest site was maternal macrophages adjacent to Nitabuch's stria, the zone of necrosis at the implantation site, and the authors conclude that macrophages are the primary source (Sharkey and colleagues 1993). The angiogenic signal at the implantation site is mostly the mother's, not the invader's.

Then the brake, which disciplines the comparison instead of decorating it. In tissue from 15 women who had a cesarean hysterectomy for placenta accreta spectrum — the condition in which the placenta implants too deeply and will not separate, typically where the decidua is missing — against 10 controls operated for uterine atony, a uterus that fails to contract after delivery, collagen I, collagen IV, fibronectin and laminin were highly expressed in decidua and myometrium, the muscular wall of the uterus. Culturing trophoblasts with those molecules induced the protease HtrA4 and its invasive program, and knocking HtrA4 down or pre-treating with recombinant HtrA1, the decidua's own inhibitor, inhibited that invasion (Chen and colleagues 2025, Placenta). The matrix itself induces the invasive program, exactly as dense fibrillar collagen induces invadopodia earlier in this part, and the host supplies the brake. Where the brake is missing the invasion does not stop — and the disease that results is an obstetric emergency, not a cancer.

One gene on that list needs its dates corrected. snail is not a 1987 discovery. Simpson described dorsal, twist and snail in Genetics in 1983, the ventral presumptive mesoderm failing to invaginate in snail homozygotes — the cells that should fold inward to make the fly's internal organs do not fold (Simpson 1983); Grau, Carteret and Simpson compared seven snail alleles in 1984 and found a graded effect along the dorsal-to-ventral axis (Grau, Carteret and Simpson 1984); and Boulay, Dennefeld and Alberga cloned it in Nature in 1987 (Boulay, Dennefeld and Alberga 1987). 1987 is the cloning. And one plain sentence: this snail is a fruit-fly gene, named in early-1980s Drosophila genetics for a reason no paper I opened states, and it has nothing to do with the freshwater snail that part 002's liver fluke grows up in. The gene, the program named after it and these dates are on the Atlas page for epithelial-to-mesenchymal transition and snail.

The program that gene names has been put in its place by one experiment. Epithelial-to-mesenchymal transition is the program by which a cell that has been part of an orderly sheet lets go of its neighbors, loses its top-to-bottom polarity, and takes on the loose, mobile character of connective tissue; Snail and Twist are the two transcription factors held responsible for it. Deleting either one in mouse models of pancreatic ductal adenocarcinoma, the common cancer of the pancreas arising from its duct lining, did not alter the emergence of invasive cancer, dissemination or metastasis; suppressing it instead raised cancer-cell proliferation and nucleoside transporters, the carriers that bring gemcitabine into a cell, which made the tumors more sensitive to that drug and lengthened survival (Zheng and colleagues 2015). In that model the program named for a snail is not rate-limiting for invasion, and switching it off made the chemotherapy work better.

And a parasite does this in the same tissue, this part's subject, not an aside. Chorionic villous explants from normal human placentas — pieces of real placental tissue kept alive outside the body — were incubated with Trypanosoma cruzi, the agent of Chagas disease, and the parasite produced destruction of the syncytiotrophoblast, disorganization of the basal lamina, which is the basement membrane under another name, and of type I collagen in the villous stroma, all in proportion to the number of parasites; the authors read that as the parasite's own proteolytic activity (Duaso and colleagues 2010). Keep the limit on it: tissue coming apart in a dish with the enzyme inferred rather than identified, though a purified T. cruzi collagenase does exist, and this part pays that correction at the end (Santana and colleagues 1997). That is human tissue, outside the body, with a protozoan taking apart both materials this part opened on: the second explant result beside Entamoeba in human colon. Trypanosoma cruzi has its own Atlas page.

So invasion is not a malignant property. It is a normal, depth-limited, host-braked program, and a tumor cell, a trophoblast and at least one protozoan each use some of its parts — so finding invasion in a tumor says nothing about who is doing the invading.

What Is Shared, What Is Only Analogy, and the Experiment Nobody Has Run

Everything genuinely shared at the molecular level comes to three things, no more — the electrical half is part 007's ledger. The substrate, the local geometry of the cutting, and one real molecular convergence in human tissue (Sabeh and colleagues 2004; Williamson and colleagues 2006; Thibeaux and colleagues 2014). Everything past them is cross-literature inference between fields using different organisms, models and readouts.

And the absences in brief, counted again in the close: no parasite MT1-MMP; no parasite lysyl oxidase, cross-linking program or mechanotransduction story; no parasite tumor-associated collagen signature; no deformability measurement for any parasitic helminth larva; and not one published experiment with a live parasite and a tumor cell in the same three-dimensional matrix.

Then the asymmetry: several parasites get through tissue without cutting anything. Invasion of dendritic cells by Toxoplasma gondii made them hypermigratory through three-dimensional collagen, and blocking antibodies against five integrins barely affected it, the shapes consistent with amoeboid movement (Kanatani and colleagues 2015). Apicomplexan gliding needs nothing cut at all (Heintzelman 2015).

Cancer cells and white blood cells do the same, inside a measured limit: protease-free migration stops when the nucleus is squeezed to about ten percent of its cross-section (Wolf and colleagues 2013). The one verified overlap is the host's enzymes conscripted by a parasite, not a parasite enzyme shared with a tumor. So if there is a convergence in the invaders' own tools, it is not enzymology — it is that both know when not to cut.

One pattern runs through both halves, part 007's charge and this part's movement: the physical observation held up and the inference about cancer did not. Cancer cells really are less negative in culture; the channel really is overexpressed in tumor samples (Hemmerlein and colleagues 2006); the sugar coat really is denser (Adams and colleagues 2018); alternating fields really do slow dividing cells in a dish (Kirson and colleagues 2004); the collagen really is aligned and stiffer around a human tumor (Conklin and colleagues 2011). Then the marker turned up in almost half of healthy women on estrogens (Ortiz and colleagues 2011); the channel's inherited gain-of-function caused overgrown gums and malformed fingertips rather than cancer (Kortüm and colleagues 2015); an antibody against a real difference did nothing (Benson and colleagues 2017); and the stroma everyone wanted stripped away was holding the cancer back (Chen and colleagues 2021). That is the shape of the error a parasite-cancer argument has to avoid.

So the verdict answers to an arithmetic. The 2026 update from IARC, the World Health Organization's cancer agency, attributes 2.3 million new cancers, 12 percent of all cancer, to twelve Group 1 infectious agents — Group 1 being the agency's top category, meaning the evidence that an agent causes cancer in people is sufficient — led by Helicobacter pylori at 760,000 and human papillomavirus at 750,000 (Rumgay and colleagues 2026); the Atlas page for IARC carries its history. The three Group 1 parasites together account for 9,700 cases: about 5,900 bladder squamous cell carcinomas for Schistosoma haematobium, about 3,800 intrahepatic cholangiocarcinomas, cancer of the bile ducts and part 002's subject, for Opisthorchis viverrini and Clonorchis sinensis. Label that on air: the 9,700 total, both subtype figures and the 19,000-case bucket they sit in come from the paper's full text rather than its indexed abstract, which carries only the 2.3 million, the 12 percent and the leading viruses and bacteria; and attributing each tumor site to its organism is a secure inference from the paper's list of twelve agents, not a sentence it prints. 9,700 is 0.42 percent of the infection-attributable total, roughly 0.05 percent of all cancer, about one seventy-eighth of the H. pylori figure.

The same paper hands over the reason to distrust that number, in its full text again rather than its abstract. Only about 19 percent of the world's population, and about 2 percent of Africa's, sits inside the high-quality registries it drew on; registries code where a tumor is and what its cells look like, never what caused it, so the subtype counts were back-derived from registry proportions applied to global totals. IARC states plainly that molecular epidemiological methods for bacterial and parasitic carcinogens are less well standardized than for viruses, and it published no uncertainty intervals at all. Set the corpus beside that: to pin the hepatitis fractions in liver cancer the same group pooled 857 publications from 81 countries (Cao and colleagues 2026), and for cirrhosis, 520 publications covering 1,376,503 patients (Alberts and colleagues 2022). Nothing like it exists for any parasite. The number is small. It is also the number of a question nobody has funded.

Here is the experiment, a protocol somebody could fund. Broad-range eukaryotic sequencing — the 18S ribosomal RNA gene or the internal transcribed spacer, the open-question test part 001 defines — with human reads subtracted, run on 100 to 200 consecutive archived formalin-fixed paraffin-embedded blocks from one tumor site. Consecutive, not selected. Batch-matched negative controls. A contamination threshold agreed in writing before the first block is opened. Per-sample reagent cost reported. One headline number: the fraction of blocks returning non-human sequence. What fixation does to that DNA is on the Atlas page for formalin fixation.

What exists is adjacent, and must be cited as such. PathoChip, a microarray with probes for every publicly available virus sequence plus hundreds of bacteria, fungi, parasites and helminths, works on DNA and RNA from fixed tumor blocks (Baldwin and colleagues 2014). Metagenomic sequencing of archived blocks hit 100 percent sensitivity and 88.24 percent specificity for mycobacterial granuloma, on 31 confirmed cases inside a 65-case series — organism DNA survives fixation, and roughly one negative in eight reads as positive (Sun and colleagues 2023). None of those is pan-eukaryotic sequencing, none used a consecutive series, and no published pilot, cost estimate or feasibility study exists for this proposal.

Which is why the controls are the design, not a formality. The one large screen of archived tumor sequence for non-human reads, Poore and colleagues in Nature in 2020, was reanalyzed and found to report microbes that were not in the samples at all (Gihawi and colleagues 2023), and Nature retracted it in July 2024 (Poore and colleagues 2024); part 005 owns that pipeline. A positive result without a threshold fixed in advance is worth nothing.

Two things I will not say. Not that no billing code exists for pathogen testing on a tumor block, because no paper documents that; what is true is that no published study addresses who would pay. And the cost of not asking is measurable, which part 005 counts in full: a median of 41 sexual-health visits before diagnosis, interquartile range 18 to 65, among the 30 seropositive women out of 51 screened in a Barcelona pilot (Roure and colleagues 2022). Then the symmetry: a negative result across a few hundred consecutive tumors would narrow this hard, and nobody has run it. A positive result would not mean a parasite caused those cancers; it would mean one was there — the discipline part 001 applied in Medellín.

What the Two Materials Prove, and the Dish Nobody Has Filled

This part is accountable to one question, not the season's: do a cancer cell leaving a breast duct and a worm larva crossing the dermis get through tissue the same way? Whether a parasite causes a cancer nobody counts is a different question, set above. This verdict is narrower, a count. Three things shared. Four things missing. And one experiment nobody has run.

First the substrate, the one identity nothing qualifies: both meet the same rope-like type I scaffold and the same thin type IV sheet, built the same way in every vertebrate host. Second, the geometry is shared and the molecules are not. The tumor cell cuts in the thin shell against its own surface with MT1-MMP, found by eliminating every other candidate in fibroblasts from gene-targeted mice (Sabeh and colleagues 2004); the hookworm larva cuts with Ac-MTP-1, an astacin from a different zinc family, fed down channels that open at its own surface, at the percentages given earlier (Williamson and colleagues 2006). And neither number came from a human: the elimination in cultured cells and animals, the hookworm figures on a recombinant enzyme and on larvae in glassware.

Third, the one true convergence in this half, in human tissue outside the body. In human colon explants, blocking matrix metalloproteinase activity abolished remodeling of the fibrillar collagen and prevented Entamoeba histolytica trophozoites from invading the mucosa; with trophozoites present the human enzymes MMP-1 and MMP-3 were overexpressed, and the parasite's own cysteine proteinase CP-A5 performed the cleavage that activates pro-MMP-3, which activates pro-MMP-1, with recombinant CP-A5 alone enough to rescue trophozoites that lacked it (Thibeaux and colleagues 2014). That is the tumor's own machinery, conscripted and operated by a parasite — though there is no tumor anywhere in that experiment, and what is shared is the host's enzyme, not anything passing between the two invaders. It is also no longer one result in one tissue. In human gingival fibroblasts, Entamoeba gingivalis, the oral amoeba of periodontal pockets, strongly activated collagenases in wild-type cells but not in cells with the host trafficking protein VAMP3 deleted by CRISPR-Cas9, the authors setting the work against intestinal amoebiasis (Rosenfeld and colleagues 2025) — conscription shown by taking away a piece of the host rather than the parasite. One disclosure: the indexed abstract strips the organism names, so the species is read from the record's index terms and its periodontal setting. Entamoeba has its own Atlas page.

Then the second protozoan, in full earlier in this part: the Trypanosoma cruzi placental explants, where syncytiotrophoblast, basal lamina and stromal type I collagen came apart as a function of parasite number, the enzyme read behind it rather than identified (Duaso and colleagues 2010). The enzyme named in the first case and inferred in the second is the whole distance between a mechanism and an observation — though T. cruzi is no longer only inferred to cut collagen: a secreted 80-kilodalton proteinase purified from T. cruzi hydrolyzed human collagen types I and IV at neutral pH, and not albumin, laminin or immunoglobulin, and cut native type I collagen in rat mesentery (Santana and colleagues 1997).

Now the absences, counted, with the searches behind them. No parasite MT1-MMP. No parasite carries a lysyl oxidase of its own and no parasite cross-linking program has been described; the one parasite result touching that family suppresses the host's LOXL4 rather than adding cross-links, so anyone who says parasites stiffen matrix the way tumors do is inventing it. No parasite collagen signature, no fiber-alignment score measured in parasitized tissue. And no deformability or elastic-modulus measurement for any parasitic helminth larva. Narrow it: atomic force microscopy has been run on nematodes — Fakhrullina and colleagues in 2016 on Caenorhabditis elegans, Akhatova and colleagues in 2018 on live C. elegans and Turbatrix aceti — both free-living, both the outside of a cuticle rather than whole-animal deformation, neither a parasite, and a title search for any parasitic larva returns nothing. Which is why the softness comparison has one side, and that side is strong: by atomic force microscopy on live metastatic cells from the pleural fluid of patients with suspected lung, breast and pancreas cancer, the cancer cells were more than 70 percent softer than the benign cells lining the same body cavity in the same sample, with a standard deviation over five times narrower (Cross and colleagues 2007). The Atlas page for collagen and invasion carries those absences, each with the search behind it.

Then the absence this section is named for, said precisely. No published experiment has put a live parasite and a tumor cell in the same three-dimensional matrix and measured how either one gets through it. Said loosely it is false: an avirulent, nonreplicating uracil-auxotroph strain of Toxoplasma gondii preferentially invaded the immunosuppressive CD11c-positive antigen-presenting cells of the ovarian carcinoma microenvironment and triggered rejection of established mouse ovarian tumors (Baird and colleagues 2013) — a parasite and a tumor in one experiment, but an immunotherapy one, with no matrix and no invasion readout, and into the tumor's antigen-presenting cells rather than its carcinoma cells. On the matrix side the nearest thing is an extract rather than an animal: soluble egg antigen from Schistosoma mansoni inhibited two human colorectal cancer cell lines dose-dependently, and 25 micrograms given locally significantly inhibited tumor growth in a mouse xenograft, human cells grown in a mouse (Pekkle Lam and colleagues 2024) — a worm product rather than a worm, pointing the opposite way from the result part 003 aired for Schistosoma haematobium on bladder urothelium, where soluble egg antigen raised proliferation. Neither is what this part needs: a living larva and a carcinoma cell in one matrix, under the same conditions on the same afternoon.

So specify it, as a design somebody could fund. Live schistosomula, or hookworm third-stage larvae, the infective stage that waits in soil, together with a human carcinoma line, in the same three-dimensional collagen gel, on the same day, in one operator's hands. Measure the pore size of that gel rather than assuming it; the gap width governs everything downstream. Run arms with and without a broad protease inhibitor cocktail, the manipulation that separated squeezing from cutting on the cancer side (Wolf and colleagues 2003). Two readouts: distance traveled, and nuclear deformation against the arrest limit the same group established, at the square-micrometer figures given earlier (Wolf and colleagues 2013). Then the arm that closes the one-sided comparison: the larva's own stiffness, by atomic force microscopy or microfluidic squeezing, on the indentation protocol run on those patients' cells. One honesty clause fixed in advance: a helminth larva is a multicellular animal with a collagenous cuticle and a muscular body wall, so whatever comes back is not a single cell's indentation stiffness, and the paper would have to say which quantity it reports.

Then the real obstacle, which is not technical. Every method in that design already exists; the atomic force microscopy has been run on worms, just not on parasitic ones (Fakhrullina and colleagues 2016). What it needs is a parasitology laboratory and a cancer laboratory in the same building on the same morning, handing living larvae in the hour they are ready to carcinoma cells waiting in a gel. Those two are funded separately and publish separately — a reading of the absence, not a sourced claim. But nothing else explains why the cheapest experiment in this part is the one nobody has done.

What a result would mean, in both directions. If a larva's progress fell off with pore size the way a tumor cell's does, that would show two organisms obeying one set of mechanics — not a parasite causing a cancer, and not shared enzymes; the inhibitor arms are there to tell cutting from squeezing, and the catalytic families already differ. If the larva would not move in the gel at all, that would not show it cannot move in a person: a gel has no vessels, no temperature gradient and no host signals. And a larva that moved exactly like a tumor cell would leave the season's question untouched, because mechanics cannot tell anyone whether a pathologist ever looked.

Then the hand-off. The drug arithmetic — what dose of an antiparasitic ever reaches a human tumor — belongs to part 009. The animals that already have this disease, and the five resemblances worth keeping, belong to part 010. What this part leaves behind is narrower and firmer than the claim it tested: the same wall, the same geometry, unrelated enzymes, one real convergence in human tissue, and a dish nobody has filled.

Sources

Every paper referenced on air, in the order it comes up. Links go to the publisher via DOI.